Abstract
Hybrid renewable energy systems (HRES) integrate multiple renewable sources and storage to improve reliability and efficiency, but their operation is challenged by stochastic generation and load variability. Deep reinforcement learning (DRL) has emerged as a promising approach for real-time optimization, yet comparative performance across different DRL algorithms and system configurations remains underexplored. This study proposes a unified DRL-based optimization framework for HRES and evaluates three algorithms—Deep Q-Network (DQN), Proximal Policy Optimization (PPO), and Soft Actor-Critic (SAC)—on a simulated microgrid comprising photovoltaic, wind turbine, battery storage, and hydrogen production subsystems. The framework formulates energy management as a Markov decision process with continuous action spaces and multi-objective rewards including cost minimization, renewable utilization, and battery longevity. Experiments conducted over one year of hourly data demonstrate that SAC achieves the highest average reward (0.92) and reduces operational costs by 18.5% compared to heuristic rule-based control, while PPO exhibits superior robustness under high variability. The results also show that DRL controllers significantly improve renewable energy curtailment reduction and battery state-of-health preservation. This work provides a comprehensive benchmark and practical guidelines for deploying DRL in HRES, highlighting the trade-offs between sample efficiency, stability, and optimality.