An online dynamic model based on Physical-Constraint Broad Learning System for extrapolation scenarios of chillers

Chillers account for the majority of energy consumption in central air-conditioning refrigeration stations. However, conventional models lack strong out-of-sample generalization, making it difficult to accurately reflect chiller performance variations under multiple operating conditions. Consequently, there is a lack of reliable multi-condition performance data to support energy-efficient regulation of chillers and optimization of refrigeration station control. To address this issue, this study proposes a physics-constrained broad learning System (PCBLS) method by introducing an error backpropagation mechanism and a customized physics-based loss function into the broad learning framework. This approach enhances the out-of-sample generalization of chiller models, enabling accurate prediction of chiller performance under unseen operating conditions based on measured data from existing conditions. The core idea is to ensure that the model’s predictions for unknown conditions remain consistent with physical laws. Experimental results demonstrate that, compared to methods without out-of-sample generalization enhancement, the proposed approach reduces MAE by 53.38%, RMSE by 55.47%, and improves R² by 19.62%, while decreasing the custom physics-based loss by approximately 99.37%. Additionally, the method maintains high accuracy while achieving fast training speeds.