Adaptive control of deep reinforcement learning and state prediction for smart microgrids in industrial parks

View at publisher Summarise with VIRAT
Authors: Yueping Xiang, Haoyuan Ma, Hao Luo, Jieying Liu, Yanqiu Hou, Tao Luo
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

When smart microgrid in industrial parks is connected with high proportion of renewable energy, it often faces the problems of lagging control response and poor robustness. In this article, an adaptive control framework combining state prediction and deep reinforcement learning is presented. By constructing the LSTM multivariable prediction model based on attention mechanism, and inputting its coding as an enhanced state into the Soft Actor-Critic (SAC) agent, the integrated control of "perception-prediction-decision" is achieved. The simulation on IEEE 13-node improved system shows that this method can reduce the maximum frequency deviation to 0.22 Hz, which is about 37% lower than that of traditional MPC. The average daily operating cost is reduced to 1496 yuan, which is more than 10% better than the comparison scheme; At the same time, it can make the frequency of energy storage action reasonable, and the number of equivalent cycles is stable at 1.9 times/day. The results show that embedding the prediction information into the state space is effective to improve the forward-looking and comprehensive performance of the control. This study provides a feasible technical path for the autonomous regulation of microgrid in complex and uncertain environment.

Fields of Study & Subjects

Microgrid Control and Optimization Frequency Control in Power Systems Smart Grid Energy Management

Publication details

Journal / source
IET conference proceedings.
Volume
2025
Issue
47
Pages
862–867
Publication date
1 March 2026
Language
EN