Abstract: Food security is a vital component of national security, and intelligent grain storage management plays a crucial role in reducing storage losses. Predicting grain silo temperatures poses challenges due to complex spatiotemporal dependencies influenced by environmental factors and heat transfer mechanisms, with accurate prediction being essential for grain storage risk prevention and control. This study proposes an approach combining Long Short-Term Memory (LSTM) networks, spatial attention mechanisms, and Adaptive Boosting (AdaBoost) ensemble learning for grain silo temperature prediction. Experimental data were collected from July 8 to November 8, 2023, at a flat-bottomed silo at Henan University of Technology. This rectangular silo (8.3m × 5.6m × 6m) features a reinforced concrete structure with insulation and an air-conditioning system, storing bulk grain. It is equipped with 22 sensor cables: 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 houses 12 temperature and humidity sensors in a three-dimensional grid, with microcontrollers managing data acquisition, processing, and transmission to ensure real-time performance and reliability. The experiment collected 784,320 data points (three-dimensional coordinates, temperature, relative humidity) updated hourly. The research methodology comprises three core components: temporal feature extraction, spatial dependency modeling, and ensemble learning optimization. First, comprehensive analysis of experimental data via the autocorrelation function revealed temporal patterns and critical 24-hour lag dependencies. Among comparative models including Transformer, LSTM, XGBoost, and CNN, the LSTM model demonstrated robust baseline performance (R²: 0.9102), capturing complex dynamic temporal features. Second, spatial distribution analysis showed non-uniform temperature variations in the silo’s three-dimensional space: outer layers of grain piles were warmer than core regions, with pronounced vertical temperature gradients. Traditional time-series models assume spatial homogeneity, failing to capture such complex dependencies, and LSTM exhibits lag and bias during abnormal temperature fluctuations. Thus, an LSTM-Attention model was introduced to adaptively learn spatial feature weights across time steps, capturing spatiotemporal interactions (R²: 0.9470). To enhance robustness and generalization, the AdaBoost meta-algorithm sequentially trains multiple LSTM-Attention weak learners, yielding the LSTM-Attention-AdaBoost model. Experimental results demonstrate that the proposed method significantly outperforms other comparative models, achieving outstanding prediction errors (MSE: 0.5588, RMSE: 0.7475, MAE: 0.4404) and excellent fitting quality (R²: 0.9589). By enabling precise temperature predictions, this model provides a technical reference for intelligent grain storage management, empowering operators to implement targeted interventions (directional ventilation, localized cooling, precise fumigation). It drives the shift from reactive to proactive management, identifying hotspots in early stages, suppressing mold growth, and interrupting pest infestation cycles. This safeguards grain quality, reduces large-scale economic losses, and offers broad application prospects in modern intelligent grain storage ecosystems, with practical significance.