Deep reinforcement learning for economic battery dispatch: A comprehensive comparison of algorithms and experiment design choices

Abstract

Deep reinforcement learning (DRL) is an increasingly popular optimization tool for the economic dispatch of battery energy systems. However, it remains largely unclear which DRL models, experiment setups, and hyperparameters are most effective, due to contradictory results in the literature and a lack of thorough benchmarks. To address this, we compare popular DRL models and experimental design choices for battery dispatch tasks. We first formulate two battery dispatch tasks that reflect the cross section of existing approaches: energy arbitrage and load following combined with improved renewable energy utilization. We then benchmark four DRL models on case studies in Canada and Germany. Besides the choice of DRL model, we conduct a comprehensive comparison of these experiment design choices: continuous and discrete action spaces, time counters added to the state space (e.g. for the hour of the day), observation stacking, and reward function penalties. Our results show that time counters and observation stacking reliably improve DRL performance. For time counters, cyclic encodings through sine and cosine waves worked best, especially if multiple counters were combined. Reward penalties were less beneficial and barely increased rewards. Deep Q-Networks (DQN) outperformed the other DRL models and significantly benefited from the tested features, increasing rewards by 51% and 20% on the respective case studies. This research highlights the need to carefully choose DRL models, experiment design, and hyperparameters. Our comparative analysis provides a solid foundation for future research on DRL for battery dispatch. By building upon our findings, researchers and practitioners can streamline the experiment design process and obtain better outcomes. • Application of deep reinforcement learning to economic battery dispatch. • Tasks analyzed include energy arbitrage, load following, renewable energy control. • Benchmark of algorithms, deep Q-networks was found to perform best. • Cyclic encoded time features and observation stacking greatly increased rewards. • With the right setup, reinforcement learning can outperform oracles.