HARNESSING DEEP LEARNING FOR ACCURATE LOAD FORECASTING IN CLUSTER MICROGRIDS

Abstract

In DC microgrids, the inherent variability of renewable energy sources (RES) poses challenges to maintaining continuous operation and voltage stability. This paper introduces a distributed forecast-based consensus control strategy designed to balance the state of charge (SoC) levels of energy storage systems (ESSs) across the microgrid. The proposed approach integrates load-supply forecasts to prioritize the charging and discharging of ESSs, thereby enhancing the microgrid's reliability and voltage stability. Each branch of the microgrid employs a long short-term memory (LSTM) deep neural network for adaptive load forecasting, which informs the optimal (dis)charging rates of ESSs to ensure operational continuity during periods of RES unavailability. To mitigate the large data demands of LSTM models, a distributed extended Kalman filter algorithm is utilized to expedite learning convergence. Experimental validation on a 380V DC microgrid hardware-in-the-loop test-bench confirms that the proposed control strategy successfully achieves its objectives, demonstrating improved microgrid endurance and voltage stability.