EA-OODRL: Ensemble attentional offline-to-online deep reinforcement learning for microgrid energy optimization

This is the authors' abstract. We don't add key points for this paper.

Optimizing microgrid scheduling is critical for addressing the unpredictability of renewable energy, dynamic load fluctuations, and the complexity of grid integration. This study proposes an Ensemble Attentional Offline-to-Online Deep Reinforcement Learning (EA-OODRL) algorithm for small-scale microgrids, aiming to enhance generalization and decision-making efficiency in scheduling tasks. The algorithm leverages a multi-head attention mechanism to improve environmental perception of sub-policies, enabling effective processing of high-dimensional, complex data and capturing diverse features. Through shared data and an ensemble resonance loss, EA-OODRL promotes collaboration among sub-policies to enhance exploration capabilities. Additionally, a diversity-regularized loss is introduced to prevent convergence to a single sub-optimal policy, ensuring adaptability to unseen states. Experiments on real-world datasets show that, while maintaining performance on par with the best baseline, our method reduces inference time by up to 41%, exhibits markedly smaller performance fluctuations when generalizing to unseen states, and achieves up to 64% lower operational cost than other baselines, demonstrating superior algorithmic decision efficiency and generalization capability.This research offers an effective solution for microgrid scheduling and provides new insights into applying deep reinforcement learning in complex systems.