AUTHOR=Liang Longwei , Shi Hui , Wang Zhaoyuan , Wang Shengjie , Li Changhong , Diao Ming 

TITLE=Research on time series prediction model for multi-factor environmental parameters in facilities based on LSTM-AT-DP model

JOURNAL=Frontiers in Plant Science

VOLUME=Volume 16 - 2025

YEAR=2025

URL=https://www.frontiersin.org/journals/plant-science/articles/10.3389/fpls.2025.1652478

DOI=10.3389/fpls.2025.1652478

ISSN=1664-462X

ABSTRACT=IntroductionExisting facility environment prediction models often suffer from low accuracy, poor timeliness, and error accumulation in long-term predictions under multifactor nonlinear coupling conditions. These limitations significantly constrain the effectiveness of precise environmental regulation in agricultural facilities.MethodsTo address these challenges, this paper proposes a novel facility environment prediction model (LSTM-AT-DP) integrating Long Short-Term Memory networks with attention mechanisms and advanced data preprocessing. The model architecture employs: (1) a Data Preprocessing (DP) module combining Wavelet Threshold Denoising (WTD) for noise elimination and Sliding Window (SW) technique for feature matrix construction; (2) an LSTM core for deep temporal modeling; and (3) an Attention Mechanism (AT) for dynamic feature weighting to enhance critical temporal feature extraction.ResultsIn 24-hour prediction tests, the model achieved determination coefficients (R²) of 0.9602 (temperature), 0.9529 (humidity), and 0.9839 (radiation), representing improvements of 3.89%, 5.53%, and 2.84% respectively over baseline LSTM models. Corresponding RMSE reductions were 0.6830, 1.8759, and 12.952 for these parameters.DiscussionThe results demonstrate that the LSTM-AT-DP model significantly enhances prediction accuracy while effectively suppressing error accumulation in long-term forecasts. This advancement provides robust technical support for precise facility environment regulation, with particular improvements observed in humidity prediction. The integrated attention mechanism proves particularly effective in identifying and weighting critical temporal features across all measured environmental parameters.