标题： 基于安全模仿学习的配电网最优储能系统调度 显示英文标题

标题： Safe Imitation Learning-based Optimal Energy Storage Systems Dispatch in Distribution Networks

摘要： 分布式能源资源（DER）的集成加剧了配电网电压幅值调节的挑战。传统的基于模型的方法依赖于复杂的顺序数学公式，难以满足实时操作需求。深度强化学习（DRL）通过利用配电网模拟器进行离线训练，然后进行实时执行，提供了一种有前景的替代方案。然而，由于探索效率有限，DRL算法往往收敛到局部最优解。此外，DRL算法无法强制执行电压幅值约束，在配电网运行中实施时可能导致潜在的操作违规。本研究通过提出一种新颖的安全模仿强化学习（IRL）框架来解决这些挑战，该框架结合了IRL和设计的安全层，旨在优化主动配电网中储能系统（ESS）的操作。所提出的安全IRL框架包括两个阶段：离线训练和在线执行。在离线阶段，使用NLP求解器收集最优的状态-动作对，指导IRL策略迭代。在在线阶段，由安全层调整训练好的IRL策略的决策，以保持安全性和约束合规性。仿真结果表明，安全IRL在平衡操作效率和安全性方面具有有效性，消除了电压违规，并在各种网络规模下保持了低操作成本误差，同时满足了实时执行要求。 显示英文摘要

摘要： The integration of distributed energy resources (DER) has escalated the challenge of voltage magnitude regulation in distribution networks. Traditional model-based approaches, which rely on complex sequential mathematical formulations, struggle to meet real-time operational demands. Deep reinforcement learning (DRL) offers a promising alternative by enabling offline training with distribution network simulators, followed by real-time execution. However, DRL algorithms tend to converge to local optima due to limited exploration efficiency. Additionally, DRL algorithms can not enforce voltage magnitude constraints, leading to potential operational violations when implemented in the distribution network operation. This study addresses these challenges by proposing a novel safe imitation reinforcement learning (IRL) framework that combines IRL and a designed safety layer, aiming to optimize the operation of Energy Storage Systems (ESSs) in active distribution networks. The proposed safe IRL framework comprises two phases: offline training and online execution. During the offline phase, optimal state-action pairs are collected using an NLP solver, guiding the IRL policy iteration. In the online phase, the trained IRL policy's decisions are adjusted by the safety layer to maintain safety and constraint compliance. Simulation results demonstrate the efficacy of Safe IRL in balancing operational efficiency and safety, eliminating voltage violations, and maintaining low operation cost errors across various network sizes, while meeting real-time execution requirements.