Robust ensemble forecasting and deep reinforcement learning for energy management on islanded microgrids
Y.F. Hsu, Yu-Hsin Hung, Chia‐Yen Lee
Abstract
Microgrids (MGs) are localized energy systems designed to integrate diverse energy sources efficiently. Energy management (EM) systems aim to optimize resource utilization, minimize waste, and enhance overall efficiency. This study examines EM strategies for an islanded MG comprising a solar photovoltaic (PV) system, a wind power system, a diesel generator (DG) set, and an energy storage system (ESS). A key aspect of the proposed model is its consideration of gradual changes required by dispatchable systems, such as DG sets, during engine start-up and power regulation—factors often overlooked in previous research. To address renewable and load variability, this study proposes a novel EM framework combining an improved adaptive robust optimization (ARO) ensemble forecast and deep reinforcement learning (DRL). An empirical study of Penghu Island in Taiwan was conducted to validate the proposed framework. The experimental results show that the improved ARO ensemble forecast demonstrates strong performance in forecasting energy loads and wind power, though it is less effective for solar power predictions. DRL models incorporating ARO forecasts significantly outperform those without forecasts, achieving a 3.4 times increase in average rewards and an 11% reduction in standard deviation. • Propose a novel framework integrating ensemble forecasting and reinforcement learning. • Integrate start-up and ramping characteristics of dispatchable devices into energy management. • Conduct an empirical study of Penghu Island in Taiwan to demonstrate the proposed framework. • Improve the resilience and adaptability on fluctuating energy demands and renewables supplies.