Cooperative Dispatch of Microgrids Community Using Risk-Sensitive Reinforcement Learning with Monotonously Improved Performance

TL;DR

This paper introduces a novel risk-sensitive reinforcement learning algorithm for microgrid cluster dispatch, balancing multiple objectives and uncertainties to improve reliability, efficiency, and computational speed in energy management.

Contribution

It proposes a new multi-objective, risk-sensitive reinforcement learning method (RS-TRPO) for microgrid dispatch, formulated as a Markov game and demonstrated to outperform traditional methods.

Findings

RS-TRPO achieves faster computation than mathematical programming.

The algorithm effectively balances multiple objectives and risk mitigation.

Demonstrated superior performance on IEEE 30-Bus Test System.

Abstract

The integration of individual microgrids (MGs) into Microgrid Clusters (MGCs) significantly improves the reliability and flexibility of energy supply, through resource sharing and ensuring backup during outages. The dispatch of MGCs is the key challenge to be tackled to ensure their secure and economic operation. Currently, there is a lack of optimization method that can achieve a trade-off among top-priority requirements of MGCs' dispatch, including fast computation speed, optimality, multiple objectives, and risk mitigation against uncertainty. In this paper, a novel Multi-Objective, Risk-Sensitive, and Online Trust Region Policy Optimization (RS-TRPO) Algorithm is proposed to tackle this problem. First, a dispatch paradigm for autonomous MGs in the MGC is proposed, enabling them sequentially implement their self-dispatch to mitigate potential conflicts. This dispatch paradigm is then formulated as a Markov Game model, which is finally solved by the RS-TRPO algorithm. This online algorithm enables MGs to spontaneously search for the Pareto Frontier considering multiple objectives and risk mitigation. The outstanding computational performance of this algorithm is demonstrated in comparison with mathematical programming methods and heuristic algorithms in a modified IEEE 30-Bus Test System integrated with four autonomous MGs.