Improved Machine Learning Model Selection Technique for Solar Energy Forecasting Applications

Abstract

Grid-Connected Photovoltaic System (GCPV) in Malaysia had become vital due to its usages and contribution to the community. One of the advanced technologies that have been implemented in the solar field is the forecasting of PV power output and comes with a great challenge to produce high accuracy. This paper focuses on developing a ranking system to evaluate the performance of selected machine learning models. In this paper four models are considered, namely Support Vector Machine (SVM), Linear Regression, Gaussian Process Regression (GPR), and Decision Tree. Utilizing high-resolution ground-based measurement of meteorological and PV system power output, evaluation metrics such as Root Mean Squared Error (RMSE), Mean Absolute Deviation (MAD), Mean Absolute Error (MAE), coefficient of determination (R²), and computation time have been recorded to evaluate the performance of machine learning forecasting methods. Results show that the computation time is the primary criterion that differentiates the performance of forecast models. Other statistical metrics show only marginal differences in terms of performance. The ranking system developed can serve as an indicator for solar power output forecasters to determine the best model for their application.