Adaptive model predictive control for solar-assisted multi-source heat pump systems using machine learning

To address the operational challenges of solar-assisted multi-source heat pump systems under fluctuating environments, this study proposes a physics-data integrated Adaptive Model Predictive Control (AMPC) framework. The methodological resides in the synergetic coupling of high-fidelity thermodynamic physical models with stochastic data-driven predictors. This hybrid architecture enables the framework to synchronize Ground Source Heat Pumps, Air Source Heat Pumps, and solar thermal collectors by capturing complex non-linear thermal couplings that pure data-driven models often overlook. During the forecasting phase, multiple machine learning algorithms are evaluated. Results indicate that Long Short-Term Memory achieves superior performance for summer cooling loads with an R2 of 0.990, while LightGBM proves most effective for handling the stochastic fluctuations of winter heating loads, yielding an R2 of 0.96. Regarding control strategy optimization, sensitivity analysis identifies a 12-h prediction horizon and a 3-h control range as the optimal configuration. This setup achieves an 8.3% energy reduction while maintaining a computational response time of 19.6 s, ensuring real-time feasibility. Full-season comparative experiments demonstrate that the proposed AMPC framework offers significant advantages over conventional PID control. The AMPC demonstrates superior robustness by reducing indoor temperature fluctuations by over 65% and lowering the MAE by at least 50% across all seasons. Ultimately, the system secures seasonal energy savings of 12.0% in summer and 8.7% in winter while consistently maintaining high levels of thermal comfort. These findings highlight the potential of AMPC in enhancing the stability and economy of multi-source energy systems, providing an effective technical solution for low-carbon building operations.