Predicting grain storage temperature with spatiotemporal heterogeneity-informed grain storage temperature prediction integrating LSTM-attention and AdaBoost

Intelligent grain storage is one of the vital components in national food security. It is often required to predict the spatiotemporal temperature field in a complex silo for grain storage risk prevention and control. However, the accurate prediction can depend mainly on the environmental factors and heat transfer. This study aims to predict grain silo temperature by integrating Long Short-Term Memory (LSTM) networks, spatial attention mechanisms, and Adaptive Boosting (AdaBoost) ensemble learning. Experimental data were collected at a full-scale flat-bottomed silo at Henan University of Technology from July 8 to November 8, 2023. The rectangular silo (8.3 m × 5.6 m × 6 m) featured a reinforced concrete structure with insulation and an air-conditioning system, thus storing bulk grain. And 22 sensor cables were equipped: Sixteen were vertically and uniformly distributed within the silo; one near the door area, one at each corner adjacent to the wall, and two in a locally anomalous zone near the southeast corner wall. Each cable contained 12 temperature and humidity sensors in a three-dimensional grid. Microcontrollers were used for the data acquisition, processing, and transmission for the real-time performance and reliability. The experiment was conducted to collect 784 320 data points (three-dimensional coordinates, temperature, and relative humidity) to update hourly. Three components included: temporal feature extraction, spatial dependency modelling, and ensemble learning optimization. 1) Autocorrelation function of experimental data revealed that there were significant temporal patterns and critical 24-hour lag dependencies. Compared with the Transformer, LSTM, XGBoost, and CNN, the LSTM model demonstrated the robust baseline performance (R²: 0.910 2), effectively capturing complex dynamic temporal features. 2) Spatial distribution analysis showed that there were non-uniform temperature variations in the silo’s three-dimensional space: The outer layers of grain piles were warmer than the core regions, with pronounced gradients of the vertical temperature. An LSTM-Attention model was introduced to adaptively learn spatial feature weights over time steps, thus capturing spatiotemporal interactions (R²: 0.947 0). The spatial heterogeneity was also assumed to reduce the noticeable lag and bias during abnormal temperature fluctuations, rather than the LSTM. The complex dependencies were better captured compared with conventional time-series models. The AdaBoost meta-algorithm was used to sequentially train multiple LSTM-Attention weak learners. The LSTM-Attention-AdaBoost model was constructed to enhance robustness and generalization. Experimental results demonstrate that the outstanding prediction performance was achieved with the relative errors (MSE: 0.558 8, RMSE: 0.747 5, and MAE: 0.440 4 ℃) and excellent fitting quality (R²: 0.958 9). Precise temperature predictions can provide a reliable technical reference for intelligent grain storage. The shift from reactive to proactive management was realized to promptly implement the target interventions (directional ventilation, localized cooling, and precise fumigation). Hotspots in early stages were identified to interrupt the mold growth and pest infestation cycles. The grain quality can be improved to reduce the large-scale storage losses. The finding can also offer broad application prospects in intelligent ecosystems of grain storage in modern agriculture.