Publicaties
A dual deep learning approach for winter temperature prediction in solar greenhouses in Northern China
Yu, Jingxin; Zhao, Jinpeng; Sun, Congcong; Zhang, Ruochen; Zheng, Wengang; Xu, Linlin; Wei, Xiaoming
Samenvatting
Achieving accurate and efficient winter temperature prediction in solar greenhouses is critically important but challenging for greenhouse cultivation, especially in northern China. To address this challenge, this study proposes three key innovations: (1) a dual deep learning architecture integrating Transformer's global modeling with BiLSTM's local processing; (2) enhanced feature extraction through relative positional encoding and hybrid sparse attention; and (3) optimized computational efficiency via time embedding strategies. The proposed model was compared with 11 typical deep learning models (including ResNet, GRU, and TCN) through systematic experiments involving multiple input feature combinations (5-18 features) and various prediction durations (6-48 h). The results indicate that the proposed model outperforms typical deep learning models across all evaluation metrics, including an average MSE of 0.0599°C and an R2 of 0.9989 in lightweight configurations, a high improvement over the best existing methods. Furthermore, the proposed model demonstrates exceptional stability (MSE standard deviation < 0.01°C) and generalization ability (R2> 0.98) under different input combinations and prediction durations. This study identifies indoor temperature, vapor pressure deficit (VPD), and indoor solar radiation as the three most critical factors influencing the prediction accuracy. Through comprehensive feature combinations, the prediction performance improves significantly with MSE reduced by 70.1% (from 0.1959°C to 0.0586°C). Among all scenarios, the 288-144 configuration (24-hour input predicting 12 h) achieves optimal balance of accuracy and efficiency. Even in challenging long-term scenarios like 576-576 (48-hour input predicting 48 h), the model still maintains good accuracy (R2=0.9665). This research provides new technological support with an accurate and efficient winter temperature prediction model for the intelligent control of greenhouse environments, offering important implications for optimal greenhouse cultivation.