Optimizing Photovoltaic-Storage Building Energy Systems: A Comparative Study of Rule-Based and Reinforcement Learning Control for Grid Stability and Self-Consumption

Abstract

With the advancement of energy transition, the adoption of photovoltaic systems in residential buildings has been increasing. However, their intermittent and unstable nature poses challenges to grid stability. Integrating energy storage batteries into building energy systems has emerged as a key solution to enhance grid reliability. Despite this, optimizing battery charging and discharging strategies to achieve self-sufficiency, peak load shaving, and supply-demand balance remains a challenge. This study introduces two battery control strategies: Rule Based Control (RBC) approach and Reinforcement Learning model using Proximal Policy Optimization (PPO). These strategies dynamically coordinate PV generation, user demand and battery operations to reduce grid dependency and minimize fluctuations. Firstly, a physics-informed machine learning model was developed to accurately predict battery energy flows under varying states, enabling informed decision-making on grid feedback or consumption. Results from experiments with real data indicate that the combined use of physics-based models and machine learning can predict building-grid energy usage with an accuracy of up to 92%. Furthermore, the study compares the effectiveness of RBC and PPO in refining battery control strategies. Performance evaluations in a case study demonstrate that both RBC (28% and 94%) and PPO (27% and 86%) significantly enhance energy self-consumption and self-sufficiency, outperforming traditional methods (15% and 38%). In terms of operational strategies, RBC exhibits superior performance over PPO in stabilizing the grid and enhancing controllability. This research offers new insights into using machine learning for optimizing building-grid interactions and supports the deployment of integrated PV-storage systems in residential applications.