Online incremental learning approach of heat pump and chiller models based on the dynamic random forests in queue structure

Machine learning has been widely utilized in modeling heat pumps and chillers to predict key parameters such as the coefficient of performance (COP) and develop physics-data hybrid-driven models. However, during the online engineering applications, factors such as equipment aging, faults, upgrades and expanded operating conditions can cause dataset shifts and dynamic changes in thermodynamic performance, leading to model failures. To address these challenges, this study introduces an online incremental learning approach based on a novel algorithm, the dynamic random forests in queue structure (DRFQS). This algorithm incorporates a universally applicable dataset shift detection method and clear stability-plasticity balance control. By integrating COP DRFQS predictors with mechanism models, heat pump and chiller models with self-evolution capabilities were developed. Simulation results including over 100 COP drop fault scenarios demonstrate that the self-evolution models outperform non-evolvable models trained via offline batch learning. The self-evolution models autonomously adapt to new time-varying and condition-dependent patterns in thermodynamic performance without interruptions, effectively overcoming dataset shift issues during long-term use. The models significantly enhance prediction accuracy for COP, thermodynamic perfection, and other parameters with prediction errors decreased by 37.1 %∼84.6 %. The study addressed the limitations of offline batch learning, offering an advanced AI-enhanced approach for HVAC system modeling.