The integration of renewable energy sources into microgrids remains challenging due to generation intermittency, storage inefficiencies, and progressive battery degradation. In this work, our main contribution is a standardized empirical benchmark of deep reinforcement learning for microgrid energy management across multiple operating regimes under a consistent evaluation protocol. To make these evaluations rigorous and comparable, we define a common state/action/reward specification that explicitly encodes battery-health proxies, dynamic pricing, and operational constraints, and we adapt a representative set of RL families accordingly: value-based methods (DoubleDQN, NoisyNet-DQN, PER-DQN, C51), policy-gradient approaches (PG), and actor–critic algorithms (PPO, A2C, SAC, DDPG, A3C-Energy). We benchmark the resulting agents on real-world driven datasets and environments (GEFCom2014, StoreNet, and a CityLearn-inspired setting), assessing performance under a consistent evaluation pipeline. Empirically, the reported results indicate that the adapted agents tend to reduce daily operating cost, smooth charge–discharge cycling, and improve battery-health proxies under realistic operating conditions, with gains that vary by dataset and operating regime. In particular, DQN-based variants exhibit consistent gains over standard exploration schemes, while methods such as PPO and actor–critic configurations maintain competitive and more stable behavior under highly volatile price signals and renewable fluctuations. Overall, the study provides scenario-conditional evidence on which RL families tend to perform well or poorly under the tested conditions, rather than a universal ranking across datasets.
Flórez et al. (2026) studied this question.