Quantification of Power Grid Frequency Regulation Capacity Demand Based on Deviation Prediction

Driven by global decarbonization goals, the rapid growth of renewable generation is increasing the frequency-regulation burden on power systems. Reliable estimates of regulation-capacity demand are therefore essential for coordinated energy and ancillary-services market clearing and secure grid operation. We develop a data-driven method that captures renewable variability while quantifying this demand accurately. The method uses renewable output, load fluctuations and meteorological conditions as predictors, with the maximum one-minute net-load deviation over each interval defining the target capacity. To address scale imbalance and redundant high-dimensional inputs, we introduce a dual screening procedure for training samples and features based on an improved weighted Euclidean distance. A deep neural network (DNN) then produces deterministic capacity forecasts, while adaptive-bandwidth kernel density estimation (ABKDE) models their residual errors. The upper confidence bound of the resulting prediction interval is used as the required regulation capacity. Tests on 2023 operating data from a provincial power grid yielded a spring-scenario mean absolute percentage error (MAPE) of 3.62% and a coefficient of determination (R2) of 98.31%. With ABKDE compensation, prediction-interval coverage probability reached 100%, while the prediction-interval normalized average width (PINAW) remained 7.46%. Performance remained robust across renewable-penetration levels and all four seasons.