Look-Ahead Unit Commitment With Adaptive Horizon Based on Deep Reinforcement Learning
Jiahao Yan, Yaping Li, Jianguo Yao, Shengchun Yang, Feng Li, Kedong Zhu
Abstract
The highly variable nature of renewable energy has led to the concept of intra-day look-ahead unit commitment (LAUC), which aims to determine the on/off status and power outputs of generating units in a rolling-horizon fashion. LAUC is traditionally performed based on look-ahead horizon (LAH) with fixed length and resolution. Such practice can neither capture the high-risk time periods, nor achieve maximum computational efficiency. To address these issues, this article proposes a LAUC method with adaptive horizon (LAUC-AH). The method uses three parameters to describe the shape of LAH, namely length, resolution, and myopia. Taking these parameters, the forecast profile of renewable energy and load demand is aggregated using a hierarchical clustering procedure to form the LAH, which is then used to construct the LAUC optimization model. Furthermore, a deep reinforcement learning-based agent is used to dynamically adjust the parameters of LAH, such that the LAUC model can adapt to different operation statuses of the power grid. The case studies are carried out on a modified IEEE-118 test case to validate the feasibility and effectiveness of the proposed method.