Deep learning-based prediction of negative air ion concentrations using TCN-attention and meteo-temporal fusion transformer

Abstract

The concentration of negative air ions (NAI) is a key indicator used for evaluating air quality and quantifying the health benefits that forest ecosystems provide. Accurate prediction of NAI concentration holds significant value for ecological tourism planning as well as public health service improvement. However, existing research has primarily analyzed factors influencing NAI concentration using conventional statistical methods, leaving a gap in fully capturing complex nonlinear interactions among meteorological drivers and in multi-step time series prediction. In this study, based on continuous observation data from 31 environmental monitoring stations in Jiangsu Province from 2021 to 2023, a 1-day prediction model TCN-Attention and a 7-day sliding prediction model MTFT (Meteo-Temporal Fusion Transformer) were developed. TCN-Attention achieved , with MAE = 167.5 ions/cm³, 2.7%–7.8% lower than those of conventional machine learning methods. MTFT achieved an overall across the 7-day forecast horizon and exceeded conventional methods by 18.6%–29.8% in relative R², indicating that the advantage of sequence-based temporal modeling increased with forecast horizon. Forecast errors of MTFT were lower in summer and autumn (seasonal MAE = 63.9 and 69.3 ions/cm³) than in winter and spring (83.1 and 85.3 ions/cm³), highlighting the temporal stability of the meteorology–NAI relationship as an important factor in seasonal predictability. To identify the effects of meteorological factors on NAI concentration, a separate SHAP analysis was conducted, revealing the inflection point of temperature effect at 25 °C, the diffusion threshold at wind speed 1.5 m/s, and optimal generation conditions under high temperature and moderate humidity (>25 °C + 70–85% RH). These meteorological response patterns were physically plausible and consistent with the descriptive results. Under the four representative meteorological scenarios, MTFT yielded MAE values of 59.5–82.1 ions/cm³, comparable in magnitude to its overall test MAE of 73.4 ions/cm³, with performance remaining comparable across scenarios. The dual-model prediction system constructed in this study has already been applied to the meteorological operational system of Jiangsu Province.

1 Introduction

Negative air ions (NAI) are molecules and light ion clusters in the atmosphere that carry negative charges (Goldstein and Arshavskaya, 1997). Recently, they have gained widespread attention due to their potential to enhance immune and respiratory functions, and offer therapeutic adjunctive effects for emotional disorders (Bowers et al., 2018). NAI can be formed in the air through ionization processes induced by ultraviolet radiation, cosmic rays, and radioactive substances in the ground. Additionally, the friction between wind and ground, discharge during thunder and lightning, splashing of waves and waterfalls, corona discharge at plant leaf tips, and photoelectric effects during photosynthesis are also important NAI generators in the atmosphere (Jiang et al., 2018). Relatively high levels of NAI can effectively reduce suspended pollutants and alleviate severe health problems caused by air pollution. Air with NAI concentrations of 1000–1500 ions/cm³ is considered clean by the World Health Organization (WHO), and NAI concentrations above 1000 ions/cm³ are beneficial to human health (Li et al., 2022). Meanwhile, NAI also exhibits obvious removal effects on indoor particulate matter, which can effectively reduce the concentrations of PM10, PM2.5, and PM1 (Nadali et al., 2020). Due to their health benefits, NAI has become an essential indicator for evaluating the health effects of urban air quality and forest ecosystems (Ling et al., 2010), and is also used in fields such as ecosystem service evaluation (Tao et al., 2023).

With the acceleration of urbanization, environmental pollution and ecosystem degradation are becoming increasingly prominent, and the air in urban areas is regarded as a major public health concern (Ling et al., 2010). Urban green spaces and forest cover can release large amounts of NAI through corona discharge and photosynthesis, providing residents with a cleaner living environment and leading to ecosystem service value (Luo et al., 2020). Human basic physiological needs can be satisfied when NAI concentration is higher than 700 ions/cm³, while NAI concentration is generally higher from summer to autumn, which is considered to be the most suitable time for health-promoting outdoor activities (Li et al., 2022). This provides important guidance for the optimal layout of ecological tourism and rational seasonal planning of health-promoting outdoor activities. However, NAI belong to the highly reactive species in the atmosphere, and their lifetime is generally short, on the scale of tens of seconds to minutes, when they collide with aerosols and disappear rapidly in dense areas, resulting in a relatively large difference in concentration data (Jiang et al., 2018; Tammet, 2006). In addition, NAI concentrations show great variation over time and space under the combined influence of meteorological (temperature, humidity, wind speed) and air quality (PM2.5, PM10 and O₃) factors (Miao et al., 2018; Wang et al., 2020).

Currently, most studies on NAI have focused on describing spatiotemporal distribution characteristics and analyzing the influencing factors, using statistical methods to reveal the control mechanism (Miao et al., 2018; Wang et al., 2020). However, predictive research has rarely involved future periods. At present, there is a gap between “retrospective analysis” and “prospective forecasting”, which limits the translation of research results. Therefore, it is of great practical significance to construct an NAI concentration prediction model to meet the practical needs of forecasting services for the health tourism industry, environmental managers, and those engaged in outdoor activities.

The spatiotemporal distribution patterns of NAI concentration and its control mechanisms constitute a central theme in current research. In the spatial dimension, NAI concentration exhibits clear landscape gradient characteristics. In forest environments, NAI concentrations can reach several thousand ions/cm³, while urban green spaces typically show concentrations below 1500 ions/cm³, and open undeveloped land exhibits only several hundred ions/cm³ (Yun et al., 2024; Li et al., 2021; Wu et al., 2011). This spatial differentiation is closely associated with differences in local land morphology, vegetation characteristics, and climatic factors (Wang et al., 2020), as well as increased anthropogenic emissions and decreased vegetation cover during urbanization (Shi et al., 2021). NAI concentration is also significantly influenced by vegetation type. Generally, coniferous and mixed forests show higher NAI concentrations than deciduous broadleaf forests and shrublands, which is attributed to the smaller radius of curvature at needle tips that provides stronger corona discharge capability (Li et al., 2021; Chen et al., 2021). In the temporal dimension, NAI concentration shows multi-scale periodic variations. Diurnal patterns vary among environments, including U-shaped distributions with high values in the morning and evening and low values at noon, ∩-shaped distributions with a midday peak, and gradually rising daytime patterns (Wang et al., 2020; Zhan and Li, 2026). These differences are closely associated with climatic characteristics and vegetation types in study regions. Seasonal variations differ by climate zone. In temperate regions, concentrations are higher in summer and autumn than in winter and spring, while subtropical regions may show inverse patterns with winter being higher than summer. This is related to the suppressive effects of high temperature on plant stomatal conductance and photosynthesis (Zhang et al., 2021). Long-term monitoring has further revealed pronounced delayed responses of NAI concentration to precipitation, with concentrations potentially rising substantially within hours after rainfall (Shi et al., 2021).

Meteorological factors constitute the primary drivers regulating NAI concentration changes, though the mechanisms of action differ substantially among factors. Relative humidity (RH) typically shows positive correlation with NAI concentration, as water molecules function as carriers for major ionic components such as OH⁻ during NAI formation. High humidity contributes to both NAI generation and extended survival time (Miao et al., 2018). Temperature effects can be nonlinear; in forests and urban green spaces, a quadratic equation provided a better fit for the relationship between air temperature and NAI concentration than a linear equation (Wang et al., 2020). Reported effects of wind speed on NAI concentration vary across environmental settings (Luo et al., 2020). Atmospheric particulate matter (PM) can negatively modulate NAI concentration through adsorption and recombination processes, particularly under stagnant boundary-layer conditions. However, on daily-to-synoptic timescales, PM variability is also strongly modulated by meteorological conditions, including boundary-layer mixing, precipitation scavenging, and wind-driven dispersion/transport (Chen et al., 2020; Miao et al., 2019). Therefore, this study focuses on meteorological drivers as the primary forcing pathway. It is important to note that these factors do not operate in isolation; their relative effects on NAI concentration vary across temporal and spatial scales (Wang et al., 2020).

In general, NAI concentration is jointly influenced by vegetation, meteorological conditions, air quality, and radiation, and the relative effects of these factors vary across temporal scales. Recent studies have used correlation analysis, regression models, random forest, and structural equation modeling to identify environmental drivers and distinguish their direct and indirect effects (Li et al., 2025). However, such studies have largely focused on factor-response analysis or contemporaneous estimation rather than forecasting future NAI concentrations. By contrast, air-pollutant forecasting has adopted multi-model fusion and various deep learning architectures, including CNN-LSTM, sparse-attention Transformer, and multi-output LSTM models (Zhang et al., 2020; Gilik et al., 2022; Zhang and Zhang, 2023; Zhou et al., 2019), while recent reviews have summarized broader opportunities and challenges for deep learning in this field (Tao et al., 2025). Nevertheless, multi-step time-series forecasting of NAI concentration using deep learning remains insufficiently explored.

The rapid development of deep learning (DL) technology has brought about a paradigm shift in ecological time-series forecasting. In closely related meteorological time-series forecasting tasks, CNN-LSTM-based deep learning models have shown competitive performance in predicting monthly atmospheric temperature and other climate variables across different cities, indicating their suitability for nonlinear environmental sequence modeling (Guo et al., 2023; Guo et al., 2024a). Although conventional machine-learning methods such as support vector regression (SVR) and random forest (RF) can model nonlinear relationships, they may be less effective than deep sequential models in capturing complex temporal features and long-term dependencies (Mohapatra et al., 2025; Cheng et al., 2025). Deep-learning models can automatically extract high-dimensional spatiotemporal features and integrate heterogeneous inputs from sources such as satellite remote sensing, ground monitoring stations, and meteorological data (Muthukumar et al., 2022). In air-quality forecasting, Long Short-Term Memory (LSTM) networks and Gated Recurrent Units (GRUs) have become widely used architectures and often achieve higher predictive accuracy than conventional approaches for pollutants such as PM2.5 and O₃ (Bernacki and Scherer, 2025). The development of hybrid architectures has further improved forecasting performance. Combining Convolutional Neural Networks (CNNs) with LSTM enables joint extraction of spatial correlations and temporal dependencies and can improve performance relative to single-architecture baselines (Tsokov et al., 2022). Attention mechanisms further strengthen the ability of models to focus on informative time steps and features, while sparse self-attention improves the capture of long-range dependencies (Cao et al., 2024). Transformer architectures consequently offer advantages for long-sequence and multi-horizon forecasting tasks (Lim et al., 2021). In addition, ensemble-learning strategies that integrate predictions from multiple base learners can improve model stability, generalization, and overall forecasting performance (Tian et al., 2024).

Multi-step forecasting is important for ecological time-series applications, but the forecasting strategy determines whether errors propagate across horizons. In recursive strategies, earlier predictions are fed back as inputs for subsequent horizons, so their errors can accumulate; direct multiple-output strategies instead estimate the future vector jointly without recursively reusing earlier forecasts (Ben Taieb et al., 2010). Signal-decomposition methods provide another means of representing nonlinear and non-stationary series: Empirical Mode Decomposition (EMD) adaptively separates a signal into intrinsic mode functions, whereas Variational Mode Decomposition (VMD) concurrently extracts band-limited modes (Huang et al., 1998; Dragomiretskiy and Zosso, 2014). Encoder–Decoder architectures provide a Sequence-to-Sequence (Seq2Seq) framework for mapping an observed multivariate sequence to a multi-step future sequence, and attention mechanisms can improve the use of temporal information during this mapping (Du et al., 2020). Temporal Fusion Transformer (TFT) integrates recurrent processing of local temporal patterns with self-attention for long-term dependencies, variable-selection mechanisms, and interpretable components for multi-horizon forecasting, and was evaluated against multiple benchmark models on several real-world datasets (Lim et al., 2021). Model interpretation methods have also become increasingly common in air-pollution prediction, where they are used to attribute model outputs to input features and characterize feature–response patterns (Houdou et al., 2024). Transfer learning offers a further strategy for data-limited settings by adapting knowledge learned from source monitoring stations to target stations with limited historical observations, which can improve predictive performance at the target sites (Ma et al., 2024).

Compared to the vigorous development in air quality prediction, research on NAI concentration forecasting has remained limited, with published time-series work including statistical approaches such as ARIMA (Li et al., 2021). Existing studies have employed methods such as multiple regression analysis and random forest to analyze NAI responses to environmental factors (Dou et al., 2025). One recent study estimated the spatial distribution of NAI concentrations by combining hybrid stacked machine-learning models with Kriging Spatiotemporal Augmentation (Liang et al., 2024). However, the primary objective of these studies is to reconstruct NAI spatial distribution based on synchronous observation data, which fundamentally differs in methodological foundation from time series prediction tasks that predict future concentrations based on historical sequences. Applications of deep learning methods in NAI time series prediction are rare, which is closely related to special challenges faced by NAI prediction modeling. NAI monitoring networks are relatively sparse, and long-term monitoring records remain limited, constraining the time-series data available for model training (Tao et al., 2023). NAI concentration is regulated by the synergistic action of multiple factors including plant physiological processes, micrometeorological conditions, and atmospheric ionization, resulting in complex dynamic variation mechanisms (Miao et al., 2018; Jiang et al., 2018). Furthermore, the short lifetime of small negative atmospheric ions causes rapid dissipation, which can contribute to high-frequency variability in measured concentrations and further complicate time-series modeling (Maruyama and Katayama, 2003). Therefore, investigating the applicability of deep learning methods in NAI time series prediction and developing prediction models that can effectively capture its dynamic characteristics represents an academic challenge worthy of in-depth research in this field.

Model interpretability is an important area in ecological forecasting research. In air-quality forecasting, SHAP and partial dependence plots have been used to identify influential pollutant and meteorological variables, characterize nonlinear response patterns, and examine interactions among predictors (Wei et al., 2025). In NAI research, several studies have instead used random-forest importance measures to rank environmental drivers under different observational settings (Luo et al., 2020; Xie et al., 2023). However, conventional random-forest importance measures primarily provide global feature rankings and do not, by themselves, reveal effect direction, response thresholds, or sample-specific contributions. SHAP can provide both global and local feature attributions (Lundberg and Lee, 2017); to our knowledge, its application to NAI forecasting remains scarcely documented.

Model robustness is another important consideration for evaluating the application value of prediction models. In air-pollution forecasting, predictive performance has been evaluated across forecast horizons and regional scales, while abrupt concentration changes remain challenging (Wen et al., 2019). Season-stratified analyses have also shown marked performance differences: daily PM2.5 forecasts were more accurate in summer but less accurate in winter, when extreme concentration peaks were more frequent, and errors also varied among city groups (Xiao et al., 2020). For NAI, the relative importance of environmental drivers can vary among seasons (Wan et al., 2024), yet the forecasting literature reviewed above has not established whether predictive performance remains stable across seasonal or meteorological regimes. Accordingly, this study evaluates deep-learning forecasts across seasons and predefined meteorological scenarios and applies SHAP to a dedicated Random Forest model using ten observed meteorological factors as predictors to quantify their contributions to the model predictions. Together, the gaps in NAI forecasting methods, multi-step prediction, interpretability, and robustness motivate the present study.

To address the deficiencies in terms of methods, temporal scales, and mechanism analysis in existing NAI prediction research, we systematically carry out spatiotemporal characteristic analysis and deep learning-based prediction modeling research on NAI based on NAI concentration and meteorological observation data of 31 environmental monitoring stations in Jiangsu Province from 2021 to 2023. The research contents include: (1) NAI concentration spatiotemporal distribution pattern characteristics are revealed through spatial interpolation and statistical analysis, and the meteorological control mechanism is quantitatively analyzed. (2) A TCN-Attention model is constructed for next-day NAI concentration prediction, which captures multi-scale temporal dependencies of historical NAI time series through dilated causal convolution and focuses on critical local historical time points using attention mechanisms. (3) An MTFT (Meteo-Temporal Fusion Transformer) model is constructed for 7-day NAI concentration sliding prediction, which establishes long-term dependencies on historical NAI time series by combining 1D convolution and multi-head self-attention mechanisms. (4) A systematic comparison is conducted on the performance of the above deep learning models and conventional methods such as ElasticNet, SVR, Random Forest, XGBoost and LightGBM.

For mechanism analysis, a dedicated Random Forest model was fitted to observed NAI concentration using ten observed meteorological factors as predictors; SHAP dependence patterns and representative-case attributions were then used to characterize nonlinear meteorological relationships, while a bivariate heatmap of observed concentrations provided complementary evidence. Based on the classification of typical meteorological conditions (high/low temperature, with/without precipitation), the model prediction stability under various meteorological condition scenarios is validated. The major contributions of this study include: (1) The deep learning method is systematically applied to NAI concentration time series prediction, and a “1-day + 7-day” multi-target prediction methodological system is established. (2) A dedicated Random Forest model using ten observed meteorological predictors, interpreted with SHAP, characterizes nonlinear meteorological relationships with NAI concentration and provides interpretive evidence for the identified meteorological response patterns. (3) The robustness of the model under different typical meteorological conditions is verified, providing technical support for the practical application of NAI concentration prediction.

2 Materials and methods

2.1 Study area

Jiangsu Province is located on the eastern coast of China, belonging to the central Yangtze River Delta (116°18′–121°57′E, 30°45′–35°20′N). The total area is 107.2 thousand km² (Figure 1). The region is positioned in the transition zone from subtropical humid monsoon climate to temperate monsoon climate, with four distinct seasons. The annual mean temperature is approximately 15.8 °C, annual precipitation is approximately 864 mm, with hot and rainy summers and mild winters with less precipitation (Zhang et al., 2024). The topography is generally flat, with elevations typically below 10 m, though exceeding 50 m in western hilly areas and southern low mountain regions. Within the region, 31 meteorological observation stations (green dots) are distributed, enabling acquisition of long-term, high-density time series data on NAI concentration and meteorological factors. Various ecosystem types are present in the study region, encompassing forests, wetlands, agricultural lands, and urban green spaces. In addition, the large seasonal differences in meteorological conditions offer a perfect experimental area for studying the NAI concentration spatiotemporal characteristics and verifying the prediction model’s stability under different meteorological conditions.

FIGURE 1

2.2 Data and preprocessing

Jiangsu Meteorological Bureau provided the data used in this study, targeting continuous daily observation records from 31 negative ion automatic monitoring stations from January 2021 to December 2023. The dataset includes basic station information (name, longitude, latitude), NAI concentration (ions/cm³), and synchronously observed meteorological factors (temperature extremes and means, relative humidity, pressure, wind speed, daily precipitation). The NAI concentration values are station-level daily mean values derived by averaging the hourly observations of each station; that is, each record corresponds to the daily mean concentration of a single monitoring station on a single day. Preprocessing was implemented in three stages. First, physically acceptable ranges for each factor were established based on “Ground Meteorological Observation Specifications” and outliers beyond these ranges were removed. Second, first-order differences between adjacent time points were calculated, and time series consistency checks using 3 standard deviations as threshold identified and labeled abrupt change anomalies. Third, minor missing values (<2%) were filled using linear interpolation. Data completeness after quality control exceeded 98%.

In feature engineering, to maximally leverage NAI concentration temporal autocorrelation and meteorological driving information, three categories of features were constructed in addition to the observed meteorological factors. (1) Multi-scale lag features: extracting historical NAI concentration values at lags of 1–14 days to represent short-term and medium-term temporal dependencies. A 21-day historical window was used for the 1-day prediction model (TCN-Attention), and a 28-day window for the 7-day sliding prediction model (MTFT). (2) Rolling statistical features: calculating 7-day moving averages and standard deviations, together with the corresponding 7-day anomaly and first-order difference terms, corresponding to the weekly variation cycle, for noise reduction and local fluctuation description; for the 7-day model, 7-day rolling means of temperature and relative humidity and 7-day cumulative precipitation were additionally included. (3) Time encoding features: mapping periodic information to continuous variables through sine/cosine transformation, avoiding discontinuities at category boundaries; the annual cycle (day of year) was encoded for both models, and the monthly cycle additionally for the 1-day model. All features were created strictly following temporal order principles using only historical information to avoid data leakage. Although spatial IDs, land-use classes, and vegetation indices were not included as explicit predictors in the final model, site-specific differences can be indirectly represented to some extent by temporal features constructed from each station’s own historical NAI sequence. Specifically, NAI lag features and rolling statistical features can reflect the relatively stable background concentration level and short-term variability of each station, while multi-day continuous meteorological sequences may also contain site-specific microclimatic differences. Therefore, even when two stations have similar temperature and humidity conditions on a given day, the model may still generate different predictions based on their preceding NAI trajectories, rolling statistical characteristics, and continuous meteorological evolution. Prior to modeling, features were standardized using RobustScaler to reduce outlier influence. To avoid information leakage, scaling parameters were estimated only from the training portion of each time-ordered split and then applied to the corresponding validation and test data. The detailed time-series evaluation and model validation protocol is described in Section 2.3.5. Specifically, hyperparameters of conventional machine learning baseline models were optimized through grid search combined with a time-aware TimeSeriesSplit procedure within the training portion of each rolling-origin split, rather than standard k-fold cross-validation. This ensured that hyperparameter tuning was performed only on temporally ordered training subsets and did not use future observations from the corresponding test window. For deep learning models, an internal validation subset was established at the tail of the training data for early stopping, preserving model parameters with the best validation performance. Table 1 summarizes the input features of the two prediction models by category.

2.3 Methods

2.3.1 1-Day prediction model: TCN-attention

To achieve high-precision next-day NAI concentration prediction, this study constructed a deep learning architecture integrating Temporal Convolutional Network (TCN) and attention mechanism. Model input is a 3-dimensional tensor with shape , where represents historical window length and represents feature dimension. Features comprise the observed meteorological factors together with three categories of derived features: lag features, rolling statistics, and time encodings.

The TCN module constitutes the model core, with its primary innovation being the efficient extraction of multi-scale temporal features using Dilated Causal Convolution (Bai et al., 2018). Unlike standard convolution, introducing dilation factor enables the convolution kernel to sample input sequences at exponential intervals, substantially expanding receptive field without increasing parameter count. For a network with kernel size and layer count , the theoretical receptive field can reach . This study employs a 6-layer TCN with dilation factors set as , in which the exponentially increasing dilations of the first five layers capture variations from daily to multi-week scales and the final layer with dilation 1 integrates the multi-scale features; the resulting 65-day receptive field fully covers the 21-day input window. Causal convolution ensures that output at time depends only on inputs prior to , strictly satisfying causality constraints in time series prediction. Furthermore, to mitigate gradient vanishing and facilitate deep feature propagation, each TCN layer combines a gated activation unit with a residual connection, with form given by (Equation 1):where is the hidden state at layer , and are the filter and gate convolution kernels of that layer, * denotes dilated causal convolution with dilation , σ(·) is the sigmoid function, and ⊙ denotes element-wise multiplication; a 1 × 1 convolution is applied to the residual path when the input and output channel dimensions differ.

The attention mechanism module applies adaptive weighting to each historical time point in the TCN output temporal feature sequence . Specifically, a context-vector attention pooling scheme is adopted (Yang et al., 2016): an alignment score is computed for each time step from its feature vector through a learnable nonlinear projection, the scores are normalized across time steps via Softmax to obtain attention weights, and the feature sequence is aggregated into a context vector by weighted summation (Equation 2):where W and v are learnable parameters and c is the resulting context vector. This mechanism enables the model to dynamically focus on critical time points, identifying and emphasizing historical moments with maximum contribution to target prediction (e.g., extreme weather or abrupt pollution changes). Finally, the context vector is mapped to a scalar output through two fully connected layers; through the residual scheme described below, this output is converted into the next-day NAI concentration prediction .

Rather than regressing the absolute concentration directly, the model adopts a residual prediction target: the network estimates the increment of the next-day concentration relative to a reference value , i.e., , where denotes the observed concentration on day t+1; the final prediction is reconstructed as . In this formulation, the reference value supplies the current concentration level, so that the network only needs to learn the day-to-day change relative to it. Two variants of the model, differing only in the reference value, were trained: in one variant the reference is the previous-day observation, ; in the other, the reference additionally incorporates the day-of-year increment of a smoothed seasonal profile, , where denotes the day of year and is the station-specific day-of-year mean profile estimated using only the training portion of each data split and smoothed with a 15-day moving average. The predictions of the two variants were averaged to produce the final model output.

For training, Huber loss was adopted to balance sensitivity between normal samples and outliers, with AdamW optimizer (initial learning rate ) for optimization; the learning rate was halved whenever the validation error plateaued. To suppress overfitting and stabilize training, early stopping (patience = 20) and gradient clipping (threshold = 1.0) were introduced. To reduce the influence of random initialization, each variant was trained with seven fixed random seeds, and its predictions were averaged over the seven runs before the two-variant averaging described above; the seed values are listed in Table 2.

2.3.2 7-Day sliding prediction model: MTFT

For medium-term (7-day) NAI concentration sliding prediction, this study proposes the Meteo-Temporal Fusion Transformer (MTFT). Unlike single-step iterative (sequential) prediction, MTFT adopts an end-to-end sequence generation paradigm, formulating multi-step prediction as direct mapping , thereby fundamentally avoiding error accumulation associated with recursive prediction (Zhang et al., 2024). Model input is a historical tensor with shape , where days, and is the feature dimension (26 in total), comprising the observed meteorological factors together with three categories of derived features: lag features, rolling statistics, and time encodings.

The architecture incorporates three modules: an input projection layer, a stacked Transformer encoder, and an output mapping layer. In the input projection stage, a linear fully connected layer, applied identically at each time step, maps the input features into a 64-dimensional representation space. Afterward, in order to introduce temporal position information, the method of positional encoding is adopted, producing position embeddings utilizing sine-cosine functions (Vaswani et al., 2017) (Equation 3):where is the time step position, is the dimension index, and is the model dimension. Positional encodings are added to the projected features, enabling the model to recognize sequential order.

The stacked Transformer encoder constitutes the MTFT core. Its design jointly exploits global and local temporal information; a similar global-local combination has been reported in recent hybrid climate prediction models, where CNN-based local feature extraction was jointly combined with LSTM/GRU-based temporal dependency modeling to improve the representation of complex climate time series (Guo et al., 2024b). In MTFT, Multi-Head Self-Attention (MHSA) captures long-range dependencies across the 28-day input window, while local short-term variation information is carried by the lag and rolling-statistics features. The number of attention heads is 4 (), each head dimension is , with computation form (Equation 4):where .

The position-encoded feature sequence is input to a deep feature extraction module consisting of two stacked standard Transformer encoder layers. Each layer contains MHSA and Feed-Forward Network (FFN) sub-layers, employing residual connections and layer normalization (Equation 5):

This hierarchical structure enhances the dimensionality of feature representations through layer-wise abstraction, while maintaining stable gradient propagation via residual connections. The encoder output at the last time step of the input window is taken as the sequence representation and, after dropout regularization, is mapped to a 7-dimensional output vector by a linear fully connected layer. As in the 1-day model, a residual prediction target is adopted: the seven outputs represent the increments (k = 1, … , 7) of the future concentrations relative to a reference sequence; two variants differing only in the reference value— in one, and in the other—were trained, and their predictions were averaged. The reconstructed concentrations constitute the predicted 7-day NAI concentration sequence, represented as (Equation 6):

This study employed the Huber loss combined with an Adam optimizer (initial learning rate 3 × 10⁻⁴) for training. Early stopping (patience = 25, restoring the best validation weights) and learning rate decay (halving the rate whenever the validation error plateaued) were combined to simultaneously ensure the stability of the training process and the optimality of the model. To further suppress overfitting, dropout (0.30 within the encoder layers and 0.20 before the output layer) and L2 weight regularization (10⁻⁴) were applied to the input projection and output layers. To reduce the influence of random initialization, each variant was trained with three fixed random seeds, and its predictions were averaged over the three runs before the two-variant averaging described above; the seed values are listed in Table 2.

2.3.3 Comparison baseline models

To systematically evaluate the performance improvement of deep learning, this study selected five representative traditional machine learning methods to construct the baseline comparison system. Their selection followed the “model lineage completeness” criteria, covering the three core paradigms of (1) linear regularization models representing the foundation of statistical learning, (2) kernel-based methods evaluating the improvement from nonlinear mappings, (3) ensemble learners estimating the upper bound of model fusion strategies’ performance. This system design ensures that the performance difference can be attributed to a specific algorithmic mechanism, rather than random factors or data split bias.

ElasticNet is a linear regularization model representing the combination of L1 (Lasso) promoting sparsity and L2 (Ridge) ensuring stability (Zou and Hastie, 2005). Its objective function can be expressed as (Equation 7):

The parameters and are used to control the strengths of L1 and L2 regularization, respectively. This model is highly effective in dealing with the multicollinearity that is often found in meteorological data. Its linear nature also makes it an excellent baseline for assessing the contributions of the linear component in NAI prediction.

Support Vector Regression (SVR) is based on the principle of structural risk minimization. It maps the input features into a high-dimensional Reproducing Kernel Hilbert Space (RKHS) using kernel functions and seeks optimal hyperplanes with -insensitive loss (Liu et al., 2017). Here, the RBF kernel was employed (Equation 8):

SVR is effective in capturing the complex response relationships between meteorological factors and NAI. However, due to its computational complexity of , it is less efficient when dealing with large-scale datasets.

Random Forest is a representative Bagging-based ensemble learner, building multiple decision trees through Bootstrap aggregation and random feature subset sampling, reducing variance by voting or averaging (Breiman, 2001). It is highly adaptive to high-dimensional features and nonlinear relationships, though the independent tree structure constrains its ability to model temporal dependencies.

XGBoost incorporates second-order Taylor expansion and regularization terms within the gradient boosting framework, effectively controlling model complexity (Chen and Guestrin, 2016). The objective function at iteration is (Equation 9):where and are first and second-order gradients of the loss function, and is the regularization term. This design demonstrates high generalization performance for structured data.

For histogram-based splitting, scanning raw feature values costs , whereas split-point search over histograms costs ( is the histogram bin count). LightGBM also uses a leaf-wise growth strategy, Gradient-based One-Side Sampling (GOSS), and Exclusive Feature Bundling (EFB) to improve computational efficiency while maintaining accuracy (Ke et al., 2017).

The above five methods span from simple linear to complex nonlinear, and from single models to ensemble strategies. Their performance differences provide clues for interpreting the intrinsic structure of NAI prediction problems. For example, good linear model performance suggests predominance of main meteorological driving effects; ensemble substantially exceeding linear indicates existence of complex nonlinear interactions. Furthermore, deep learning gains exceeding ensemble reflect added value of temporal dependency modeling. All models used identical time-aware data splitting methods, with hyperparameters of the conventional machine learning models selected by grid search using TimeSeriesSplit within the training data of each rolling-origin split; the final hyperparameters of all models are summarized in Table 2.

2.3.4 Model interpretability analysis

SHAP (SHapley Additive exPlanations) quantifies the contribution of each input feature to model predictions using Shapley values from cooperative game theory, thereby providing both global and sample-level feature attributions (Lundberg and Lee, 2017). Previous applications in daily PM₂.₅ forecasting have used SHAP to quantify contributions from atmospheric pollutants and meteorological variables (He et al., 2026). In this study, a dedicated Random Forest model was fitted to observed NAI concentration using the ten observed meteorological factors as predictors. TreeSHAP was then applied to the fitted Random Forest model to quantify each meteorological factor’s contribution to its predictions. For model and sample , the SHAP explanation model is defined as (Equation 10):where represents presence/absence of simplified input features, is total feature count, is the baseline prediction (the expected model output over the training samples), and is the Shapley value of feature . The Shapley value for each feature is given as the average marginal contribution across all feature subsets (Equation 11):where is the full input feature set, is a subset not containing feature , and is the model prediction value when using only subset for sample . SHAP value sign directly reflects influence direction: increases prediction, while decreases it. In this study, SHAP was not applied to the TCN-Attention or MTFT forecasting models. The purpose of the interpretability analysis was to quantify the direct influence of meteorological factors on NAI concentration. However, NAI concentration is strongly autocorrelated, so in the forecasting models a substantial share of the predictive information is carried by the autoregressive lag terms; attribution methods applied to these models would consequently assign most of the importance to the lag terms and thereby obscure the meteorological effects of interest. Accordingly, a Random Forest model dedicated to this attribution analysis was constructed, with its input restricted to the ten observed meteorological factors (daily maximum, minimum, and mean temperature; daily maximum, minimum, and mean air pressure; daily minimum and mean relative humidity; and daily mean and maximum wind speed), and trained on the full observational dataset. SHAP values were computed with TreeExplainer, which is designed for tree-based models and calculates the baseline prediction directly from the model’s training data, so that no separate background dataset was required; the computation was performed on 500 randomly selected samples. On this basis, SHAP dependence plots reveal the nonlinear effects and threshold characteristics of individual factors, and a bivariate temperature–humidity heatmap of observed concentrations corroborates the synergistic effect between the two factors, providing interpretable support for meteorological control mechanisms on NAI concentration.

2.3.5 Model evaluation

Model performance was comprehensively evaluated using coefficient of determination (), Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Mean Absolute Percentage Error (MAPE). With and representing observed and predicted values respectively, the mean of observations, and the sample count, each metric is defined as (Equations 12–15): is used to assess the model’s explanatory power of the observed variance, MAE and RMSE are used to evaluate the absolute magnitude of error (Chai and Draxler, 2014), and MAPE is used to evaluate the scale-independent relative error.

To account for the temporal dependence of daily NAI observations, model evaluation adopted an expanding-window rolling-origin cross-validation scheme, a time-series-aware strategy for predictor evaluation with ordered observations (Bergmeir and Benítez, 2012). Three consecutive half-year test windows were defined along the chronological axis—July–December 2022, January–June 2023, and July–December 2023—with the training set in each split expanding from January 2021 up to the day immediately preceding the test window. In each split, only historical data preceding the test window were used for model training and hyperparameter tuning, while the test window was kept completely independent until final evaluation. For the conventional machine learning models, hyperparameters were selected by grid search using TimeSeriesSplit within the training data. For the deep learning models, hyperparameters were selected on an internal validation subset carved from the tail of the training split, which was also used for early stopping. This design preserved chronological order during scaling, tuning, validation, and testing, thereby reducing future-information leakage and improving the seasonal representativeness of model evaluation.

The two forecasting tasks were organized at spatial scales consistent with their intended applications: the 1-day task was modeled and evaluated on the station-level daily series of the 31 monitoring stations to support station-scale applications, whereas the 7-day task was modeled and evaluated on the provincial-mean daily series averaged over the same 31 stations to provide regional-scale guidance for medium-term planning. Because the splits were defined purely along the time axis, all 31 stations are represented in both the training and test sets of every split for the 1-day task, and no station-based selection was involved. Under this scheme, the three test windows contain 5,703, 5,608, and 5,677 station-day test samples for the 1-day task. For the 7-day task, each test sample is a forecast window covering seven consecutive days from its forecast origin; the three test windows contain 178, 175, and 178 forecast windows. Evaluation metrics were pooled across stations and time: for the 1-day task over all station-day samples, and for the 7-day task over all individual daily observations; overall metrics were computed on the three test windows combined, and the day-specific metrics of the 7-day task were reported by lead day.

To assess robustness for operational application, the 7-day MTFT forecasts were further evaluated under different meteorological scenarios. Scenarios were defined along two axes, temperature and precipitation, according to the meteorological conditions on the forecast start date of each forecast window in the rolling-origin test periods. For temperature, two levels were defined by daily mean temperature >25 °C (high) or <15 °C (low)—thresholds corresponding to the major inflection points of the temperature effect identified in the SHAP interpretability analysis—and windows with intermediate temperatures (15 °C–25 °C) were excluded. For precipitation, presence or absence was determined using 0.1 mm as the threshold (the common meteorological threshold for the occurrence of precipitation). The combinations of the two axes constituted four mutually exclusive meteorological scenarios. Within each scenario, evaluation metrics were computed over all individual daily observations (forecast window × Day 1–7), in the same manner as the overall test metrics. Furthermore, seasonal evaluation of the 7-day forecasts was conducted using the typical meteorological season division (spring: March–May, summer: June–August, autumn: September–November, winter: December–February).

Table 2 summarizes the final hyperparameters and training settings of all models used in this study. All experiments were conducted on a Windows workstation equipped with an Intel Core i9 processor, 64 GB of RAM, and an NVIDIA GeForce RTX 4090 GPU. The TCN-Attention model was implemented in PyTorch, the MTFT model in TensorFlow/Keras, and the conventional machine learning models and the SHAP analysis in scikit-learn, XGBoost, LightGBM, and SHAP. The code is available from the corresponding author upon reasonable request.

3 Results

3.1 Spatiotemporal characteristics and meteorological driving mechanisms of NAI concentration

Figure 2 displays the spatial distribution and statistical characteristics of NAI concentration at 31 observation stations in Jiangsu Province. From Figure (A), concentration exhibits a latitudinal gradient, with highest values in northern Jiangsu (approximately 4000–4300 ions/cm³), moderate levels in central region (approximately 3000–3500 ions/cm³), and lowest in southern region (approximately 2200–3000 ions/cm³). Panel (B) presents the observed frequency distribution of daily-averaged concentrations, which is right-skewed (mean: 3124.6 ions/cm³, median: 2927.0 ions/cm³, and standard deviation: 830.1 ions/cm³), and the peak of the kernel density curve coincides with the median. The tail in the high concentration region (greater than 4000 ions/cm³) deviates substantially from the normal distribution, as the observed frequency values are much larger than the theoretical ones.

FIGURE 2

One-way ANOVA in Figure (C) confirmed extremely significant differences among three regions (F = 3646.83, p < 0.001). Northern Jiangsu median is approximately 3800 ions/cm³ (interquartile range 3200–4200 ions/cm³), central region approximately 3000 ions/cm³ (2600–3400 ions/cm³), and southern region approximately 2800 ions/cm³ (2500–3100 ions/cm³). Significant anomalies present in the southern region of Jiangsu can be ascribed to either specific locations with dense vegetation or brief meteorological events, such as the aftermath of rain or strong gusts of wind. As illustrated in Figure (D), linear regression analysis reveals that latitude is not a strong predictor of concentration variations (R² = 0.101), explaining merely 10.1% of spatial variation. In the latitude band of 31.0–31.5°N, the annual mean concentration range among southernmost stations is 2200 ions/cm³. Meanwhile, in the 33.0–33.5°N band, the corresponding range for northernmost stations is close to 900 ions/cm³. These findings indicate that intra-zonal latitudinal variability of NAI is more pronounced at lower latitude belts. In conclusion, while NAI concentration across Jiangsu Province displays significant spatial disparities, the latitudinal gradient does not adequately account for these differences, and the decisive influences are localized environmental factors.

As illustrated in Figure 3, the spatiotemporal patterns of NAI concentration in Jiangsu Province from 2021 to 2023 and their variation characteristics at different timescales are depicted. On the annual scale, NAI concentration shows a distinct seasonal periodicity, being highest in summer (about 3400–3500 ions/cm³), lowest in winter (about 2800–3000 ions/cm³), and with a difference exceeding 600 ions/cm³. The 7-day moving average line clearly outlines the periodic curve, and the standard deviation shading area reveals that the year-to-year spread of daily concentrations is narrower in summer and autumn than in winter and spring. This means that the summer peak recurs stably from year to year, whereas concentrations in the colder months are more strongly modulated by inter-annual differences in weather conditions. According to the spatial distribution in Figure (B), the province exhibits significant regional heterogeneity, with high-value regions concentrated within 119°E–119.5°E and 33°N–33.5°N, with peaks exceeding 3900 ions/cm³. This is the hilly and wooded part of the province, in the center and north. The eastern coastal plains, as well as the more heavily urbanized areas in the south, maintain relatively low NAI levels, between 2640 and 3000 ions/cm³. This spatial distribution is in line with patterns of vegetation cover and anthropogenic urbanization.

FIGURE 3

As shown in Figure (C), diurnal variation displays extremely marked seasonal differences. In summer, there is a broad afternoon peak at 14–15:00 (approximately 3500 ions/cm³) owing to increased photochemical reactions and transpiration of vegetation after noon. In spring, the diurnal curve is relatively gentle, with a broad peak at 12–15:00 (roughly 2900–2950 ions/cm³). In autumn, concentrations remain at a high level slightly below that of summer for most of the day and substantially higher than in spring, with more noticeable fluctuations: the concentration declines to a mid-morning trough around 9:00 and then rises to an afternoon peak at 14–15:00 (approximately 3500 ions/cm³). In winter, concentrations stay at the lowest level of the year (approximately 2750–2900 ions/cm³), with only a small uptick around noon to early afternoon (12–15:00). Weak solar radiation and low vegetation activity have caused a decrease in ion generation capacity. In all four seasons, concentrations decline from the afternoon peak toward the evening. Seasonal statistics in Figure (D) quantitatively demonstrate extremely significant differences among four seasons (Kruskal-Wallis H-test = 1675.13, p < 0.001). Median is maximum in summer (approximately 3200 ions/cm³) and minimum in winter (approximately 2800 ions/cm³). Furthermore, summer box plot “box” width is largest, with extreme values potentially reaching 7,000 ions/cm³, indicating maximum variability during this period.

In summary, NAI concentration in Jiangsu Province is regulated by combined influences of seasonal climate variation and regional characteristics of lower boundary surface, forming spatiotemporal structure of “high in summer and low in winter; high inland and low coastal”. Seasonal differences in diurnal variation demonstrate ion generation control mechanisms by meteorological factors such as solar radiation, temperature, and humidity, while spatial heterogeneity emphasizes key roles of vegetation types and anthropogenic activities. These results provide scientific foundation for regional ecological environment quality assessment and health tourism resource development.

Figure 4 demonstrates influence mechanisms of meteorological factors on air ion concentration and their nonlinear interaction effects. Correlation analysis in Figure (A) shows temperature exhibits significant positive correlation with NAI concentration, while pressure shows significant negative correlation, whereas linear correlations of humidity and wind speed are weak. Figure (B) obtains quantitative relationship of main effect through regression (y = 21.0x+2766, R² = 0.631, p < 0.001), indicating NAI concentration increases approximately 21 ions/cm³ for each 1 °C temperature rise. Random forest-based feature importance evaluation in Figure (C) reveals nonlinear mechanisms of factors, with temperature main effect importance reaching 0.394. Additionally, temperature × humidity (0.289) and temperature × wind speed (0.144) interactions contribute substantially, while main effect contributions of humidity (0.040) and wind speed (0.027) are extremely small. This aligns with the weak linear relationships for humidity and wind speed seen in Figure (A), reinforcing that their associations with NAI concentration are mainly due to interactions with temperature instead of being independent linear predictors.

FIGURE 4

The temperature × humidity heatmap shown in Figure (D) shows how these factors act synergistically and depend on temperature. The optimal result is seen when there is a combination of high temperature and moderate humidity (>25 °C + 70–85% RH), which gives the maximum concentration (3400 ions/cm³). In high temperature range, moderate humidity is most favorable for ion generation, whereas in low temperature range (<10 °C), high humidity actually causes concentration reduction, potentially related to enhanced deposition and recombination processes under low temperature. Temperature × wind speed box plot in Figure (E) shows low wind speed (<1.5 m/s) under high temperature conditions contributes most to NAI accumulation, while wind speed effects are small in low temperature range, further supporting nonlinearity of interaction. Multi-year climate data in Figure (F) shows NAI concentration varies synchronously with temperature, with summer peak in July–August corresponding to high temperature and moderate humidity, and winter minimum corresponding to low temperature and high humidity combination. The phenomenon of spring NAI increase lagging temperature rise and autumn NAI decrease likewise lagging temperature decline is consistent with temperature-dependent temperature × humidity interaction shown in Figure (D). Overall, NAI concentration is regulated by temperature, but interactions with humidity and wind speed substantially amplify seasonal amplitude, with the summer combination of high temperature and moderate humidity being the primary cause of peak formation.

3.2 Deep learning model construction and performance evaluation

Figure 5 shows that the effectiveness of the feature engineering strategy was systematically evaluated through a multi-dimensional validation system encompassing four aspects: (1) feature selection, (2) data preprocessing, (3) temporal dependency analysis, and (4) noise reduction. Permutation Importance in Figure (A) reveals hierarchical structure of feature contributions. NAI lag feature Lag1 importance reaches 0.409, substantially exceeding other features, indicating air ion concentration possesses strong temporal dependency. Rolling statistical features (Rollμ7d, etc.) occupy five items among top 20, suggesting effectiveness of capturing local fluctuation patterns through temporal window aggregation. Interaction features among meteorological variables, though relatively lower in importance, are included in top 20, confirming modeling value of nonlinear relationships. Normalization method comparison in Figure (B) shows substantial scale differences among variables (e.g., NAI 2565–3646 ions/cm³, wind speed 0.6–3.5 m/s), with RobustScaler based on median and interquartile range narrowing distribution more than Z-score and demonstrating higher capability to suppress outlier influence. This is important for time series including extreme weather. Autocorrelation analysis in Figure (C) provides theoretical basis for lag selection. ACF(1) = 0.963 indicates strong first-order dependence, PACF identifies lags within 1–7 days as the optimal lag window, while partial autocorrelations essentially decline to the noise level from day 8 onward. This explains why Lag1–Lag6 dominate in importance ranking. Noise reduction analysis in Figure (D) shows 7-day and 14-day moving averages effectively suppress short-term random fluctuations while preserving seasonal characteristics (high in June–August, low in December–February). In summary, dominant role of temporal lags, necessity of normalization, scientific basis of lag orders, and noise reduction capability of rolling statistics collectively form the foundation for constructing high-precision prediction models.

FIGURE 5

Figure 6 systematically evaluates TCN-Attention model’s 1-day prediction capability under the rolling-origin evaluation protocol across three aspects: (1) time series tracking, (2) scatter plot approximation, and (3) error distribution. In Figure (A), prediction curve exhibits high concordance with observation curve over the 18 consecutive months merged from the three rolling-origin test windows, with model capturing the complete seasonal cycle—the summer high-concentration periods of 2022 and 2023, the intervening winter minimum, and the spring recovery—demonstrating high robustness to non-stationary time series. In Figure (B), point cloud is densely distributed along 1:1 reference line, with regression line remaining close to the reference line, showing limited systematic deviation. Figure (C) displays error statistical characteristics, with distribution exhibiting an approximately symmetric shape and peak located near zero error, confirming prediction is largely unbiased. In summary, integrating three metrics (R² = 0.917, MAE = 167.5 ions/cm³, RMSE = 248.8 ions/cm³), TCN-Attention demonstrates strong performance in NAI concentration 1-day prediction task under the time-aware evaluation protocol.

FIGURE 6

Figure 7 compares the 1-day prediction accuracy of TCN-Attention and conventional machine learning methods across four metrics under the rolling-origin evaluation protocol. From Figure (A), TCN-Attention attains the highest R² (0.917), exceeding conventional methods (0.903–0.915), reflecting its advantage in capturing complex temporal dependencies. In Figures (B) and (C), TCN-Attention achieves the lowest MAE (167.5 ions/cm³) and RMSE (248.8 ions/cm³), which are 2.7%–7.8% and 1.2%–7.2% lower than those of conventional methods (best conventional method ElasticNet: MAE = 172.1 ions/cm³, RMSE = 251.9 ions/cm³), respectively. In Figure (D), TCN-Attention MAPE (5.45%) is likewise the lowest, 3.7%–7.2% below conventional methods (5.66%–5.87%), indicating high relative accuracy across varying concentration levels. Meanwhile, differences among conventional methods remain relatively limited (MAE variation = 9.5 ions/cm³), suggesting that conventional machine learning, constrained by its reliance on predefined input features, remains limited in characterizing the nonlinear relationships among multiple interdependent factors, whereas TCN-Attention learns temporal representations directly from the sequences and thereby overcomes this limitation. Overall, through multi-scale temporal feature extraction via dilated causal convolution and focus on critical time points via attention mechanism, TCN-Attention consistently ranks first across all four metrics under the time-aware evaluation protocol, supporting the methodological rationale for adopting the deep learning architecture in this study.

FIGURE 7

Figure 8 evaluates MTFT 7-day sliding prediction capability under the rolling-origin evaluation protocol across four aspects: (1) time series tracking, (2) error distribution, (3) accuracy degradation with increasing steps, and (4) training convergence. In Figure (A), although deviation from observations gradually increases with prediction steps, the overall 7-day prediction performance remains stable (R² = 0.866, MAE = 73.4 ions/cm³, RMSE = 92.8 ions/cm³), maintaining high explanatory power in medium-term tasks. Error distribution in Figure (B) approximates a normal shape, with the mean error close to zero (mean = 2.0 ions/cm³), indicating that the predictions are largely free of systematic bias. Figure (C) shows that error dispersion generally increases at longer forecast horizons, while median errors remain near zero across prediction steps, suggesting that the model maintains broadly stable central tendency despite increased uncertainty. Figure (D) quantifies performance degradation regularity, with R² generally decreasing and MAE generally increasing as the prediction step extends, while the curves become comparatively smoother after the first few forecast days. Specifically, R² decreases from 0.943 (Day1) to 0.824 (Day7), and MAE increases from 46.5 to 87.9 ions/cm³. That is, error accumulation progresses mainly in early prediction steps and becomes relatively stable in the medium-term interval. In Figure (E), the validation loss declines rapidly within the initial epochs and thereafter remains within a narrow range, while the training loss decreases steadily throughout, indicating stable convergence of the training process. In summary, MTFT demonstrates controlled error accumulation rate and good medium-term stability in 7-day prediction under the time-aware evaluation protocol, positioning it as a reliable technical solution for weekly prediction.

FIGURE 8

Figure 9 demonstrates MTFT 7-day prediction capability under the rolling-origin evaluation protocol from two aspects: representative period cases and performance changes with increasing forecast horizon. Four cases in Figure (A) cover all four seasons across the annual cycle, with prediction bars showing high consistency with observation bars across all periods. The MAE of the four representative cases ranges from 56.8 to 77.6 ions/cm³, with the summer case (A-2) performing best (MAE = 56.8 ions/cm³), while the spring case (A-1) shows relatively larger error (MAE = 77.6 ions/cm³). This difference is broadly consistent with the seasonal predictability pattern analyzed in Figure 11. Figure (B) quantifies performance changes with increasing prediction steps, with R² decreasing from 0.943 (Day1) to 0.824 (Day7), and MAE increasing from 46.5 to 87.9 ions/cm³. Notably, the day-to-day increment of MAE diminishes steadily with forecast horizon, narrowing from 15.8 ions/cm³ (Day1 to Day2) to 2.2 ions/cm³ (Day6 to Day7). This pattern suggests that error accumulation occurs mainly in the early prediction steps, while relatively stable performance is maintained at the medium-term prediction stage. Overall, MTFT demonstrates stable prediction capability under varying seasons and concentration levels, maintaining R² > 0.82 explanatory power even at day 7 under the time-aware evaluation protocol, providing reliable support for weekly NAI prediction.

FIGURE 9

Figure 10 systematically compares results of MTFT and five conventional machine learning methods in the 7-day prediction task using four metrics under the rolling-origin evaluation protocol. In Figure (A), MTFT R² is 0.866, exceeding conventional methods (0.667–0.730) by 18.6%–29.8% in relative terms. This demonstrates a clear advantage of the deep time series architecture in capturing complex dependencies, an advantage that is more pronounced than in the 1-day task (Figure 7) as the forecast horizon extends. Error metrics in Figures (B) and (C) similarly show MTFT MAE is 73.4 ions/cm³ and RMSE is 92.8 ions/cm³, lower than conventional methods (MAE 104.7–122.3 ions/cm³, RMSE 131.5–146.1 ions/cm³), with error reduction reaching 29.9%–40.0% for MAE and 29.4%–36.5% for RMSE. In Figure (D), MTFT MAPE is 2.34%, 30.1%–39.5% lower than conventional methods (3.35%–3.87%), demonstrating stable relative accuracy under varying concentration levels. It is worth mentioning that, among the traditional approaches, ElasticNet achieves the best results (R² = 0.730), followed by Random Forest (R² = 0.691) and the gradient boosting models (LightGBM at R² = 0.686 and XGBoost at R² = 0.680), whereas SVR results in the lowest value (R² = 0.667). This finding suggests that linear components remain informative in this task, while conventional models remain less effective than MTFT in extracting sequential information from the time series. Overall, by modeling temporal dependencies and integrating multiple meteorological factors, MTFT continues to outperform conventional approaches across all four metrics under the time-aware evaluation protocol, supporting the methodological rationale for deep learning in mid-range NAI prediction.

FIGURE 10

Figure 11 demonstrates prediction stability under the rolling-origin evaluation protocol from two aspects: (1) degree of error dispersion, and (2) seasonal differences. From Figure (A), Std increases rapidly from 59.3 ions/cm³ at day 1–91.4 ions/cm³ at day 3, after which the growth slows markedly, with the day-to-day increment narrowing from 20.1 ions/cm³ (Day1 to Day2) to 2.5 ions/cm³ (Day6 to Day7) and Std reaching 107.5 ions/cm³ at day 7. This indicates that prediction uncertainty accumulation progresses primarily at the short-term stage, with the increase rate clearly decelerating at medium-to-long forecast horizons. Simultaneously, CV remains within the range of 1.2–1.3 across all steps, indicating that the relative dispersion of prediction errors remains broadly stable across time horizons. Seasonal heatmap in Figure (B) shows clear seasonal differences in prediction difficulty. MAE is lowest in summer (average 63.9 ions/cm³), followed by autumn (average 69.3 ions/cm³), whereas winter (average 83.1 ions/cm³) and spring (average 85.3 ions/cm³) exhibit relatively larger errors throughout the 7-day horizon. This suggests that the summer high-concentration period, dominated by comparatively stable radiation- and temperature-driven diurnal variations, is easier to predict, whereas the more frequent synoptic transitions in winter and spring weaken the day-to-day persistence of NAI concentration and thereby increase prediction difficulty. Overall, prediction stability is primarily regulated by prediction steps, while seasonal differences more strongly reflect differences in the predictability of NAI concentration itself.

FIGURE 11

3.3 Model interpretability and robustness validation

Figure 12 characterizes nonlinear meteorological relationships with NAI concentration through SHAP dependence analyses, an observed-concentration bivariate heatmap, and SHAP attributions for representative cases. Panel (A) demonstrates that the temperature contribution transitions from negative (<15 °C) to positive at around 25 °C, exhibiting a threshold-like pattern in the present dataset. Furthermore, variation in SHAP values across humidity levels at similar temperatures suggests that the temperature contribution is humidity-dependent. Panel (B) shows a positive contribution under low wind speed and a rapid decrease in SHAP values above approximately 1.5 m/s, a pattern consistent with local accumulation under low-wind conditions and enhanced dispersion at higher wind speeds. Panel (C), an observed-concentration bivariate heatmap rather than a SHAP interaction plot, shows that high concentrations are associated with high temperature and moderate humidity (>25 °C + 70–85% RH), indicating a joint meteorological pattern rather than an effect attributable to a single factor. Panel (D) compares the contribution of each factor in representative high- and low-NAI cases in summer. Temperature-related features contribute positively to both cases, but their contributions are substantially larger in the high-concentration case, whereas the contributions of the remaining features differ only modestly, indicating that temperature-related contributions account for most of the contrast between these two representative cases.

FIGURE 12

In summary, the combined SHAP and observational analyses identified three dataset-specific meteorological response patterns: (1) temperature contributions became strongly positive above approximately 25 °C; (2) wind-speed contributions declined above approximately 1.5 m/s, consistent with enhanced dispersion; and (3) high observed NAI concentrations were associated with high temperature and moderate humidity, while the representative summer cases differed mainly in the magnitude of temperature-related SHAP contributions. These findings provide interpretable evidence for meteorology–NAI relationships and physical context for the forecasting results, without implying direct interpretation of the internal decision processes of TCN-Attention or MTFT.

Meteorological condition stratification was performed in Figure 13 to systematically study the stability of MTFT 7-day predictions under different meteorological scenarios. The seven-day forecast windows from the rolling-origin test periods (531 in total) were classified according to the meteorological conditions on their forecast start dates, of which 347 windows fall into the four scenarios; within each scenario the evaluation metrics were computed directly over all individual daily observations (forecast window × Day 1–7), identical in calculation to the overall test metrics. The rationale for classification is as follows: for temperature, 25 °C and 15 °C were used as thresholds, where 25 °C is the critical point marking the sharp increase in temperature effect during SHAP analysis in Figure 12, and 15 °C is the threshold value at the boundary of the negative contribution region. For precipitation, the threshold was set to 0.1 mm, which is consistent with the meteorological definition of “effective precipitation”. Compared with conventional single-factor classification or prioritization, this scheme has three advantages: (1) the four groups are fully mutually exclusive, eliminating classification ambiguity; (2) the factor levels are symmetric, allowing independent evaluation of the effects of temperature and precipitation; and (3) diagonal comparison allows for interaction effects. Samples in the temperature intermediate zone (15 °C–25 °C) were excluded from the analyses (184 of the 531 windows) to clarify the comparison of temperature effects.

FIGURE 13

From the perspective of prediction accuracy, the MTFT model maintains consistent performance across all four scenarios, with MAE ranging from 59.5 to 82.1 ions/cm³ and RMSE from 73.2 to 103.3 ions/cm³, a level comparable to the overall test error (MAE = 73.4 ions/cm³), indicating that no meteorological scenario suffers a disproportionate loss of accuracy. In terms of temperature effects, the mean NAI concentration of the high temperature groups (A + B) (ca. 3420–3470 ions/cm³) is substantially higher than the low temperature groups (C + D, ca. 2880–2980 ions/cm³), supporting the positive temperature driving mechanism shown in Figure 12. In terms of precipitation effects, precipitation does not lead to a marked increase in prediction error: the high-temperature rainy scenario (B) attains the lowest MAE (59.5 ions/cm³) among the four scenarios, and the MAE of the low-temperature rainy scenario (D, 82.1 ions/cm³) is essentially the same as that of its no-precipitation counterpart (C, 81.2 ions/cm³). This suggests that although precipitation processes complicate the meteorological situation, the air ion action mechanism per se is relatively stable and does not result in a marked increase in the complexity of the prediction. The relatively larger errors of the two low-temperature scenarios (81.2–82.1 ions/cm³, versus 59.5–70.6 ions/cm³ for the high-temperature scenarios) likely reflect the more frequent synoptic transitions in the cold season, which weaken the day-to-day persistence of NAI concentration and thereby increase prediction difficulty. Meanwhile, the mean prediction error of each scenario remains within ±16 ions/cm³, less than 0.5% of the corresponding mean concentration level, indicating that predictions are largely free of systematic bias in each of the four scenarios. In conclusion, validation based on typical meteorological conditions demonstrates that the model maintains stable accuracy, with errors of the same magnitude as the overall test level, across the four representative scenarios, confirming its prediction robustness for operational application.

4 Discussion

4.1 Analysis of factors contributing to deep learning model performance advantages

Understanding the task’s inherent structure is key to explaining the disparate model performances in predicting negative air ion (NAI) concentrations. For the seven-day forecasting task, comparison among five conventional machine learning approaches showed that ElasticNet (R² = 0.730) performed better than ensemble learning methods such as Random Forest (R² = 0.691) and XGBoost (R² = 0.680). This result suggests that NAI prediction has a dual-faceted structure. The strong performance of ElasticNet suggests that linear main effects account for a substantial proportion of the predictable variation in NAI concentration, with the high concordance between temperature and NAI concentration (R² = 0.631, Figure 4B) providing supporting evidence. However, as the temperature × humidity interaction heatmap in Figure 4D demonstrates, the direction of the humidity response can reverse across temperature ranges. At high temperatures (>25 °C), moderate humidity is optimal, whereas at low temperatures (<15 °C), high humidity suppresses NAI concentration. These nonlinear conditional dependencies remain relevant, representing a structural limitation of linear models and an important source of the gains achieved by deep learning architectures.

TCN-Attention achieved R² = 0.917 in next-day prediction, with MAE and RMSE 2.7%–7.8% and 1.2%–7.2% lower than those of the evaluated conventional methods, respectively. Its superiority lies in its ability to efficiently model the temporal dependence structure of NAI concentration. Autocorrelation analysis in Figure 5C demonstrates that NAI concentration possesses strong first-order dependence (ACF(1) = 0.963) and maintains significant correlation at lags of 1–7 days. Therefore, high-precision prediction requires the simultaneous integration of short-term inertia and multi-day temporal dependence. TCN-Attention employs 21-day input sequences and expands its receptive field exponentially through dilated causal convolutions, thereby capturing daily-scale variations in shallow layers and extracting weekly-scale trends in deeper layers. Furthermore, the attention mechanism dynamically weights critical historical time points (Vaswani et al., 2017). Compared with recurrent neural networks, TCN provides parallel computation and stable gradient propagation, offering advantages for long-sequence modeling (Bai et al., 2018). While conventional machine learning can also incorporate lagged features, reducing a temporal task to static regression inadequately preserves the sequential coupling and dynamic evolution between time points, which helps explain the methodological advantage of deep learning. Consistent evidence has been reported in daily PM₂.₅ forecasting, where a wavelet-based W-CNN-BiGRU-BiLSTM model combining meteorological variables, air-pollutant indicators, and PM₂.₅ lag terms (t−1 to t−6) achieved high predictive accuracy in Guangzhou, supporting the value of integrating short-term historical sequences with meteorological drivers in daily atmospheric forecasting (He et al., 2025). For seven-day forecasting, MTFT achieved R² = 0.866. As described in Section 2.3.5, the 7-day task is evaluated on the provincial-mean daily series, whose day-to-day variability is much smaller than that of individual station series; its absolute errors are therefore systematically smaller than the station-level errors of the 1-day task and are not directly comparable in magnitude. Its direct multi-output design addresses recursive error accumulation, a core challenge in multi-step prediction. Conventional autoregressive decoding generates forecasts sequentially, so errors at one step can become input noise for the next step, allowing errors to propagate across forecast steps. In contrast, MTFT adopts an end-to-end sequence-generation paradigm that directly outputs a seven-dimensional vector, thereby avoiding recursive feedback of earlier predictions into later forecast steps. Within the same MTFT model, R² decreased from 0.943 on day 1 to 0.824 on day 7, while MAE increased from 46.5 to 87.9 ions/cm³. However, the daily MAE increment narrowed from 15.8 ions/cm³ between days 1 and 2 to 2.2 ions/cm³ between days 6 and 7, indicating that performance degradation was concentrated mainly in the early forecast steps and subsequently became more gradual. MTFT’s meteo-temporal fusion architecture extracts local meteorological patterns through one-dimensional convolution and captures long-range dependencies through multi-head self-attention, enabling it to identify key signals within the 28-day historical window that are associated with NAI concentration changes over the following week. The methodological advantage of hybrid architectures over individual deep learning models has also been observed in an independent PM₂.₅ forecasting study: a CNN-GRU-LSTM model integrating local feature extraction with sequential temporal-dependence modeling outperformed CNN, GRU, and LSTM alone in monthly ambient PM₂.₅ forecasting, with the RMSE values of CNN, GRU, and LSTM being 57.00%, 35.98%, and 32.78% higher than that of the hybrid model, respectively (He and Guo, 2024). Seasonal analysis in Figure 11B shows lower MAE in summer (63.9 ions/cm³) and autumn (69.3 ions/cm³) than in winter (83.1 ions/cm³) and spring (85.3 ions/cm³). The higher errors in winter and spring may be associated with stronger day-to-day meteorological variability during the cold and transitional seasons. Previous studies have reported greater day-to-day temperature variability in winter and spring across China, active cold-frontal processes over the Yangtze River Delta in winter, and a climatological peak in East Asian cyclone and frontal activity in early spring (Kang et al., 2019; Okajima et al., 2025; Wang et al., 2025). Together with the broader interannual variability of winter and spring NAI concentration shown in Figure 3A, these meteorological characteristics indicate less stable environmental conditions during the colder seasons and may partly explain their higher forecast errors. From an application perspective, the seven-day horizon can support health-tourism and forest-recreation planning. The overall model fit (R² = 0.866), together with a day-7 MAE of 87.9 ions/cm³—equivalent to approximately 2.8% of the pooled mean concentration (∼3200 ions/cm³)—supports the model’s potential use in such planning applications.

It is therefore useful to distinguish recursive error propagation from other sources of uncertainty in seven-day forecasting. By directly outputting all seven forecast horizons, MTFT avoids using earlier predictions recursively as inputs and thereby reduces error accumulation introduced by the forecasting strategy itself. Nevertheless, the deterioration with increasing lead time may jointly reflect weakening predictive information from historical inputs, uncertainty in future environmental evolution, and residual model limitations rather than a single physical source. MTFT therefore reduces one algorithmic source of error propagation but does not eliminate the broader predictability limits associated with longer forecast horizons.

Within the seven-day MTFT forecasts, increasing lead time was associated with greater uncertainty. As Figure 11A shows, the error standard deviation increased from 59.3 ions/cm³ on day 1–91.4 ions/cm³ on day 3 and 107.5 ions/cm³ on day 7, while the daily increment narrowed from 20.1 to 2.5 ions/cm³. This initially steep and subsequently gradual increase is consistent with the finite predictability of weather-related processes at longer lead times (Slingo and Palmer, 2011). Moreover, the negative-side outliers in Figure 8C indicate that the model tended to overestimate some extremely low-concentration observations, identifying these conditions as an area of reduced predictive accuracy. This pattern may reflect abrupt environmental changes, limited representation of rare events, and residual model limitations.

Importantly, MTFT performance was further evaluated across four typical meteorological scenarios. The robustness analysis in Figure 13 shows comparable error levels under high-temperature dry, high-temperature rainy, low-temperature dry, and low-temperature rainy conditions, with scenario MAE ranging from 59.5 to 82.1 ions/cm³, of the same magnitude as the overall test MAE (73.4 ions/cm³), and with no pronounced deterioration in any single scenario. These results support the robustness of MTFT within the current monitoring network and test periods and indicate its potential for operational application.

4.2 Consistency between interpretability results and physical mechanisms

While high predictive accuracy is important, environmental applications also require the meteorological relationships used to interpret NAI variation to be physically plausible. As described in Section 2.3.4, SHAP was applied to a dedicated Random Forest interpretability model based exclusively on observed meteorological factors, rather than to the TCN-Attention or MTFT forecasting models. The analysis was therefore used to characterize nonlinear and condition-dependent meteorological relationships with NAI concentration and to evaluate their consistency with the descriptive results in Figure 4 and established physical knowledge.

Figure 12A shows that the temperature contribution is predominantly negative at lower temperatures, particularly below approximately 15 °C (roughly −50 to −100), and increases markedly with temperature, becoming strongly positive in the high-temperature range above 25 °C, where some contributions reach approximately 150–200. This pattern indicates that temperature is not only a major driver of the overall variation in NAI concentration, as shown by the full-sample relationship and feature-importance results in Figures 4B,C, but also exerts a distinctly nonlinear influence. A response of this kind is physically plausible because vegetation activity and meteorological conditions can jointly influence NAI production, with their relative pathways varying among seasons in warm-temperate forests (Cai et al., 2024). Warm, high-radiation growing-season conditions may also favor plant-related photoelectric processes, whereas ion production through the Lenard effect specifically requires droplet fragmentation associated with rainfall, splashing, or other water agitation (Jiang et al., 2018). Previous studies have likewise reported nonlinear and seasonally or environmentally dependent temperature responses (Wang et al., 2020; Wang et al., 2020). Accordingly, 25 °C is interpreted here as a reference marking the high-temperature response regime in the present dataset rather than as a universal physiological threshold.

Figure 12B shows that wind-speed contributions are generally positive under low-wind conditions, with some contributions reaching roughly 20–40, and the overall trend declines toward zero or negative values as wind speed increases beyond approximately 1.5 m/s. Because NAIs are unstable and short-lived, the declining contribution observed here is interpreted as a local transport-and-dispersion pattern rather than as a universal wind-speed response; attachment to aerosols and related processes further contribute to ion loss (Jiang et al., 2018). This pattern is consistent with Figure 4E, in which the high-temperature and low-wind combination corresponds to comparatively high NAI concentration, indicating that wind speed primarily regulates local accumulation and dispersion rather than acting as the principal generation factor. The two representative summer cases in Figure 12D further show that temperature-related features provide the dominant positive contributions in both cases, with substantially stronger contributions in the high-concentration case (3539 ions/cm³) than in the low-concentration case (3181 ions/cm³). Contributions from the remaining meteorological factors are comparatively smaller, and the wind-speed contributions have the same direction in both cases. Thus, the contrast between these two cases is associated primarily with the strength of temperature-related forcing rather than with wind speed alone.

The observational bivariate heatmap in Figure 12C, which was constructed from sampled observations rather than from SHAP interaction values, shows a prominent high-concentration region under high temperature (>25 °C) and moderate relative humidity (70%–85%). This pattern indicates that high NAI concentration is associated with a favorable combination of meteorological conditions rather than with a single factor. Temperature provides the principal positive background, while adequate water availability can facilitate ion formation and persistence; the influence of humidity nevertheless remains conditional on temperature and other environmental processes (Miao et al., 2018). Figure 4D, calculated from the full observational dataset using fixed temperature and humidity categories, shows the same broad high-temperature and moderate-humidity pattern.

Comparison with previous studies suggests that meteorological influences on NAI concentration recur across different environments, including urban parks, urban green spaces, forests, and national parks (Luo et al., 2020; Chen et al., 2021; Shi et al., 2021; Li et al., 2021). However, their relative importance and even response direction vary with vegetation, season, regional climate, and analytical scale: temperature is the central meteorological driver in the present dataset, whereas humidity, water availability, particulate matter, or site conditions are more influential in some other environments. Temperature responses can be nonlinear and seasonally variable, and wind-speed effects may be significant in some settings but weak in others. Thus, the comparatively general feature is not temperature dominance or a single universal numerical threshold, but a recurrent multivariate, nonlinear, and context-dependent meteorological response structure.

In summary, temperature is a central meteorological driver of NAI concentration in the present dataset, while humidity and wind speed modify its influence through condition-dependent relationships. High NAI concentrations tend to occur under a combination of high temperature, moderate humidity, and limited wind-driven dispersion rather than being controlled by any single factor alone. The agreement among full-sample statistics, feature-importance analysis, SHAP dependence patterns, bivariate grouping, seasonal variation, and representative cases provides complementary evidence for the physical plausibility of this meteorology–NAI response structure. These relationships also provide a physical context for the seasonal predictability pattern: the comparatively regular summer and autumn regimes shown in Figure 3A are associated with lower forecast errors in Figure 11, whereas the broader variability and more frequent environmental transitions in winter and spring correspond to greater forecasting difficulty. This cross-figure correspondence suggests, but does not directly prove, that seasonal prediction performance may depend more on the temporal stability of the meteorology–NAI relationship than on the absolute concentration level. It supports the physical interpretation of meteorological response patterns that provide context for the forecasting results, but does not imply that SHAP directly explained the internal decision processes of TCN-Attention or MTFT.

4.3 Limitations and future research directions

Primary limitations of this study stem from objective constraints at data level. First, this study is based on 2021–2023 data from 31 observation stations in Jiangsu Province, which represents the longest continuous, quality-controlled NAI record currently available across this monitoring network. Station density and temporal coverage remain insufficient to completely describe spatial heterogeneity and inter-annual variation of NAI concentration, and a 3-year window cannot fully capture climate teleconnections at multi-year scales (e.g., ENSO, PDO) that may modulate NAI variability. However, NAI monitoring network construction and operation require substantial financial investment and long-term data accumulation, with expansion of station numbers and observation periods depending on continuous advancement of national and local ecological environment monitoring systems, thus not resolvable within single research framework. Second, model inputs are limited to standard meteorological elements and do not include potentially relevant covariates such as boundary-layer conditions, vegetation and stand characteristics, soil moisture, and particulate-matter indicators (PM2.5 and PM10). Recent field studies show that stand structure is associated with NAI variability (Li et al., 2024), while soil-moisture-modulated radon exhalation, boundary-layer conditions, and aerosol loading affect related atmospheric small-ion concentrations (Shabek et al., 2025; Zhang et al., 2025). These data either require field observations with specialized equipment (e.g., boundary layer vertical observations) or lack public products covering study region, with availability constraints exceeding individual researcher capabilities. Specifically, none of the 31 NAI monitoring stations in our network is co-located with an air-quality monitoring station, and cross-network spatial matching from the national CNEMC air-quality network would introduce substantial representativeness error. However, quantifying the independent contribution of particulate matter to NAI variability beyond the meteorologically mediated component captured by our feature set requires a dedicated study design beyond the present scope. Such data-level “bottlenecks” constrain to some degree models’ ability to describe complex ecological processes, while indicating important directions for future information infrastructure development.

In addition, the reported forecasting accuracy applies to stations within the current monitoring network. Because the models rely on lag and rolling features computed from each station’s own observation history, and the rolling-origin evaluation splits the data purely along the time axis with all 31 stations represented in every training set (Section 2.3.5), the evaluation quantifies temporal generalization to future periods rather than spatial generalization to unseen stations. Assessing transferability to stations outside the network would require leave-one-station-out or leave-one-region-out validation, which is left for future work together with the transfer-learning direction discussed below.

Future research can develop in the following directions: integrate multi-source data including boundary layer observations, remote sensing-derived vegetation indices, air-quality covariates (PM2.5, PM10, SO₂, NO₂, O₃) through co-deployed sensors or validated gridded reanalysis products (e.g., MAIAC-AOD, CAMS, CHAP), and numerical weather prediction products, imposing physical constraints on models; introduce spatial interpolation and regional modeling, advancing from observation station-level time series prediction to regional-scale spatial prediction; explore transfer learning strategies to enhance generalization capability in data-scarce regions. In particular, recent daily O₃ forecasting using LSTM in Liaocheng City has shown that O₃ can be effectively modeled within a comparable deep-learning sequence-prediction framework, providing a useful methodological reference for future joint NAI-PM-O₃ prediction once co-located air-quality observations become available (Guo et al., 2025). The framework should also be continuously updated as new operational data become available, and long-term trend analysis should be conducted once a sufficiently long record is accumulated, enabling explicit assessment of inter-annual climate teleconnections.

5 Conclusion

In this study, based on continuous observation data from 31 ecological environment monitoring stations in Jiangsu Province from 2021 to 2023, spatiotemporal characteristics of negative air ion concentration were systematically analyzed, and deep learning-based prediction modeling was implemented. Major conclusions are as follows:

NAI concentration exhibits pronounced spatiotemporal differentiation. Spatially, a “high north, low south” latitudinal gradient is observed, with northern region median approximately 3800 ions/cm³, approximately 1000 ions/cm³ higher than southern region. Temporally, “summer maximum, winter minimum” seasonal variation is observed, with difference between summer peak (3400–3500 ions/cm³) and winter minimum (2800–3000 ions/cm³) exceeding 600 ions/cm³.

Temperature is the primary driver of NAI concentration. Main effect contribution proportion reaches 39.4%, with NAI concentration increasing approximately 21 ions/cm³ for each 1 °C temperature rise. Temperature × humidity and temperature × wind speed interactions account for 28.9% and 14.4% respectively, demonstrating synergistic patterns of meteorological factors.

TCN-Attention achieved high accuracy in daily prediction for stations within the current monitoring network (, MAE = 167.5 ions/cm³), with MAE 2.7%–7.8% lower than those of the conventional machine learning methods. MTFT maintained an overall across the 7-day forecast horizon. Moreover, the advantage of the deep learning models over conventional methods widened as the forecast horizon extended, from a modest margin in the 1-day task to a relative R² improvement of 18.6%–29.8% in the 7-day task, indicating that sequence-based temporal modeling becomes increasingly valuable for longer-range NAI prediction.

Prediction difficulty exhibits a systematic seasonal structure. Forecast errors of the 7-day model were lower in summer and autumn (seasonal MAE of 63.9 and 69.3 ions/cm³) than in winter and spring (83.1 and 85.3 ions/cm³). Because summer combines the highest concentrations with the lowest forecast errors, this pattern suggests that seasonal predictability depends more on the temporal stability of the meteorology–NAI relationship than on the absolute concentration level, with the more frequent synoptic transitions in the colder seasons weakening the day-to-day persistence of NAI concentration.

Models possess good robustness under the four representative meteorological scenarios. The scenario MAE of the 7-day model ranged from 59.5 to 82.1 ions/cm³, of the same magnitude as the overall test error (MAE = 73.4 ions/cm³), with no pronounced deterioration in any single scenario.

Meteorological effects obtained through SHAP analysis are highly consistent with physical mechanisms of NAI generation. Temperature effect has inflection point at 25 °C, wind speed 1.5 m/s is critical value regulating balance between diffusion and accumulation, and combination of high temperature (>25 °C) and moderate humidity (70%–85% RH) forms optimal conditions for NAI generation.

Currently, the dual-model framework of “daily prediction by TCN-Attention + 7-day sliding prediction by MTFT” constructed in this study has been deployed in Jiangsu Province meteorological operational system, providing technical support for NAI resource assessment, forest health promotion activity planning, and ecotourism development.

References

• 1

  BaiS.KolterJ. Z.KoltunV. (2018). An empirical evaluation of generic convolutional and recurrent networks for sequence modeling. arXiv preprint arXiv:1803.01271.

  • Google Scholar
  • View reference in article

• 2

  Ben TaiebS.SorjamaaA.BontempiG. (2010). Multiple-output modeling for multi-step-ahead time series forecasting. Neurocomputing73 (10–12), 1950–1957. 10.1016/j.neucom.2009.11.030

  • CrossRef
  • Google Scholar
  • View reference in article

• 3

  BergmeirC.BenítezJ. M. (2012). On the use of cross-validation for time series predictor evaluation. Inf. Sci.191, 192–213. 10.1016/j.ins.2011.12.028

  • CrossRef
  • Google Scholar
  • View reference in article

• 4

  BernackiJ.SchererR. (2025). A comprehensive review of data-driven techniques for air pollution concentration forecasting. Sensors25 (19), 6044. 10.3390/s25196044

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 5

  BowersB.FloryR.AmetepeJ.StaleyL.PatrickA.CarringtonH. (2018). Controlled trial evaluation of exposure duration to negative air ions for the treatment of seasonal affective disorder. Psychiatry Res.259, 7–14. 10.1016/j.psychres.2017.08.040

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 6

  BreimanL. (2001). Random forests. Mach. Learn.45 (1), 5–32. 10.1023/A:1010933404324

  • CrossRef
  • Google Scholar
  • View reference in article

• 7

  CaiL.ShiG.-Y.ZhangJ.DuL.NiX.HuY.et al (2024). Empirical analysis of the influence mechanism of vegetation and environment on negative air ion in warm temperate forest ecosystems. Environ. Pollut.363, 125273. 10.1016/j.envpol.2024.125273

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 8

  CaoW.QiW.LuP. (2024). Air quality prediction based on time series decomposition and convolutional sparse self-attention mechanism transformer model. IEEE Access12, 155340–155350. 10.1109/ACCESS.2024.3484579

  • CrossRef
  • Google Scholar
  • View reference in article

• 9

  ChaiT.DraxlerR. R. (2014). Root mean square error (RMSE) or mean absolute error (MAE)? Arguments against avoiding RMSE in the literature. Geosci. Model Dev.7 (3), 1247–1250. 10.5194/gmd-7-1247-2014

  • CrossRef
  • Google Scholar
  • View reference in article

• 10

  ChenT.GuestrinC. (2016). “XGBoost: a scalable tree boosting system,” in Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 785–794. 10.1145/2939672.2939785

  • CrossRef
  • Google Scholar
  • View reference in article

• 11

  ChenZ.ChenD.ZhaoC.KwanM. P.CaiJ.ZhuangY.et al (2020). Influence of meteorological conditions on PM2.5 concentrations across China: a review of methodology and mechanism. Environ. Int.139, 105558. 10.1016/j.envint.2020.105558

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 12

  ChenQ.WangR.ZhangX.LiuJ.WangD. (2021). Effects of different site conditions on the concentration of negative air ions in Mountain forest based on an orthogonal experimental study. Sustainability13 (21), 12012. 10.3390/su132112012

  • CrossRef
  • Google Scholar
  • View reference in article

• 13

  ChengP.WeiC.ZhangJ.WangH. (2025). Air quality index forecasting based on quadratic decomposition and Transformer-BiLSTM—A case study of beijing. Atmosphere16 (12), 1334. 10.3390/atmos16121334

  • CrossRef
  • Google Scholar
  • View reference in article

• 14

  DouX.HouS.ShenY.PengQ.LiuH.YiQ.et al (2025). Spatiotemporal distribution characteristics and influencing factors of negative air ion (NAI) concentrations in Yunnan. Aerosol Air Qual. Res.25, 49. 10.1007/s44408-025-00054-6

  • CrossRef
  • Google Scholar

• 15

  DragomiretskiyK.ZossoD. (2014). Variational mode decomposition. IEEE Trans. Signal Process.62 (3), 531–544. 10.1109/TSP.2013.2288675

  • CrossRef
  • Google Scholar

• 16

  DuS.LiT.YangY.HorngS.-J. (2020). Multivariate time series forecasting via attention-based encoder–decoder framework. Neurocomputing388, 269–279. 10.1016/j.neucom.2019.12.118

  • CrossRef
  • Google Scholar

• 17

  GilikA.OgrenciA. S.OzmenA. (2022). Air quality prediction using CNN+LSTM-Based hybrid deep learning architecture. Environ. Sci. Pollut. Res.29 (8), 11920–11938. 10.1007/s11356-021-16227-w

  • CrossRef
  • Google Scholar
  • View reference in article

• 18

  GoldsteinN.ArshavskayaT. V. (1997). Is atmospheric superoxide vitally necessary? Accelerated death of animals in a quasi-neutral electric atmosphere. Z. Naturforsch. C J. Biosci.52 (5-6), 396–404. 10.1515/znc-1997-5-619

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 19

  GuoQ.HeZ.WangZ. (2023). Prediction of monthly average and extreme atmospheric temperatures in zhengzhou based on artificial neural network and deep learning models. Front. For. Glob. Change.6, 1249300. 10.3389/ffgc.2023.1249300

  • CrossRef
  • Google Scholar
  • View reference in article

• 20

  GuoQ.HeZ.WangZ. (2024a). Monthly climate prediction using deep convolutional neural network and long short-term memory. Sci. Rep.14, 17748. 10.1038/s41598-024-68906-6

  • CrossRef
  • Google Scholar
  • View reference in article

• 21

  GuoQ.HeZ.WangZ.QiaoS.ZhuJ.ChenJ. (2024b). A performance comparison study on climate prediction in weifang city using different deep learning models. Water16 (19), 2870. 10.3390/w16192870

  • CrossRef
  • Google Scholar
  • View reference in article

• 22

  GuoQ.HeZ.WangZ. (2025). Assessing the effectiveness of long short-term memory and artificial neural network in predicting daily ozone concentrations in liaocheng city. Sci. Rep.15, 6798. 10.1038/s41598-025-91329-w

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 23

  HeZ.GuoQ. (2024). Comparative analysis of multiple deep learning models for forecasting monthly ambient PM2.5 concentrations: a case study in dezhou city, China. Atmosphere15 (12), 1432. 10.3390/atmos15121432

  • CrossRef
  • Google Scholar
  • View reference in article

• 24

  HeZ.GuoQ.WangZ.LiX. (2025). A hybrid wavelet-based deep learning model for accurate prediction of daily surface PM2.5 concentrations in guangzhou city. Toxics13 (4), 254. 10.3390/toxics13040254

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 25

  HeZ.GuoQ.ZhangZ.FengG.QiaoS.WangZ. (2026). Forecasting daily ambient PM2.5 concentrations in qingdao city using deep learning and hybrid interpretable models and analysis of driving factors using SHAP. Toxics14 (1), 44. 10.3390/toxics14010044

  • CrossRef
  • Google Scholar
  • View reference in article

• 26

  HoudouA.El BadisyI.KhomsiK.AbdalaS. A.AbdullaF.NajmiH.et al (2024). Interpretable machine learning approaches for forecasting and predicting air pollution: a systematic review. Aerosol Air Qual. Res.24, 230151. 10.4209/aaqr.230151

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 27

  HuangN. E.ShenZ.LongS. R.WuM. C.ShihH. H.ZhengQ.et al (1998). The empirical mode decomposition and the hilbert spectrum for nonlinear and non-stationary time series analysis. Proc. R. Soc. Lond. A454 (1971), 903–995. 10.1098/rspa.1998.0193

  • CrossRef
  • Google Scholar
  • View reference in article

• 28

  JiangS. Y.MaA.RamachandranS. (2018). Negative air ions and their effects on human health and air quality improvement. Int. J. Mol. Sci.19 (10), 2966. 10.3390/ijms19102966

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 29

  KangH.ZhuB.GaoJ.HeY.WangH.SuJ.et al (2019). Potential impacts of cold frontal passage on air quality over the yangtze river Delta, China. Atmos. Chem. Phys.19, 3673–3685. 10.5194/acp-19-3673-2019

  • CrossRef
  • Google Scholar
  • View reference in article

• 30

  KeG.MengQ.FinleyT.WangT.ChenW.MaW.et al (2017). LightGBM: a highly efficient gradient boosting decision tree. Adv. Neural Inf. Process. Syst.30, 3149–3157.

  • Google Scholar
  • View reference in article

• 31

  LiC.XieZ.ChenB.KuangK.XuD.LiuJ.et al (2021). Different time scale distribution of negative air ions concentrations in mount wuyi national park. Int. J. Environ. Res. Public Health18 (9), 5037. 10.3390/ijerph18095037

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 32

  LiA.LiQ.ZhouB.GeX.CaoY. (2022). Temporal dynamics of negative air ion concentration and its relationship with environmental factors: results from long-term on-site monitoring. Sci. Total Environ.832, 155057. 10.1016/j.scitotenv.2022.155057

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 33

  LiA.LiQ.YangY.HuY.XiaoS.LiZ.et al (2024). Stand structure and environment jointly determine negative air ion concentrations in forests: evidence from concurrent on-site monitoring in four typical subtropical forests during the growing season. Environ. Exp. Bot.220, 105684. 10.1016/j.envexpbot.2024.105684

  • CrossRef
  • Google Scholar
  • View reference in article

• 34

  LiC.ZongC.ChenB.YangX.HuangM.HongY.et al (2025). Temporal dynamics and relationship between negative air ions and environmental factors in subtropical forests, China. Sci. Rep.15, 12228. 10.1038/s41598-025-96762-5

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 35

  LiangZ.ZengZ.KadaM.LuoY.KaufholdL.XuL.et al (2024). Seamless hourly estimation of negative air ion concentrations: integrating hybrid stacked machine learning models with kriging spatiotemporal augmentation. Geophys. Res. Lett.51, e2024GL109870. 10.1029/2024GL109870

  • CrossRef
  • Google Scholar
  • View reference in article

• 36

  LimB.ArıkS. Ö.LoeffN.PfisterT. (2021). Temporal fusion transformers for interpretable multi-horizon time series forecasting. Int. J. Forecast.37 (4), 1748–1764. 10.1016/j.ijforecast.2021.03.012

  • CrossRef
  • Google Scholar
  • View reference in article

• 37

  LingX.JayaratneR.MorawskaL. (2010). Air ion concentrations in various urban outdoor environments. Atmos. Environ.44 (18), 2186–2193. 10.1016/j.atmosenv.2010.03.026

  • CrossRef
  • Google Scholar
  • View reference in article

• 38

  LiuB. C.BinaykiaA.ChangP. C.TiwariM. K.TsaoC. C. (2017). Urban air quality forecasting based on multi-dimensional collaborative support vector regression (SVR): a case study of beijing-tianjin-shijiazhuang. PLoS ONE12 (7), e0179763. 10.1371/journal.pone.0179763

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 39

  LundbergS. M.LeeS. I. (2017). A unified approach to interpreting model predictions. Adv. Neural Inf. Process. Syst.30, 4765–4774.

  • Google Scholar
  • View reference in article

• 40

  LuoL.SunW.HanY.ZhangW.LiuC.YinS. (2020). Importance evaluation based on random forest algorithms: insights into the relationship between negative air ions variability and environmental factors in urban green spaces. Atmosphere11 (7), 706. 10.3390/atmos11070706

  • CrossRef
  • Google Scholar
  • View reference in article

• 41

  MaZ.WangB.LuoW.JiangJ.LiuD.WeiH.et al (2024). Air pollutant prediction model based on transfer learning two-stage attention mechanism. Sci. Rep.14, 7385. 10.1038/s41598-024-57784-7

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 42

  MaruyamaT.KatayamaT. (2003). Transfer of small negative atmospheric ions. J. Appl. Phys.94 (11), 7365–7369. 10.1063/1.1625438

  • CrossRef
  • Google Scholar
  • View reference in article

• 43

  MiaoS.ZhangX.HanY.SunW.LiuC.YinS. (2018). Random forest algorithm for the relationship between negative air ions and environmental factors in an urban park. Atmosphere9 (12), 463. 10.3390/atmos9120463

  • CrossRef
  • Google Scholar
  • View reference in article

• 44

  MiaoY.LiJ.MiaoS.CheH.WangY.ZhangX.et al (2019). Interaction between planetary boundary layer and PM2.5 pollution in megacities in China: a review. Curr. Pollut. Rep.5 (4), 261–271. 10.1007/s40726-019-00124-5

  • CrossRef
  • Google Scholar

• 45

  MohapatraE.DasM.RathS. (2025). Deep learning-based AQI forecasting: a CNN-LSTM model with visual insights from SHAP-LIME and PDP. Discov. Appl. Sci.7, 1326. 10.1007/s42452-025-07845-x

  • CrossRef
  • Google Scholar
  • View reference in article

• 46

  MuthukumarP.NagrechaK.ComerD.CalvertC. F.AminiN.HolmJ.et al (2022). PM2.5 air pollution prediction through deep learning using multisource meteorological, wildfire, and heat data. Atmosphere13 (5), 822. 10.3390/atmos13050822

  • CrossRef
  • Google Scholar
  • View reference in article

• 47

  NadaliA.ArfaeiniaH.AsadgolZ.FahiminiaM. (2020). Indoor and outdoor concentration of PM10, PM2.5 and PM1 in residential building and evaluation of negative air ions (NAIs) in indoor PM removal. Environ. Pollut. Bioavailab.32 (1), 47–55. 10.1080/26395940.2020.1728198

  • CrossRef
  • Google Scholar
  • View reference in article

• 48

  OkajimaS.NakamuraH.Kuwano-YoshidaA.ParfittR. (2025). Mechanisms for an early spring peak of extratropical cyclone activity in east Asia. J. Clim.38 (9), 1981–1997. 10.1175/JCLI-D-24-0203.1

  • CrossRef
  • Google Scholar
  • View reference in article

• 49

  ShabekF.KolaržP.ČelikovićI.ĆurčićM.JaničijevićA. (2025). Interaction between radon, air ions, and ultrafine particles under contrasting atmospheric conditions in belgrade. Serbia. Atmos.16 (7), 808. 10.3390/atmos16070808

  • CrossRef
  • Google Scholar

• 50

  ShiG.ZhouY.SangY.HuangH.ZhangJ.MengP.et al (2021). Modeling the response of negative air ions to environmental factors using multiple linear regression and random forest. Ecol. Inf.66, 101464. 10.1016/j.ecoinf.2021.101464

  • CrossRef
  • Google Scholar

• 51

  SlingoJ.PalmerT. (2011). Uncertainty in weather and climate prediction. Philos. Trans. R. Soc. A Math. Phys. Eng. Sci.369 (1956), 4751–4767. 10.1098/rsta.2011.0161

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 52

  TammetH. (2006). Continuous scanning of the mobility and size distribution of charged clusters and nanometer particles in atmospheric air and the balanced scanning mobility analyzer BSMA. Atmos. Res.82 (3–4), 523–535. 10.1016/j.atmosres.2006.02.009

  • CrossRef
  • Google Scholar

• 53

  TaoS.SunZ.LinX.ZhangZ.WuC.ZhangZ.et al (2023). Negative air ion (NAI) dynamics over Zhejiang Province, China, based on multivariate remote sensing products. Remote Sens.15 (3), 738. 10.3390/rs15030738

  • CrossRef
  • Google Scholar

• 54

  TaoC.WangY.WangY.ZhengZ.ZhangH. (2025). Deep learning for air pollutant forecasting: opportunities, challenges, and future directions. Front. Environ. Sci. Eng.19 (12), 172. 10.1007/s11783-025-2092-6

  • CrossRef
  • Google Scholar

• 55

  TianH.KongH.WongC. (2024). A novel stacking ensemble learning approach for predicting PM2.5 levels in dense urban environments using meteorological variables: a case study in Macau. Appl. Sci.14 (12), 5062. 10.3390/app14125062

  • CrossRef
  • Google Scholar

• 56

  TsokovS.LazarovaM.Aleksieva-PetrovaA. (2022). A hybrid spatiotemporal deep model based on CNN and LSTM for air pollution prediction. Sustainability14 (9), 5104. 10.3390/su14095104

  • CrossRef
  • Google Scholar

• 57

  VaswaniA.ShazeerN.ParmarN.UszkoreitJ.JonesL.GomezA. N.et al (2017). Attention is all you need. Adv. Neural Inf. Process. Syst.30, 5998–6008.

  • Google Scholar

• 58

  WanX.ZhouR.LiL.YangC.LianJ.ZhangJ.et al (2024). Factors influencing the concentration of negative air ions in urban forests of the zhuyu Bay scenic area in yangzhou, China. Atmosphere15 (3), 316. 10.3390/atmos15030316

  • CrossRef
  • Google Scholar

• 59

  WangH.WangB.NiuX.SongQ.LiM.LuoY.et al (2020). Study on the change of negative air ion concentration and its influencing factors at different spatio-temporal scales. Glob. Ecol. Conserv.23, e01008. 10.1016/j.gecco.2020.e01008

  • CrossRef
  • Google Scholar

• 60

  WangY.NiZ.WuD.FanC.LuJ.XiaB. (2020). Factors influencing the concentration of negative air ions during the year in forests and urban green spaces of the dapeng peninsula in shenzhen, China. J. For. Res.31 (6), 2537–2547. 10.1007/s11676-019-01047-z

  • CrossRef
  • Google Scholar

• 61

  WangX.WangK.MaoY.WuG. (2025). Day-to-day temperature variability in meteorological observations and reanalysis data over China. Earth Space Sci.12, e2025EA004573. 10.1029/2025EA004573

  • CrossRef
  • Google Scholar

• 62

  WeiQ.ChenY.ZhangH.JiaZ.YangJ.NiuB. (2025). Simulation and prediction of PM2.5 concentrations and analysis of driving factors using interpretable tree-based models in shanghai, China. Environ. Res.270, 121003. 10.1016/j.envres.2025.121003

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 63

  WenC.LiuS.YaoX.PengL.LiX.HuY.et al (2019). A novel spatiotemporal convolutional long short-term neural network for air pollution prediction. Sci. Total Environ.654, 1091–1099. 10.1016/j.scitotenv.2018.11.086

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 64

  WuC. F.LaiC. H.ChuH. J.LinW. H. (2011). Evaluating and mapping of spatial air ion quality patterns in a residential garden using a geostatistic method. Int. J. Environ. Res. Public Health8 (6), 2304–2319. 10.3390/ijerph8062304

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 65

  XiaoF.YangM.FanH.FanG.Al-QanessM. A. A. (2020). An improved deep learning model for predicting daily PM2.5 concentration. Sci. Rep.10, 20988. 10.1038/s41598-020-77757-w

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 66

  XieZ.LiC.LinY.LiuJ.HeZ. (2023). The lagging effect of precipitation on NAIs concentrations on rainy days in wuyi Mountain national park, China. Atmosphere14 (2), 377. 10.3390/atmos14020377

  • CrossRef
  • Google Scholar

• 67

  YangZ.YangD.DyerC.HeX.SmolaA.HovyE. (2016). “Hierarchical attention networks for document classification,” in Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, 1480–1489. 10.18653/v1/N16-1174

  • CrossRef
  • Google Scholar

• 68

  YunJ.YaoW.WangX.ChenF.WangS. (2024). Daily dynamics of forest air negative ion concentration in spring and the relationship of influencing factors: results of field monitoring. Air Qual. Atmos. Health17 (3), 501–511. 10.1007/s11869-023-01458-2

  • CrossRef
  • Google Scholar

• 69

  ZhanH.LiZ.CuiH. (2026). Spatial and diurnal variations of negative air ions concentrations and their influencing factors in guancen Mountain. Front. Environ. Sci.14, 1719189. 10.3389/fenvs.2026.1719189

  • CrossRef
  • Google Scholar

• 70

  ZhangZ.ZhangS. (2023). Modeling air quality PM2.5 forecasting using deep sparse attention-based transformer networks. Int. J. Environ. Sci. Technol.20 (12), 13535–13550. 10.1007/s13762-023-04900-1

  • CrossRef
  • Google Scholar

• 71

  ZhangY.ZhangR.MaQ.WangY.WangQ.HuangZ.et al (2020). A feature selection and multi-model fusion-based approach of predicting air quality. ISA Trans.100, 210–220. 10.1016/j.isatra.2019.11.023

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 72

  ZhangZ.TaoS.ZhouB.ZhangX.ZhaoZ. (2021). Plant stomatal conductance determined transpiration and photosynthesis both contribute to the enhanced negative air ion (NAI). Ecol. Indic.130, 108114. 10.1016/j.ecolind.2021.108114

  • CrossRef
  • Google Scholar

• 73

  ZhangZ.ZhangS.ChenC.YuanJ. (2024). A systematic survey of air quality prediction based on deep learning. Alex. Eng. J.93, 128–141. 10.1016/j.aej.2024.03.031

  • CrossRef
  • Google Scholar

• 74

  ZhangT.QiX.LampilahtiJ.ChenL.ChiX.NieW.et al (2025). Differential characterization of air ions in boreal forest of Finland and a megacity of eastern China. Atmos. Chem. Phys.25, 10027–10048. 10.5194/acp-25-10027-2025

  • CrossRef
  • Google Scholar

• 75

  ZhouY.ChangF.-J.ChangL.-C.KaoI.-F.WangY.-S. (2019). Explore a deep learning multi-output neural network for regional multi-step-ahead air quality forecasts. J. Clean. Prod.209, 134–145. 10.1016/j.jclepro.2018.10.243

  • CrossRef
  • Google Scholar

• 76

  ZouH.HastieT. (2005). Regularization and variable selection via the elastic net. J. R. Stat. Soc. Ser. B Stat. Methodol.67 (2), 301–320. 10.1111/j.1467-9868.2005.00503.x

  • CrossRef
  • Google Scholar