A Comparative Study on the Operational Effectiveness of Machine Learning Models in Solar Power Forecasting

Abstract

Accurate solar power forecasting is crucial for optimizing grid operations and balancing energy supply and demand. Due to the high variability of solar radiation, advanced machine learning methods are needed to enhance forecasting accuracy. This study compares three models: Extreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), and a single-hidden-layer Bidirectional Gated Recurrent Unit (BiGRU). XGBoost and LightGBM are decision tree-based boosting models known for fast training and high accuracy, while BiGRU is a recurrent neural network designed for time-series data but prone to overfitting. Experimental results show that XGBoost and LightGBM train significantly faster and achieve lower errors (Normalized Mean Absolute Percentage Error-NMAPE is lower than 5%), demonstrating superior generalization. In contrast, BiGRU exhibits overfitting with NMAPE equal to 23.986% and Root Mean Squared Error (RMSE) equal to 18,763.12 kW on June 30, 2021. Notably, on December 31, 2021, XGBoost and LightGBM closely followed actual power generation trends, whereas BiGRU struggled to capture variations, further indicating its generalization issues. The findings highlight XGBoost and LightGBM as more suitable models for solar power forecasting, providing valuable insights for researchers and engineers in power grid management.