PHÂN TÍCH ẢNH HƯỞNG CÁC SIÊU THAM SỐ CỦA MÔ HÌNH LIGHTGBM ĐẾN DỰ BÁO CÔNG SUẤT ĐIỆN MẶT TRỜI

Abstract

This paper presents a study on the impact of certain hyperparameters in the LightGBM model on solar power generation forecasting accuracy. The considered hyperparameters include the maximum number of leaves in decision trees (num_leaves), learning rate (learning_rate), and the number of boosting rounds (n_estimators). Ten scenarios with different combinations of these hyperparameters were implemented and compared based on error metrics: RMSE, MAPE, and NMAPE, as well as training and inference time. The results show that adjusting these parameters could improve the forecasting performance of the model, as reflected in a slight reduction in forecasting errors for instance, the MAPE decreased from 90.67% to 82.94% when increasing num_leaves from 30 to 60. However, the improvements are insignificant, the error metrics only vary within a narrow range across scenarios. This indicates that the LightGBM model is relatively robust to changes in hyperparameters within the tested range, and moderate tuning of num_leaves, learning_rate, and n_estimators does not lead to dramatic changes in forecasting accuracy