Author: Huimin Yao (The Hong Kong Polytechnic University) - Central chilled water systems exhibit complex, coupled hydraulic and thermal behaviors, which poses a significant challenge for conventional predictive models. This study introduces a novel spatio-temporal framework that integrates Graph Neural Networks (GNNs) to capture the system's physical topology, with Recurrent Neural Networks (RNNs) to model its temporal dynamics. Validated on a high-fidelity simulation of a campus-wide chilled water system, the proposed model demonstrates substantially improved performance over traditional baseline models, achieving over a 50% reduction in Mean Absolute Error (MAE) for temperature prediction. Furthermore, to address practical deployment challenges, this study validates robust strategies for handling incomplete sensor data. The results show that cross-sectional mean imputation is a pragmatic and effective approach for maintaining model reliability. This research establishes a high-fidelity and robust modeling framework, enhancing the accuracy and practicality needed for advanced predictive control in large-scale building energy systems.
