A multi-feature prediction method for grain pile temperature based on regional modeling and deep spatio-temporal learning

Accurate grain pile temperature prediction is essential for maintaining grain quality and supporting aeration management in grain storage systems. This study proposes a spatio-temporal forecasting method for grain pile temperature prediction that integrates a spatial partitioning strategy with deep time-series modeling. According to the actual three-dimensional sensor layout, temperature measurements are organized into spatial blocks for region-level modeling. Spatial relationships among blocks are extracted through convolutional neural networks and neighborhood-based operations, while temporal dependencies are modeled using a hybrid framework combining LSTM, iTransformer, and gated fusion. The proposed model performs multi-feature prediction of block-level average temperature, temperature variance, and three-dimensional temperature gradients, providing a structured representation of the regional thermal state. Experiments were conducted on real grain storage data from the Jianghuai region of China and compared with Informer, PatchTST, Transformer, LSTM, and STGCN. Across all 30 spatial blocks, GLiT achieved relatively low RMSE and MAE for average temperature, temperature variance, and three-dimensional gradients. The results indicate that the proposed framework provides stable multi-feature forecasting performance while reducing the modeling dimension through block-level representation. This study provides a data-driven approach to regional thermal-state forecasting and a basis for further grain storage analysis and intelligent management.