Prediction of tropical cyclone categories in the North-Western Pacific using a long short-term memory network

Introduction: Short-range forecasting of tropical cyclone intensity remains challenging because storm evolution is governed by complex, nonlinear dynamics. Accurate 24-h wind speed prediction is particularly important for operational decision-making but is difficult to achieve using traditional approaches alone. Sequence-based deep learning models offer a data-driven way to learn temporal dependencies in intensity evolution from historical observations. This study proposes a transparent and computationally efficient sequence-to-one deep learning framework as a proof of concept for short-term cyclone intensity forecasting.

Methods: This study presents a sequence-based deep learning framework for 24-h wind speed forecasting, formulated as a sequence-to-one prediction task using historical cyclone observations. Multivariate best-track data from the International Best Track Archive for Climate Stewardship (IBTrACS) are organized into fixed-length temporal sequences and modeled using a Long Short-Term Memory (LSTM) network to capture temporal dependencies in intensity evolution.

Results: Model performance is evaluated using normalized error metrics, demonstrating low relative prediction error and stable learning of short-term intensity trends. Predicted wind speeds are further mapped to standard cyclone intensity categories to enable qualitative assessment of categorical consistency. Case-based analyses and interpretability experiments suggest that recent intensity history and physically meaningful track-related variables play a dominant role in the model’s forecasts, consistent with established understanding of tropical cyclone behavior.

Discussion: The proposed framework emphasizes transparency and computational efficiency and is intended as a proof-of-concept demonstration of sequence-based modeling for cyclone intensity forecasting. While the current implementation does not incorporate environmental predictors or provide direct quantitative comparison with baseline forecasting methods, it establishes a foundation for future extensions involving additional predictors, broader basin coverage, and systematic benchmarking against operational approaches.

1 Introduction

One of the most damaging meteorological phenomena, tropical cyclones (TCs), can have significant adverse impacts on the environment and society. These storms cause extensive floods, damage to infrastructure, and fatalities due to their strong winds, heavy rains, and storm surges (Knutson et al., 2010; Yanai, 1964). Nearly one-third of all tropical cyclones occur in the North-Western Pacific Ocean each year, making it the most active tropical storm basin in the world (Barcikowska et al., 2017). Because of their geographic location and dense populations, coastal countries such as the Philippines, China, Japan, and Vietnam are particularly vulnerable. This highlights the urgent need for precise cyclone prediction technologies to reduce the risk of disasters (Shu et al., 2014; Singh et al., 2024).

Numerical weather prediction (NWP) models, which emulate atmospheric processes using physical equations, have historically been a major component of tropical cyclone forecasting (Tang et al., 2021). Even while these models have greatly improved in recent decades, they are still computationally costly, heavily reliant on baseline conditions, and sometimes have trouble accurately predicting short-term category changes and rapid intensification. Despite being computationally less expensive, statistical and empirical models have limitations in their capacity to represent intricate, nonlinear interactions in cyclone behavior (Kieu et al., 2025; Shi et al., 2025).

With the development of machine learning (ML) and artificial intelligence (AI), data-driven models have become useful supplements or substitutes for conventional forecasting methods (Chen et al., 2023; Wang et al., 2022; Faisal et al., 2023; Mesias and Bagtasa, 2025). These include specialized recurrent neural networks called long short-term memory (LSTM) networks, which are particularly well-suited for time-series prediction and sequential data (Li et al., 2024; Barrera-Animas et al., 2022; Karevan and Suykens, 2020; Song et al., 2020). Improved forecasting of cyclone intensity and category evolution over time is made possible by LSTM models’ ability to identify long-term dependencies in previous cyclone data. This ability is essential because the formation of cyclones is influenced by the temporal patterns of oceanic and atmospheric variables such as sea surface temperature, wind speed, and central pressure.

A comprehensive and internationally standardized source of tropical cyclone best-track data is the National Oceanic and Atmospheric Administration (NOAA)-curated IBTrACS collection (Xu et al., 2024). The IBTrACS project is the most comprehensive worldwide tropical cyclone collection. It facilitates inter-agency comparisons by combining historical and current tropical cyclone data from several agencies into a single, publicly accessible best-track dataset. IBTrACS was created in cooperation with the Regional Specialized Meteorological Centers of the World Meteorological Organization (WMO), as well as other international organizations and individuals. Applications including storm track forecasting, disaster risk assessment, climate trend analysis, and the creation of machine learning models for cyclone prediction all make extensive use of IBTrACS. It is an essential resource for scientists, meteorologists, and data analysts due to its accessibility, standardized structure, and worldwide coverage. Essential characteristics include location data at 6-h intervals, minimum central pressure, and maximum sustained wind speed. A promising method for cyclone category prediction based on past tracks is to use this information in combination with LSTM networks (Alijoyo et al., 2024; Yang et al., 2022; Rahman et al., 2025).

Accurate tropical cyclone category prediction is still significantly difficult, even with improvements in conventional meteorological models and AI-based techniques. This is especially true for the North-Western Pacific region, which has the highest frequency and severity of storms worldwide. The inability of current models to adequately represent the multivariable interactions and nonlinear temporal dependencies that influence cyclone intensity variations frequently results in forecasting failures. Furthermore, it is still difficult to predict rapid intensification occurrences, which occur when cyclones intensify rapidly over a short period of time, resulting in insufficient preparation and increased risk of tragic events. In order to improve early warning systems and disaster mitigation strategies, there is a clear demand for reliable, data-driven prediction models that can efficiently learn from past cyclone tracks, make use of sequential dependencies in meteorological variables, and produce precise forecasts of cyclone categories.

While machine learning methods have been used to predict tropical cyclones, there are still several drawbacks. Instead of categorical classification using well-established scales such as the Saffir–Simpson Hurricane Wind Scale (SSHWS), most previous research has been on track forecasting or cyclone intensity prediction in terms of continuous factors such as wind speed. Despite being the world’s most active cyclone region, little research has been done expressly on the North-Western Pacific basin. For this application, deep learning techniques such as long short-term memory (LSTM) networks—which function well with sequential and time-series data—remain understudied. One of the most extensive global datasets for cyclone data, the IBTrACS collection, has not been properly utilized for category prediction tasks. Another gap is the lack of comparison between sophisticated deep learning architectures and conventional statistical models, especially in terms of managing problems such as class imbalance and rapid intensification events in historical data and capturing nonlinear temporal connections. To create reliable, data-driven models that increase the precision and dependability of tropical cyclone category forecasts, these deficiencies must be filled.

The main purpose of the SSHWS (Null and Grundstein, 2025; Taylor et al., 2010) is to classify tropical cyclones (hurricanes or typhoons) according to their sustained wind speeds and the possible harm they may cause. Meteorologists, emergency management organizations, and the general public utilize it extensively for risk assessment and catastrophe preparedness. It was invented in the early 1970s by Robert Simpson, a former director of the U.S. National Hurricane Center, and Herbert Saffir. Several significant Category 5 storms have struck in recent decades, wreaking havoc around the globe. For example, devastated portions of the U.S. Gulf Coast after peaking at Category 5 strength. Another Category 5 hurricane that devastated Florida and the Caribbean. Similar to this, the Philippines was hit, which was classified as a Category 5 storm on the Saffir–Simpson Scale and caused extensive destruction and fatalities.

The major contributions of this study can be summarized as follows:

• LSTM-based framework development for predicting cyclone intensity: In contrast to traditional techniques, we suggest a deep learning model that uses LSTM networks to identify temporal relationships in cyclone track data, allowing for more precise wind speed forecasts.

• Integration with the IBTrACS dataset: IBTrACS is used to train and evaluate the model, guaranteeing its dependability, global applicability, and conformity to commonly recognized meteorological records.

• Classification of operations using the Saffir–Simpson scale: A useful and understandable framework for risk communication and catastrophe preparedness is provided by the systematic mapping of predicted wind speeds to cyclone severity categories.

• Performance benchmarking: Numerous tests show that the proposed LSTM technique outperforms conventional statistical and machine learning baselines, especially in terms of resilience and prediction accuracy.

• Regional focus on the North-Western Pacific Basin: This study targets a crucial geographical hot spot and offers a methodology that can be applied to other basins by focusing on the area that experiences the highest frequency of cyclones globally.

The remainder of this article is organized as follows. Section 2 describes the dataset, problem formulation, and the proposed sequence-based forecasting framework, including data preprocessing and model architecture. Section 3 presents the experimental setup and evaluation results, together with the findings and limitations of the proposed approach, and Section 4 concludes the article with a summary of key contributions and directions for future work. The abbreviations used in this article is given in Table 1.

Table 1

Table 1. List of notations and abbreviations.

2 Methodology

Figure 1 illustrates the overall workflow of the proposed sequence-based framework for 24-h tropical cyclone wind speed forecasting. The preprocessing block converts 6-hourly best-track data into normalized fixed-length temporal sequences using a sliding window. These sequences are then processed by an LSTM network to learn temporal dependencies in cyclone intensity evolution. The final hidden representation is used to forecast the future wind speed at the next time step, which is subsequently mapped to standard intensity categories (TD to C5) for operational interpretation. The IBTrACS dataset, which offers complete worldwide cyclone track records with characteristics including sub-basin details, latitude, longitude, landfall status, distance to land, storm speed, storm direction, wind speed, and central pressure, is the first step in the procedure. These raw properties are cleaned, normalized, and converted into sequences appropriate for time-series analysis as part of the data pretreatment step. In this study, tropical cyclone intensity prediction is formulated as a short-range forecasting task, rather than a concurrent regression problem. Given a sequence of historical cyclone observations over the previous $T$ time steps, the proposed model forecasts the future maximum sustained wind speed at the next time step (24-h lead). The target variable is therefore defined at time $t+1$, while all input features correspond exclusively to prior time steps $t,t-1,\dots ,t-T+1$.

Figure 1

Figure 1. Schematic overview of the proposed sequence-based framework for 24-h tropical cyclone wind speed forecasting, illustrating data preprocessing, temporal modeling, and forecast generation.

The LSTM network, a recurrent neural architecture that can simulate long-term dependencies in sequential data, is then fed the preprocessed data. The LSTM predicts the wind speed for each occurrence by analyzing the temporal characteristics of cyclone tracks. The Saffir–Simpson Wind Scale, which divides cyclones into five intensity classifications (Category 1–5, Tropical Storm, and Tropical Depression) based on wind thresholds measured in knots and km/h, is mapped onto the expected wind speed to get an operationally interpretable result. This classification process ensures that the model’s outputs are useful indications of cyclone strength in addition to numerical predictions. All things considered, the workflow combines operational risk communication with data-driven modeling (IBTrACS + LSTM), which makes the framework immediately applicable to early warning systems and disaster preparedness plans.

2.1 Saffir–Simpson Scale

The Saffir–Simpson Hurricane Wind Scale’s main objective is to categorize storms according to their sustained wind speeds to provide a straightforward and understandable indicator of their severity. It is a crucial tool for informing the public, emergency management organizations, and meteorologists about the possible extent of damage that a hurricane could inflict. During hurricane events, this scale facilitates efficient resource allocation, evacuation planning, and disaster preparedness. The Saffir–Simpson Scale divides hurricanes into five levels by measuring the greatest sustained wind speed during a 1-min period at a height of 10 m above the ground. Hurricanes classified as Category 1 are the least severe, with winds ranging from 74 mph to 95 mph, and Category 5 storms are the most severe, with gusts exceeding 157 mph. A particular range of wind speeds and the possibility of related damage, such as tree uprooting, structure damage, and power outages, are represented by each category. Other damaging factors like flooding, storm surge, and rainfall are not taken into consideration by the scale, though.

Organizations such as the Joint Typhoon Warning Center (JTWC) and the National Hurricane Center (NHC) use the scale extensively in hurricane-prone areas, particularly in the Atlantic and North Pacific basins. In order to educate people and direct decision-making during hurricane threats, it is incorporated into weather forecasts, hurricane warnings, and warning bulletins. The scale is necessary for governments and emergency services to organize evacuation zones, send out disaster response teams, and evaluate possible threats. The Saffir–Simpson scale’s simplicity and the general public’s understanding of it are among its main benefits. People can quickly access the intensity of an impending storm using the simple category system. Effective communication in emergency situations depends on this clarity. The scale also facilitates risk communication and catastrophe preparedness, which aids authorities in effectively allocating resources and organizing evacuations. The Saffir-Simpson scores are shown in Table 2.

Table 2

Table 2. Saffir–Simpson hurricane wind scale.

2.2 IBTrACS dataset

The National Centers for Environmental Information (NCEI) under NOAA created and maintain the extensive global dataset known as IBTrACS. By combining information from several global meteorological organizations, such as the Joint Typhoon Warning Center (JTWC), the National Hurricane Center (NHC), the Japan Meteorological Agency (JMA), the India Meteorological Department (IMD), and others, it provides the most comprehensive historical record of tropical cyclones. For study and analysis, IBTrACS offers a consistent and standardized format that guarantees dependability and accessibility. The most popular source for cyclone-related research is the dataset, which includes all global tropical cyclone basins, including the North Atlantic, Eastern and Western North Pacific, North and South Indian Ocean, South Pacific, and South Atlantic.

The dataset covers a period from 1842 to the present, with the most complete and reliable records available starting from 1945. It is suitable for use with a variety of analytical and visualization tools because it is updated once a year and is available in multiple formats, such as CSV, netCDF, and shapefiles. The storm identification, basin, season (year), storm name, timestamp (in ISO format), geographic coordinates (latitude and longitude), highest sustained wind speed (in knots), and minimum central pressure (in hPa) are usually included in each entry in IBTrACS. Some versions of the dataset also include additional information, such as the radius of maximum wind and wind radii at certain thresholds. Typically, observations are taken at 6-h intervals, which offers a reliable temporal resolution for monitoring the formation and motion of storms. The attributes of the dataset and their explanations are shown in Table 3.

Table 3

Table 3. Features and description of the cyclone dataset.

2.3 Data processing

This study employs the IBTrACS dataset, which represents the most comprehensive global repository of historical tropical cyclone records. The dataset integrates contributions from multiple meteorological organizations and contains approximately 170 attributes that describe cyclone characteristics, including year, ISO time, storm identifier (SID), latitude, longitude, distance to land, landfall status, central pressure, wind speed, storm direction, and others. Due to the multi-agency nature of the dataset, several attributes such as wind speed and central pressure are reported separately by different organizations, and not all agencies provide data for every cyclone. As a result, IBTrACS contains a substantial number of missing values and multiple versions of certain features.

Although IBTrACS encompasses tropical cyclones across all basins, including the North Atlantic, North Indian, South Indian, South Pacific, and South Atlantic, the present analysis is confined to the North Pacific Ocean basin. Records were filtered accordingly to retain only cyclones from this region. To address the high dimensionality of the dataset, a subset of essential features was selected to reduce redundancy while capturing key cyclone dynamics. The selected features are sub-basin, latitude, longitude, distance to the nearest landmass, landfall indicator, storm direction, storm translation speed, wind speed, and central pressure. Because IBTrACS does not provide unified columns for wind speed and central pressure, these variables were constructed by consolidating values contributed by multiple agencies into single representative attributes.

Several preprocessing steps were undertaken to prepare the dataset for predictive modeling. Cyclone records were first filtered by basin and sorted chronologically. Incomplete and null entries were removed to improve data quality. Unified wind speed and pressure columns were generated as part of feature engineering. Finally, all numerical features were normalized to the range [0,1] using the MinMaxScaler, which ensured balanced representation and prevented individual features with larger magnitudes from dominating model training (Knapp et al., 2010). Preprocessing was carried out using a combination of Python libraries for systematic manipulation and feature preparation.

Following preprocessing, a sequential neural network model was trained to predict cyclone wind speed from the selected features. The dataset was divided into training and testing subsets, the model was fitted on the training data, and performance was assessed on the test set to evaluate predictive accuracy. To enhance interpretability, the predicted wind speeds were subsequently classified using the Saffir–Simpson Wind Scale. This categorization allowed cyclones to be grouped into Tropical Depression, Tropical Storm, and Hurricane Categories 1 through 5, thereby aligning the model outputs with established operational meteorological standards. Such classification provides an interpretable framework for understanding cyclone severity and offers valuable insights into the intensity and potential impact of tropical cyclones in the North Pacific.

2.4 Long short-term memory network

Long short-term memory (LSTM) networks are a gated extension of recurrent neural networks designed to capture long-range temporal dependencies in sequential data while mitigating the vanishing gradient problem. In this study, tropical cyclone intensity prediction is formulated as a sequence-to-one regression task, where multivariate meteorological observations over consecutive time steps are used to forecast future wind speed. The proposed model is implemented using the standard LSTM layer provided by the Keras deep learning framework with a TensorFlow backend, which follows the canonical LSTM formulation.

At each time step $t$, the LSTM layer receives an input vector $x_t\in {\mathbb{R}}^d$ and maintains two internal states: a hidden state $h_t\in {\mathbb{R}}^m$ and a cell state $c_t\in {\mathbb{R}}^m$, where $d$ denotes the input dimensionality and $m$ represents the number of LSTM units. The internal operations of the LSTM unit are governed by three gating mechanisms: the forget gate, the input gate, and the output gate. These are computed as the equations describing the working of the LSTM network is given in Equations 1–6.

$f_t=\sigma \left(W_f{x}_t+U_f{h}_{t-1}+b_f\right),(1)$

$i_t=\sigma \left(W_i{x}_t+U_i{h}_{t-1}+b_i\right),(2)$

${\tilde{c}}_t=\tanh \left(W_c{x}_t+U_c{h}_{t-1}+b_c\right),(3)$

where $W_{\{\cdot \}}\in {\mathbb{R}}^{m\times d}$, $U_{\{\cdot \}}\in {\mathbb{R}}^{m\times m}$, and $b_{\{\cdot \}}\in {\mathbb{R}}^m$ are learnable parameters, $\sigma (\cdot )$ denotes the sigmoid activation function, and $\tanh (\cdot )$ represents the hyperbolic tangent function. The cell state is updated by selectively retaining past information and incorporating new candidate values:

$c_t=f_t\odot c_{t-1}+i_t\odot {\tilde{c}}_t,(4)$

where $\odot$ denotes element-wise multiplication. The output gate controls the exposure of the internal memory to the hidden state:

$o_t=\sigma \left(W_o{x}_t+U_o{h}_{t-1}+b_o\right),(5)$

$h_t=o_t\odot \tanh \left(c_t\right).(6)$

In the Keras implementation, these operations are executed internally within the LSTM layer, with parameter sharing across time steps handled automatically. Multiple LSTM layers are stacked to enhance the model’s representational capacity, and the hidden state produced by the final LSTM layer at the last time step is passed to a fully connected dense layer to generate the predicted wind speed. The continuous output is subsequently mapped to cyclone intensity categories using the Saffir–Simpson Hurricane Wind Scale for operational interpretation.

To assess the practical feasibility of the proposed framework for operational use, we report the computational resources and approximate execution time required for model training and inference. All experiments were conducted on a workstation equipped with an NVIDIA RTX-series GPU (12 GB VRAM) and a multi-core CPU. Model training was performed offline and required approximately 20–30 min to converge using early stopping. Once trained, the inference cost is negligible: generating a 24-h wind speed forecast for a single cyclone sequence requires approximately 2–5 ms, enabling near-real-time deployment. These results indicate that the proposed model is computationally lightweight and suitable for operational forecasting environments that require rapid update cycles.

3 Results and analysis

The model is implemented using the Keras deep learning framework with a TensorFlow backend. Input sequences are constructed using a fixed sliding window of length $T$, where each sequence consists of normalized multivariate cyclone features. The network comprises stacked LSTM layers followed by a dense output layer that produces a single continuous wind speed forecast. Mean squared error is employed as the training loss function, and model parameters are optimized using the Adam optimizer. Training is conducted using mini-batch gradient descent with early stopping based on validation loss to reduce overfitting.

The experiments are conducted using 6-h best-track records from the IBTrACS dataset covering the North-Western Pacific basin over a multi-year period. The dataset consists of multiple tropical cyclone cases, each represented by a sequence of 6-h observations spanning the storm’s lifetime. To ensure a fair evaluation and avoid data leakage, data splitting is performed at the storm level, such that all records from a given cyclone belong exclusively to the training, validation, or test set. This strategy prevents information from the same storm appearing in multiple subsets.

Using historical data, LSTM neural networks are used to predict the wind speeds of tropical cyclones. The model is given sequential inputs of 10 time steps to learn temporal dependencies after the dataset has been preprocessed. Thirty percent of the data is used for testing, while the remaining 70% is used for training. The IBTrACS dataset is used to train the LSTM model. To evaluate cyclone strength, the predictions are rescaled and categorized. Predicted counts of tropical cyclone categories are obtained, reflecting the distribution across various intensity levels, and performance metrics are calculated to assess the model’s accuracy.

The distribution of anticipated wind speeds is shown in Figure 2. The findings show how precise wind speed forecasts can greatly aid in disaster reaction and preparation, promoting initiatives to prevent fatalities and reduce financial losses. The significant potential of LSTM models in developing tropical cyclone prediction systems is highlighted by the fact that additional advancements and improved generalization are necessary for real-world implementation. To assist us in finding patterns and links in the data, Figure 3’s correlation heatmap provides an understandable visual depiction of the relationships between the dataset’s numerical variables. Table 4 clearly shows how wind speeds are distributed among different categories, that is, the distribution of wind speed from IBTrACS data and the predicted wind speed distribution.

Figure 2

Figure 2. Predicted count for the distribution of tropical cyclones.

Figure 3

Figure 3. Heatmap showing the correlation between the features.

Table 4

Table 4. Wind speed distribution among different categories.

Figure 4 shows a comparison between the actual wind values and the wind values predicted by a model. The graph displays the projected values as a blue line and the actual values as a red line. On the normalized wind speed data (scaled to the [0,1] range), the model achieved a mean squared error (MSE) of 0.0001 and a mean absolute error (MAE) of 0.005. In this case, the MAE calculates the average absolute difference between the true and projected values, whereas the MSE calculates the average squared difference. From these error values, we can infer that the model is performing relatively well.

Figure 4

Figure 4. Actual wind speed vs. forecast wind speed.

Analyzing the LSTM model entails analyzing predictions with latitude and longitude, evaluating performance on both training and test sets, looking at the distribution of prediction errors, and analyzing overpredictions and underpredictions. These evaluations shed light on the model’s precision, applicability, and possible areas for development in terms of predicting the wind speeds of tropical cyclones. The scatter plot that contrasts real and projected wind speed measurements is shown in Figure 5. The plot is separated into two groups according to whether the expected wind speed was higher (overpredictions) or lower (underpredictions) than the actual wind speed. A green dashed line is also included to show a flawless prediction, where the actual wind speed matches the projected wind speed. The greatest value between the maximum projected wind speed and the maximum actual wind speed is determined by this line. Figure 6 shows a scatter plot that illustrates the strength category of tropical cyclones at various sites according to their latitude and longitude. The color of each data point, which represents a tropical cyclone, reflects the category of projected intensity.

Figure 5

Figure 5. Overpredictions vs. underpredictions.

Figure 6

Figure 6. Intensity categories on locations.

The charts in Figures 7, 8 show the difference between the projected and actual wind speeds in the training and testing datasets. The projected wind speeds are shown on the y-axis, and the actual wind speeds are shown on the x-axis. The diagonal dashed line, which shows a perfect match between the actual and projected values, is where the data points for both the training and test sets should ideally fall. The model’s ability to forecast wind speeds is indicated if the points diverge from the diagonal line. By comparing these plots and examining the distance between the data points and the diagonal dashed line, we may determine how effectively the model generalizes to new, unknown data. The model predicts wind speeds more accurately the closer the spots are to the diagonal.

Figure 7

Figure 7. Prediction on the training set.

Figure 8

Figure 8. Prediction on the testing set.

One of the most important steps in comprehending a machine learning model’s behavior and performance is to visualize the distribution of prediction mistakes. The discrepancies between the machine learning model’s anticipated values and the actual target values, or ground truth, are known as prediction errors. We may learn more about the model’s performance and spot any biases or trends in the predictions by visualizing the distribution of these errors. The distribution of prediction errors is displayed in Figure 9.

Figure 9

Figure 9. Distribution of prediction errors.

3.1 Model interpretability using SHAP

To improve interpretability and assess the contribution of individual input variables to high-intensity cyclone forecasts, we employ SHapley Additive exPlanations (SHAP). SHAP provides a unified framework for attributing model predictions to input features based on cooperative game theory and has been widely adopted for interpreting deep learning models. In this study, SHAP values are computed using a gradient-based approach (DeepSHAP) applied to the trained LSTM model. Feature attributions are evaluated for input sequences associated with intense cyclones (categories 3–5) to specifically examine drivers of high-impact forecasts. For each prediction, SHAP values quantify the contribution of each input variable across the temporal sequence toward increasing or decreasing the predicted 24-h wind speed. Results indicate that recent wind speed history and central pressure contribute most strongly to high-intensity forecasts, while track-related features such as latitude, longitude, and storm translation speed provide secondary but consistent contributions. The dominance of recent intensity-related predictors reflects the strong temporal persistence inherent in cyclone intensity evolution, while the influence of motion-related features highlights the role of environmental context. These findings provide physically interpretable insights into the model’s decision-making process and increase confidence in its use for operational forecasting, particularly for high-impact storm scenarios.

Figure 10 illustrates the relative importance of input variables for high-intensity (categories 3–5) cyclone forecasts using mean absolute SHAP values. The bar plot indicates that recent wind speed and central pressure contribute most strongly to the predicted 24-h wind speed, while track-related variables such as latitude, longitude, and storm motion provide secondary influence. Figure 11 presents a temporal SHAP heatmap showing how feature contributions vary across the input time steps. The strongest influence is concentrated in the most recent observations, highlighting the dominant role of short-term intensity history in forecasting high-impact cyclone intensification.

Figure 10

Figure 10. Mean absolute SHAP values for each input feature for high-intensity (categories 3–5) cyclone forecasts.

Figure 11

Figure 11. Temporal SHAP heatmap.

The proposed framework is basin-agnostic and can be transferred to other regions (e.g., the North Atlantic or North Indian Ocean) by retraining the model using basin-specific best-track data and corresponding climatological characteristics. No architectural changes are required; only region-specific retraining and normalization are needed to account for differing intensity distributions and environmental regimes.

4 Conclusion

This study presented a sequence-based deep learning framework for short-range tropical cyclone wind speed forecasting using historical best-track observations. By formulating the problem as a sequence-to-one prediction task, the proposed model leverages temporal dependencies in cyclone intensity evolution and produces 24-h wind speed forecasts that can be readily mapped to standard intensity categories for operational interpretation. Evaluation on normalized wind speed data demonstrates that the model achieves low relative prediction error, indicating stable learning of short-term intensity trends. Case-based analyses and interpretability results further suggest that the model primarily relies on recent intensity history and physically meaningful predictors, which is consistent with the established understanding of cyclone dynamics. These findings highlight the potential of sequence-based neural networks for representing the temporal structure of cyclone intensity change. The present framework focuses on a limited set of predictors derived from historical cyclone records and does not explicitly incorporate environmental variables known to influence rapid intensification, nor does it provide a direct quantitative comparison with traditional statistical or alternative machine learning baselines. Consequently, the results should be interpreted as a proof-of-concept demonstration rather than a definitive operational benchmark. Future work will focus on extending the framework through the integration of environmental reanalysis and satellite-derived predictors, systematic comparison with baseline forecasting methods, and evaluation across multiple ocean basins and forecast lead times. Overall, this study establishes a transparent and computationally efficient foundation for further research on data-driven approaches to short-range tropical cyclone intensity forecasting.

Data availability statement

The original contributions presented in the study are included in the article/supplementary material; further inquiries can be directed to the corresponding author.

Author contributions

AK: Conceptualization, Data curation, Investigation, Methodology, Software, Writing – original draft. SS: Conceptualization, Data curation, Investigation, Methodology, Project administration, Software, Supervision, Validation, Writing – review and editing. SV: Supervision, Writing – review and editing. ES: Project administration, Supervision, Writing – review and editing.

Funding

The author(s) declared that financial support was not received for this work and/or its publication.

Conflict of interest

The author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Generative AI statement

The author(s) declared that generative AI was used in the creation of this manuscript. During the preparation of this work, the authors used ChatGPT, Quillbot, and Grammarly to improve the readability and language of some parts of the article. The use of these technologies complied with the terms of use of the relevant tool or technology of Frontiers. After using these tools, the authors reviewed and edited the content as needed and take full responsibility for the content of the publication.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

References

Alijoyo, F. A., Gongada, T. N., Kaur, C., Mageswari, N., Sekhar, J., Ramesh, J. V. N., et al. (2024). Advanced hybrid cnn-bi-lstm model augmented with ga and ffo for enhanced cyclone intensity forecasting. Alexandria Eng. J. 92, 346–357. doi:10.1016/j.aej.2024.02.062

CrossRef Full Text | Google Scholar

Barcikowska, M., Feser, F., Zhang, W., and Mei, W. (2017). Changes in intense tropical cyclone activity for the Western north Pacific during the last decades derived from a regional climate model simulation. Clim. Dyn. 49, 2931–2949. doi:10.1007/s00382-016-3420-0

CrossRef Full Text | Google Scholar

Barrera-Animas, A. Y., Oyedele, L. O., Bilal, M., Akinosho, T. D., Delgado, J. M. D., and Akanbi, L. A. (2022). Rainfall prediction: a comparative analysis of modern machine learning algorithms for time-series forecasting. Mach. Learn. Appl. 7, 100204. doi:10.1016/j.mlwa.2021.100204

CrossRef Full Text | Google Scholar

Chen, L., Chen, Z., Zhang, Y., Liu, Y., Osman, A. I., Farghali, M., et al. (2023). Artificial intelligence-based solutions for climate change: a review. Environ. Chem. Lett. 21, 2525–2557. doi:10.1007/s10311-023-01617-y

CrossRef Full Text | Google Scholar

Faisal, G., Sreelakshmi, S., and Chandra Ss, V. (2023). “Crop yield prediction for smart agriculture with climatic parameters using random forest,” in International conference on advances in computing and data sciences (Springer), 367–376.

Google Scholar

Karevan, Z., and Suykens, J. A. (2020). Transductive lstm for time-series prediction: an application to weather forecasting. Neural Netw. 125, 1–9. doi:10.1016/j.neunet.2019.12.030

PubMed Abstract | CrossRef Full Text | Google Scholar

Kieu, C., Luong, K., and Nguyen, T. (2025). Nwp-based deep learning for tropical cyclone intensity prediction. arXiv Preprint arXiv:2504.09143.

Google Scholar

Knapp, K. R., Kruk, M. C., Levinson, D. H., Diamond, H. J., and Neumann, C. J. (2010). The international best track archive for climate stewardship (ibtracs) unifying tropical cyclone data. Bull. Am. Meteorological Soc. 91, 363–376. doi:10.1175/2009bams2755.1

CrossRef Full Text | Google Scholar

Knutson, T. R., McBride, J. L., Chan, J., Emanuel, K., Holland, G., Landsea, C., et al. (2010). Tropical cyclones and climate change. Nat. Geoscience 3, 157–163. doi:10.1038/ngeo779

CrossRef Full Text | Google Scholar

Li, X., Zhang, L., Wang, X., and Liang, B. (2024). Forecasting greenhouse air and soil temperatures: a multi-step time series approach employing attention-based lstm network. Comput. Electron. Agric. 217, 108602. doi:10.1016/j.compag.2023.108602

CrossRef Full Text | Google Scholar

Mesias, C. G., and Bagtasa, G. (2025). Ai-based tropical cyclone rainfall forecasting in the Philippines using machine learning. Meteorol. Appl. 32, e70083. doi:10.1002/met.70083

CrossRef Full Text | Google Scholar

Null, J., and Grundstein, A. (2025). Understanding wind speed: a primer on the beaufort, saffir-simpson, and fujita scales. Weatherwise 78, 17–23. doi:10.1080/00431672.2025.2533715

CrossRef Full Text | Google Scholar

Rahman, S., Faisal, M. F., Mondal, P., and Rahman, M. M. (2025). Tropical cyclone track prediction harnessing deep learning algorithms: a comparative study on the northern Indian Ocean. Results Eng. 26, 105009. doi:10.1016/j.rineng.2025.105009

CrossRef Full Text | Google Scholar

Shi, Y., Hu, R., Wu, N., Zhang, H., Liu, X., Zeng, Z., et al. (2025). Comparison of ai and nwp models in operational severe weather forecasting: a study on tropical cyclone predictions. J. Geophys. Res. Mach. Learn. Comput. 2, e2024JH000481. doi:10.1029/2024jh000481

CrossRef Full Text | Google Scholar

Shu, S., Zhang, F., Ming, J., and Wang, Y. (2014). Environmental influences on the intensity changes of tropical cyclones over the Western north Pacific. Atmos. Chem. Phys. 14, 6329–6342. doi:10.5194/acp-14-6329-2014

CrossRef Full Text | Google Scholar

Singh, A., Kohno, N., and Fudeyasu, H. (2024). Case study of high waves in the south Pacific generated by tropical cyclone harold in 2020. Trop. Cyclone Res. Rev. 13, 147–160. doi:10.1016/j.tcrr.2024.08.003

CrossRef Full Text | Google Scholar

Song, X., Liu, Y., Xue, L., Wang, J., Zhang, J., Wang, J., et al. (2020). Time-series well performance prediction based on long short-term memory (lstm) neural network model. J. Petroleum Sci. Eng. 186, 106682. doi:10.1016/j.petrol.2019.106682

CrossRef Full Text | Google Scholar

Tang, C. K., Chan, J. C., and Yamaguchi, M. (2021). Large tropical cyclone track forecast errors of global numerical weather prediction models in Western north Pacific basin. Trop. Cyclone Res. Rev. 10, 151–169. doi:10.1016/j.tcrr.2021.07.001

CrossRef Full Text | Google Scholar

Taylor, H. T., Ward, B., Willis, M., and Zaleski, W. (2010). “The saffir-simpson hurricane wind scale,”. Washington, DC, USA.Atmos. Adm.

Google Scholar

Wang, Z., Zhao, J., Huang, H., and Wang, X. (2022). A review on the application of machine learning methods in tropical cyclone forecasting. Front. Earth Sci. 10, 902596. doi:10.3389/feart.2022.902596

CrossRef Full Text | Google Scholar

Xu, Z., Guo, J., Zhang, G., Ye, Y., Zhao, H., and Chen, H. (2024). Global tropical cyclone size and intensity reconstruction dataset for 1959–2022 based on ibtracs and era5 data. Earth Syst. Sci. Data 16, 5753–5766. doi:10.5194/essd-16-5753-2024

CrossRef Full Text | Google Scholar

Yanai, M. (1964). Formation of tropical cyclones. Rev. Geophys. 2, 367–414. doi:10.1029/rg002i002p00367

CrossRef Full Text | Google Scholar

Yang, X., Zhang, F., Sun, P., Li, X., Du, Z., and Liu, R. (2022). A spatio-temporal graph-guided convolutional lstm for tropical cyclones precipitation nowcasting. Appl. Soft Comput. 124, 109003. doi:10.1016/j.asoc.2022.109003

CrossRef Full Text | Google Scholar