Intelligent reinforcement training optimisation of dispatch strategy for provincial power grids with multi-agent systems: Considering operational risks and backup availability

Abstract

In order to optimise resource allocation within the province, a two-stage scheduling model for provincial-level power grids, encompassing day-ahead and intra-day stages is proposed. Firstly, a Conditional Generative Adversarial Network is employed to generate scenarios for load and new energy output. Based on the generated scenario set, the model takes into account the uncertainty and permissible error intervals of new energy and load, utilising conditional value at risk to measure the system scheduling risk. In the day-ahead stage, an optimisation model is proposed, considering intra-provincial power purchase demands, with the goal of minimising system operating costs, including risk costs. It optimises day-ahead scheduling and contingency plans to ensure economic efficiency and robustness of the system based on extreme scenarios. During the training phase, the dataset is enhanced using Conditional Generative Adversarial Network and updated daily, improving the training effectiveness of the multi-agent proximal policy optimisation intra-day scheduling model. In the intra-day stage, the intra-day scheduling model utilises ultra-short-term forecasting data as input to generate contingency plans for dispatching reserve units. Experiments conducted on the IEEE 39-node system validate the feasibility and effectiveness of the proposed approach.