Data-Driven Model Predictive Control-Based Proactive Scheduling for Commercial Microgrid Considering Anomaly Detection

Abstract

One of the challenging tasks for commercial microgrids with distributed generation is to ensure critical loads resilience during high-impact, low-frequency extreme events. This article proposes a data-driven model predictive control-based proactive scheduling strategy for microgrids subject to extreme weather events. The goal is to maximize the utility value of the building by realizing the difference between the comfort level of critical loads and the total operating cost of the commercial microgrid during the planning horizon. A two-layer framework is proposed to solve the considered problem. First, the aggregate demand load measured by smart meters is processed using a combined long short-term memory and fuzzy neural network algorithm to eliminate anomaly data. Then, the obtained real data are applied through the deep-Q network method to seek appropriate proactive scheduling solutions. However, the rewards obtained from the deep-Q network method are random because of the uncertainties in extreme events. Therefore, robust optimization namely, robust deep reinforcement learning, is incorporated to enhance the robustness of the solutions. The proposed framework ensures the optimality and feasibility of the commercial microgrid's proactive scheduling strategy within the uncertainties in extreme events and the different intensity levels of anomalous data. Extensive simulation results prove the effectiveness of the proposed optimization model and algorithm for enhancing commercial microgrid resilience in extreme events. Furthermore, it also shows good ability in detecting smart meters anomalous data, with a stable convergence value of the performance index, to secure commercial microgrid smart meter data.