Reinforcement learning-based adaptive coordination of solar energy for load, battery, electric vehicle, and hydrogen storage

The intermittency of solar photovoltaic (PV) systems creates critical challenges in real-time energy coordination, multi-energy routing, and long-duration storage management that conventional rule-based controllers fail to resolve under uncertainty. Battery storage capacity limitations necessitate complementary hydrogen storage for long-duration energy resilience, yet no unified framework has effectively coordinated both with PV generation, load demand, and grid interaction in real time. To address these gaps, a hybrid energy-management framework is proposed that integrates PV generation, battery storage, hydrogen production, and grid interaction within a unified smart-home architecture at one-minute control resolution. A random forest (RF)-based PV forecasting model (MAE: 0.3213, R 2 : 0.9881) and an artificial neural network (ANN)-based hydrogen prediction model (MAE: 0.0013, R 2 : 0.9987) enable proactive, forecast-informed energy allocation under variable irradiance. A multi-agent reinforcement learning with proximal policy optimization (MARL-PPO) strategy achieves optimal real-time dispatch, contributing 56.2% of total energy from PV (532.80 kWh), reducing grid dependency to 30.2%, and routing 248.02 kWh of surplus PV to produce 4.72 kg of green hydrogen with zero curtailment. Compared with rule-based and greedy baseline controllers, MARL-PPO achieves superior hydrogen yield (4.72 kg) and improves energy efficiency from 72.6% to 86.3% over without-hydrogen configuration. Monte Carlo robustness analysis (N=50) confirms a tight distribution of grid dependency (30.20%–30.40%, Std: 0.015%), and forecast error sensitivity analysis shows only a 0.38 percentage-point increase in grid dependency as PV forecast error rises from 0% to 15%, validating the proposed framework for reliable grid-independent smart-home energy management.