To address the insufficient energy efficiency prediction accuracy caused by topological coupling in the parallel operation of multiple chillers, this study proposes a physics-guided spatiotemporal fusion model combining Long Short-Term Memory (LSTM) and Graph Convolutional Network (GCN). The LSTM module extracts temporal features from single-unit energy consumption sequences; the GCN module captures spatial dependencies through constructed topological graphs representing cooling water networks and load distribution relationships. Based on the operational data from a large-scale data center cooling station, the model is trained and tested using 128 million raw records along with 8000 h of simulation data. Experimental results demonstrate that the GCN-LSTM model achieves a 19.4% reduction in Root Mean Square Error and a 2.06% decrease in Mean Absolute Percentage Error compared to standalone LSTM models at 30-min prediction intervals. The ablation experiment further verifies the necessity of GCN for spatial modeling, and its absence led to a significant increase in prediction error. The proposed GCN-LSTM architecture surpasses traditional single-model performance limitations and provides an adaptable solution for multi-equipment collaborative optimization. This approach establishes a new technical path for industrial energy conservation and intelligent control systems.
Liu et al. (Thu,) studied this question.