Spatiotemporal self-attention network ST-GranNet for granary temperature prediction.

Area of Science:

• Agricultural Engineering
• Data Science
• Artificial Intelligence

Background:

• Accurate food storage temperature prediction is vital for supply chain safety and waste reduction.
• Traditional methods struggle with complex nonlinearities and spatiotemporal feature extraction in granary environments.
• Existing deep learning models lack sufficient interaction and external variable integration for granary temperature prediction.

Purpose of the Study:

• To develop an innovative spatiotemporal attention framework, ST-GranNet, for high-precision granary temperature prediction.
• To address limitations in capturing complex sensor relationships and integrating external variables in existing models.
• To enhance food safety and reduce food waste through improved temperature monitoring.

Main Methods:

• Proposed ST-GranNet, a spatiotemporal attention framework utilizing self-attention mechanisms.
• Employed channel embedding and point embedding for spatial correlations and long-term dependencies.
• Developed a novel fusion approach to integrate spatiotemporal features, validated on a real corn storage dataset.

Main Results:

• ST-GranNet outperformed conventional models (LSTM, GRU) and emerging self-attention models in prediction accuracy (MAE, MSE).
• Hyperparameter optimization and ablation studies confirmed the model's effectiveness.
• Visualization experiments elucidated the self-attention mechanism's role in prediction.

Conclusions:

• ST-GranNet offers a high-precision tool for intelligent grain storage management and risk warning.
• The model has significant potential to reduce grain loss and optimize storage conditions.
• This research contributes to safer food supply chains and more efficient agricultural practices.