Toward zero-shot generalization of deep reinforcement learning for hybrid energy system dispatch

Hybrid energy system (HES) dispatch requires sequential operational decisions under uncertainty in renewable generation, demand, and market conditions. Deep reinforcement learning (DRL) is increasingly applied to this problem because it can learn dispatch policies from data without requiring full system models or future information. However, most existing studies train and evaluate DRL agents on the same environments, leaving their ability to generalize to unseen operating conditions largely untested. This study systematically investigates zero-shot generalization of DRL for HES dispatch. We propose training and evaluation protocols for two practically relevant settings: generalization to unseen time-series data and to unseen plant configurations, including interpolation and extrapolation. In addition, we introduce adaptive actions, a modification to discrete action spaces in which agents select decision types rather than exact power levels, while the corresponding power values are computed through predefined logic. The proposed methods are evaluated on two HES case studies with dispatchable generators, battery storage, and power-to-gas storage under renewable and market uncertainty. Results show that agents trained on diverse datasets can achieve generalization gaps below 5%, while adaptive actions significantly improve both training and test performance. Min–max observation scaling further enables robust generalization. These findings highlight the potential of DRL for robust dispatch and early stages of HES design.