A Machine Learning-Based SVG Parameter Identification Framework Using Hardware-in-the-Loop Testbed

Static Var Generator (SVG) can effectively compensate reactive power for maintaining acceptable bus voltage levels before and after major disturbances. Accurate model parameters of SVG controllers are essential to ensure reliable simulation of SVG dynamic behavior for power system planning and operational decision makings. Targeting the known issues of traditional parameter identification methods, this paper presents a novel machine learning-based framework of SVG parameter identification using actual measurement data collected from the hardware-in-the-loop (HIL) testbed. Two types of state-of-the-art algorithms are adopted to optimize SVG parameters so that the model performance can better match actual measurements. First, the actual measurements of SVG in various cases are obtained through the RTDS HIL testbed. Then, the trajectory sensitivity analysis of SVG controller parameters is carried out to identify the key parameters for calibration. Next, the parameter calibration problem can be solved by the convolutional neural network (CNN) and soft actor-critic (SAC) algorithms. The effectiveness of the proposed framework is verified on an actual SVG device using RTDS measurements, which outperforms the results obtained by the particle swarm optimization (PSO) algorithm in both accuracy and speed.