An interpretable temperature prediction method for grain in storage based on improved temporal Fusion Transformers

Accurate temperature prediction for grain in storage is crucial for safety monitoring and early warning of abnormal conditions. Existing methods for predicting grain temperature face several challenges, such as neglecting spatial dependencies between grain temperatures at different locations within the grain pile, low accuracy, insufficient consideration of multifactorial influences, poor generalization capabilities, and lack of interpretability. To address these challenges, an interpretable temperature prediction model for grain in storage is proposed based on Temporal Fusion Transformers (TFT) integrated with a Graph Convolutional Network (GCN) module. This integration enables the model to simultaneously capture both spatial and temporal dependencies of grain temperatures. The model processes historical grain temperatures, meteorological data, granary internal air temperature and humidity, grain moisture content, grain varieties, and other granary information, categorizing these into static and dynamic variables. The inclusion of weather forecast data for the granary location as known future variables significantly improves prediction accuracy. The interpretability of the model allows for the visualization of input variable importance rankings. Comparative experiments demonstrate that the proposed GCN-TFT model outperforms other comparable models. Practical application experiments further confirm the model’s applicability and effectiveness in predicting grain temperatures. Furthermore, the use of an interpretable model signifies a significant advancement in grain temperature prediction. The interpretable results are expected to assist granary managers in developing effective grain storage management strategies.