Inertia Estimation Through Local Frequency Measurements and Convolutional Neural Networks

Increasing penetration of non-conventional energy sources into traditional power systems introduces variability in system inertia over time. Converter-dominated power grids exhibit dynamic behaviors that significantly differ from those of traditional synchronous systems, rendering existing power system models inadequate for precise inertia estimation. In this paper, a Convolutional Neural Networks (CNN)-based methodology is proposed for inertia estimation, utilizing local frequency measurements to provide an adaptive and robust solution. A pulse excitation signal is applied to induce system perturbations, with frequency deviations captured through a Phase-Locked Loop (PLL). The proposed CNN model demonstrated an accuracy of 94.35%, with estimated inertia values deviating by less than 10% from actual measurements and achieving a Root Mean Square Error (RMSE) of 0.3865. When validated on previously unseen frequency data, the model maintained a comparable performance, yielding an RMSE of 0.3996. This data-driven, model-free approach presents a substantial advancement over conventional inertia estimation techniques, particularly in addressing complexities introduced by renewable energy sources and their impact on system dynamics.