Energy Management in Grid-Connected Photovoltaic Battery Systems Using Degradation-Aware Model Predictive Control-Based Reinforcement Learning

Abstract

This thesis presents a degradation-aware Model Predictive Control–based Reinforcement Learning (MPC-RL) framework for optimal energy management in residential grid-connected photovoltaic (PV) battery systems. The goal is to minimize total operating costs while accounting for battery degradation, grid price variability, and forecast uncertainty. The MPC formulation incorporates a parametrized stage cost that includes grid energy costs, battery degradation, and state-of-charge (SoC) penalty window, with parameters tuned using a Q-learning algorithm. Gaussian Process Regression (GPR) is used to forecast the net residual power (PV production minus demand) based on temporal and historical features, enabling a realistic simulation environment and improved MPC decision-making. The proposed method is benchmarked against a baseline non-parametrized MPC controller, purely RL-based strategies from literature, and a simple Gaussian residual model. Results show that while the proposed MPC-RL approach does not outperform all RL benchmarks in short-term simulations, it achieves competitive performance with reduced degradation and improved constraint handling. GPR-based residual forecasts outperform the Gaussian baseline, reducing total monthly costs by up to 2.2 % compared to the random-sampling approach, and closely tracking actual residual trends. These findings suggest that MPC-RL with GPR prediction offers a promising and degradationconscious approach for residential microgrid energy management.