Online Optimal Power Scheduling of a Microgrid via Imitation Learning
Online Optimal Power Scheduling of a Microgrid via Imitation Learning
复制标题
通过模仿学习的微电网在线优化功率调度
DOI：
--
复制
发表时间：
2022
影响因子：
9.6
通讯作者：
T. Lee
中科院分区：
文献类型：
--
作者：
Shuhua Gao;Cheng Xiang;Ming Yu;K. T. Tan;T. Lee
本文研究了具有多种分布式能源的微电网的经济运行。考虑到可再生能源发电的间歇性以及市场价格和负荷的高随机性，在线电力调度方法利用实时信息处理不确定性的能力被普遍采用。传统的在线方法，如模型预测控制，需要一个单独的预测器，而最近基于强化学习（RL）的方法可以直接从历史数据中学习策略。然而，RL方法经常受到由连续状态和动作空间、复杂约束和缓慢训练引起的维度问题的困扰。我们提出了一种新的基于模仿学习的数据驱动在线方法，通过问题分解克服了这些限制，更重要的是，模仿混合整数线性规划（MILP）求解器而不是从头开始学习。由MILP专家演示的策略用深度神经网络逼近。我们的方法大大减少了训练时间，即使是在一个小的微电网中，与Q-learning方法相比，实现了17倍的加速。此外，在各种不确定因素的影响下，我们的方法所获得的运行成本接近于理论最小值。对模拟和真实数据的大量数值研究表明，与其他常用方法相比，所提出的方法具有性能优势。
This paper investigates the economic operation of a microgrid with a variety of distributed energy resources. Given the intermittency of renewable generation and the high stochasticity in market prices and loads, online power scheduling approaches are generally preferred for their uncertainty handling capacity by exploiting real-time information. Traditional online methods like model predictive control require a separate forecaster, while recent reinforcement learning (RL) based methods can learn a policy from historical data directly. However, RL methods often suffer from dimensionality issues arising from the continuous state and action space, complex constraints, and sluggish training. We propose a novel data-driven online approach based on imitation learning instead, which overcomes these limitations through problem decomposition, and more importantly, mimicking a mixed-integer linear programming (MILP) solver rather than learn from scratch. The policy demonstrated by the MILP expert is approximated with a deep neural network. Our approach reduces the training time dramatically even in a small microgrid, achieving a 17-times speedup in contrast to a Q-learning method. Moreover, the operation cost achieved by our approach subject to various uncertainties is close to the theoretical minimum value. Extensive numerical studies on both simulated and real-world data highlight the performance advantage of the proposed approach as compared to other common methods.