A DRL-Aided Multi-Layer Stability Model Calibration Platform Considering Multiple Events

摘要

Maintaining accurate stability models for power system planning and operational analysis is of great importance. Calibrating problematic parameters using PMU measurements that work well for multiple events remains a challenging problem. To tackle the known issues, this paper presents a novel and generalized deep-reinforcement-learning (DRL)-aided platform for automated parameter calibration with an adaptive multilayer dueling Deep Q Network (D-DQN) algorithm that searches optimal parameter sets for multiple events simultaneously. This platform leverages state-of-the-art DRL algorithms and supports various types of stability models used in software vendors’ transient stability programs. To help improve the efficiency of parameter calibration, a hierarchical structure with coarse-fine layers and adaptive steps is adopted when training effective DRL agents. It provides a systematic way to calibrate stability model parameters, which can save tremendous labor efforts for maintaining model accuracy and complying with industry standards. The effectiveness of the proposed approach is verified through numerical experiments on a realistic power plant model considering multiple system events.