An Adaptive Multi-Objective Decision System With Reinforcement Learning for PV-ESS Sizing and Operation for Grid-Connected Industrial PV-ESS Systems in Distribution Networks

The integration of photovoltaic (PV) systems with energy storage systems (ESS) introduces complex design and operational challenges due to the stochastic nature of solar generation and the dynamic behavior of energy markets. This paper presents a reinforcement learning -enhanced Decision Support System (DSS). The system jointly addresses the optimal sizing and real-time control of PV-ESS configurations. The proposed framework combines a multi-objective optimization core with a policy-gradient RL agent. This agent jointly optimizes operational decisions and dynamically adjusts the weights in the multi-objective function—based on real-time feedback and system state—to reflect the relative importance of competing objectives such as cost, emissions, curtailment and battery degradation. These adjustments reflect evolving environmental, market and load conditions. Historical and real-time data from utility-scale PV parks and industrial consumers inform the learning process, ensuring policy relevance and robustness. The system is designed for grid-connected industrial sites embedded in distribution networks, where node-level optimization can indirectly support peak reduction, load smoothing and demand-side flexibility at the network level. The system utilizes a policy-gradient reinforcement learning agent with adaptive weight tuning to balance conflicting objectives. Simulation results over a two-year horizon demonstrate that the proposed RL-DSS reduces average energy costs by up to 35%, increases PV self-consumption above 70% and lowers curtailed energy below 8%, outperforming static and rule-based strategies. The results validate the potential of RL-based decision systems to enable scalable, flexible and economically viable energy management under real-world uncertainty. Future work includes integration with probabilistic forecasting models and distributed multi-agent coordination, where each industrial node may act as an autonomous RL agent contributing to cooperative grid-level objectives.