Enhancing robustness of cooling load forecasting in data-scarce scenarios: An integrated framework combining Bayesian diagnosis and Residual-domain Kalman Correction

Accurate short-term cooling load forecasting is a prerequisite for optimizing HVAC system control and energy efficiency. However, in real-world applications, the reliability of data-driven models is frequently compromised by two major issues: distribution shifts caused by undetected system faults and the scarcity of high-quality training data. To address these challenges, this study proposes a robust integrated framework that combines system state diagnosis with a novel forecasting mechanism. First, an interpretable Bayesian Network (BN) is employed to diagnose the HVAC system’s operating state, acting as a ‘gatekeeper’ to ensure the validity of the forecasting task. Upon confirming normal operation, a Residual-domain Kalman Correction (RKC) model performs the load prediction. This hybrid model adopts a ‘base-correction’ paradigm, fusing deep learning with Kalman filtering to dynamically correct prediction residuals, thereby significantly enhancing noise suppression. Based on hybrid datasets generated by simulating fault scenarios under real-world operating conditions, the diagnosis module achieves 98.75% accuracy in identifying system states. Furthermore, the RKC model outperforms mainstream benchmarks with a coefficient of determination (R²) of 96.4% and a 17.5% reduction in RMSE. Importantly, under data-scarce conditions, the model exhibits exceptional robustness, with performance degrading by only 18.1% compared to over 60% for standard models. This framework provides a reliable solution for deploying data-driven forecasting in complex, data-limited engineering environments.