Optimization Scheduling Strategy with Multi-Agent Training Data Rolling Enhancement for Regional Power Grid Considering Operation Risk and Reserve Availability

Abstract

This study proposes an optimal dispatching method for regional grids, considering reserve availability and operational risks, while incorporating renewable energy and load uncertainties during day-ahead dispatch. The approach uses rolling augmentation of multi-intelligence training data, ensuring effectiveness and economy in intra-day scheduling, while improving model convergence speed and accuracy. First, a novel approach is proposed to measure system scheduling risk using Conditional Value-at Risk (CVaR). The proposed approach employs the Conditional Generative Adversarial Network (CGAN) to generate novel sets of load and energy output scenarios along with admissible error bounds. By utilizing these sets and intervals, the proposed approach can accurately and efficiently estimate the scheduling risk of the system. A day- ahead optimization model is proposed to minimize system operation cost, including risk cost, while optimizing active scheduling and backup plans to ensure system economy and robustness based on limit scenarios. To improve the effectiveness of the training data for the Multi-agent Proximal Policy Optimization (MAPPO) intra-day scheduling model, the dataset is enhanced using CGAN and updated daily on a rolling basis, optimizing the model's training effect. During the intra-day phase, the intra-day dispatch model utilizes ultra-short-term forecast data as input to generate real-time dispatch plans for standby units. The proposed approach is validated for its feasibility and effectiveness through experiments conducted on the IEEE39 node system.