Research on Moisture Prediction Based on Spatio-Temporal Graph Neural Network

To address issues such as significant control deviations and insufficient stability caused by fluctuations in incoming materials, environmental changes, and measurement delays during the control of outlet moisture content in the loosening and conditioning process, this study developed a spatio-temporal graph neural network model based on Adaptive Graph Convolutional Recurrent Network(AGCRN) for moisture prediction. By adaptively learning the spatial topological relationships among sensors and integrating gated recurrent units to capture temporal dependencies, the model achieves accurate prediction of outlet moisture content. In response to the variability of industrial scenarios, a dynamic graph construction mechanism and a multi-objective loss function were introduced during training to further enhance the model’s generalization and adaptive capabilities. Practical applications demonstrate that after system optimization, the control deviation of outlet moisture content was stably reduced from the original ± 1.2% to within ± 0.5%, and the process capability index CPK increased by 31%, effectively improving the control stability and quality consistency of the loosening and conditioning process.