A Thermal State Field Forecasting Framework for Intelligent Ventilation Decision-Making in Granary Based on a Hierarchical TCN-Mamba Network

Temperature forecasting from distributed temperature sensing supports thermal characterization and ventilation decisions in grain storage. However, existing studies mainly focus on individual sensor points and are insufficient to describe spatial thermal evolution. This study proposes a thermal state field forecasting framework integrating a Temporal Convolutional Network and Mamba. Sensor observations are aggregated into spatial blocks to construct average temperature, temperature variance, and three-dimensional temperature gradients. Spatial attention captures local correlations, while the hierarchical TCN and Mamba encoder and Cross-Scale Trend Attention extract multi-scale temporal features and long-range evolution patterns. Across all 30 spatial blocks, the proposed model achieved RMSE values of 0.2988 ± 0.0374, 0.5573 ± 0.0112, 0.0380 ± 0.0009, 0.0442 ± 0.0031, and 0.0788 ± 0.0050, and MAE values of 0.2591 ± 0.0330, 0.4270 ± 0.0093, 0.0344 ± 0.0008, 0.0379 ± 0.0031, and 0.0684 ± 0.0053 for average temperature, temperature variance, and gradients in the X, Y, and Z directions, respectively. Based on the predicted thermal state field, an illustrative forecast-driven decision-support framework provides event-triggered ventilation decision support and regional priority recommendations. The decision analysis further illustrates the relationship between predicted thermal evolution and ventilation timing and regional prioritization under predefined rules.