Balancing supply and demand constitutes the most important and challenging task in an isolated microgrid. Accordingly, it is essential to develop an optimization scheduling strategy for an energy management system of an isolated microgrid operation. In this study, a novel forecast-driven stochastic scheduling strategy was devised for the optimal operation of an isolated hydrogen microgrid. First, the change in wind power and load over 24 h was forecast using a bidirectional and long short-term memory convolutional neural network modeled end-to-end. To the best of the authors’ knowledge, this is the first application of end-to-end modeling for wind-power forecasting. Based on the forecast results, the stochastic optimization scheduling of the energy management system was resolved through deep reinforcement learning to minimize the microgrid lifecycle cost. Deep reinforcement learning combines the advantages of deep learning and reinforcement learning and uses statistical models to effectively solve sequence decisions of features of high-dimensional spaces. In addition, stochastic scenarios were generated using Monte Carlo simulations to analyze the uncertainties in wind and load. Furthermore, the energy capacity degradation of the energy storage system was considered. Finally, the effectiveness of the proposed approach was validated based on comparisons of different benchmark models and the latest models. The proposed scheduling strategy can realize high operational efficiency and reliable energy management system scheduling and is expected to serve as a reference for future research in this area.

平衡供需是孤立微电网中最重要和最具挑战性的任务。因此，有必要为孤立的微电网运行的能量管理系统制定优化调度策略。在这项研究中，设计了一种新的预测驱动的随机调度策略，用于孤立氢微电网的优化运行。首先，使用端到端建模的双向长短期记忆卷积神经网络预测 24 小时内风力和负荷的变化。据作者所知，这是风电预测端到端建模的首次应用。根据预测结果，通过深度强化学习解决能源管理系统的随机优化调度，最大限度地降低微电网生命周期成本。深度强化学习结合了深度学习和强化学习的优点，利用统计模型有效解决高维空间特征的序列决策问题。此外，使用蒙特卡洛模拟生成随机情景来分析风和负载的不确定性。此外，还考虑了储能系统的能量容量退化。最后，基于不同基准模型和最新模型的比较，验证了所提出方法的有效性。