A Framework of Using Machine Learning Approaches for Short-Term Solar Power Forecasting

Abstract

Various machine learning approaches are widely applied for short-term solar power forecasting, which is highly demanded for renewable energy integration and power system planning. However, appropriate selection of machine learning models and data features is a significant challenge. In this study, a framework is developed to quantitatively evaluate various models and feature selection methods, and the best combination for short-term solar power forecasting is discovered. More specifically, the machine learning methods include the random forest, artificial neural network and extreme gradient boosting (XGBoost), and the feature selection techniques include the feature importance and principle component analysis (PCA). All possible combinations of these machine learning and feature selection methods are developed and evaluated for solar power forecasting. The best ensemble of machine learning methods and feature selection techniques is identified for solar power forecasting in Hawaii, US. Simulation results show that the XGBoost method with features selected by the PCA method outperforms the other approaches. In addition, the random forest and XGBoost models have rarely been used for short-term solar forecasting. This framework can be used to select appropriate machine learning approaches for short-term solar power forecasting and the simulation results can be used as a baseline for comparison.