Feature selection of power system data for frequency prediction with BO-LightGBM-Boruta

Accurate grid frequency prediction is essential for the stable operation and optimized dispatch of modern power systems. However, efficiently identifying critical features that influence grid frequency prediction from high-dimensional, complex power system data remains a significant challenge. To enhance the accuracy of grid frequency predictions, a feature selection method that integrates Bayesian Optimization (BO), LightGBM, and the Boruta algorithm (BO-LightGBM-Boruta) was proposed. The data collected under various fault scenarios from the New England 10-machine 39-bus test system were used as experimental samples to train the grid frequency prediction models and evaluate their predictive performance, thereby verifying the effectiveness and practical value of the proposed feature selection approach. The performance of the BO-LightGBM-Boruta method is compared with Lasso regression and Recursive Feature Elimination (RFE) under three prediction models: E3D-LSTM, LSTM, and ConvLSTM. The results show that the BO-LightGBM-Boruta method provides significantly more accurate predictions than the other two feature selection approaches in grid frequency prediction tasks. Notably, the combination of the proposed method and E3D-LSTM, which achieves the highest accuracy across different fault conditions, demonstrates outstanding predictive performance.