A hybrid machine learning framework for optimizing heat pump-driven domestic hot water systems based on user behaviour and control strategies

Abstract

• Heterogeneous dynamics challenge conventional rule-based DHW control systems. • A multi-output machine learning model is proposed for heat pump-driven DHW systems. • The framework correlates user behaviour and weather with optimal control strategies. • High-fidelity targets are created via combined optimization and dynamic simulations. • The predictive model using ensemble regressors serves as a data-driven surrogate. Regulatory and technological advancements have reduced space heating demand, positioning domestic hot water (DHW) systems as the next bottleneck in the decarbonization of residential buildings. Increasingly dynamic energy profiles, on-site generation and storage, and heterogeneous system interactions driven by occupant behaviour and environmental variability pose significant challenges to conventional control strategies. Given the importance of accurately capturing the probabilistic relationship between user behaviour, boundary conditions, and intelligent control strategies in heat pump-driven DHW networks, this study develops a multi-output machine learning (ML) framework to enhance energy performance, operational flexibility, and user-centric responsiveness. A hybrid methodology is proposed to explicitly correlate control strategies with occupant behaviour and ambient conditions, formulating five strategies as design variables within a statistical multi-objective optimization to capture trade-offs across four response objectives. The forecasting ML model, using three ensemble-based multi-output regressors, serves as a data-driven surrogate for optimal control strategy selection, allowing for robust learning of complex control interactions. A rigorous integration of RSM-CCD algorithm, TRNSYS dynamic simulation, and a stochastic Gaussian demand generator developed in a MATLAB code underpins the methodological core of this study, yielding a performance-optimized dataset through combined statistical and high-fidelity modelling. The results indicate that the established framework provides a meaningful avenue to optimized energy efficiency and user-centric responsiveness, and that the predictive model exhibits a technically sound performance. An optimized charging schedule reduces DHW thermal losses by over 34 %, while increasing the annual renewable energy fraction by up to 25 %. The trained Random Forest (RF) regressor ( R 2 = 0.987) outperforms other models across various metrics, followed by the Bayesian-optimized XGBoost. The correlation matrix of features reveals strong positive alignment (+0.85) between heat pump charging and recirculating loop operation, along with negative correlation with DHW consumption. The hybrid approach proposed in this study leverages the accuracy of physics-based simulations with the adaptability and scalability of ML prediction, enabling reliable forecasting of optimal control strategies applicable to real subsequent inputs beyond the training dataset.

Topics & Concepts