Solar Photovoltaic Power Output Prediction Using Machine Learning-Based Regressors

Abstract

This study proposes a framework for predicting solar photovoltaic (solar PV) power output using Machine Learning-based regressors for short-, medium-, and long-term prediction horizons. To identify the most effective regressor, we propose a comparison framework to evaluate the performance of several types of regressor models. This evaluation will include Neural Networks, Boosting and Bagging Ensembles, and a baseline assessment using a linear regressor family. In this study, we implement the grid search method to improve model performance by fine-tuning hyperparameters, as does the K-fold shuffle split cross-validation method. We consider large spatial and long temporal historical datasets for the case study. A 5 km x 5 km gridded hourly temporal-based 1 MW modelled Solar PV dataset consisting of direct and diffuse irradiation, temperature, and power output during 2013-2022 in the Java-Bali region, Indonesia, is used as a case study. The grid search-optimized Neural Networks family, the Multilayer Perceptron model, can accurately predict power output from short-, medium-, and long-term horizons, with an average MAE of 0.248 kW and an average RMSE of 0.306 kW, followed by Random Forest, a grid search optimized Bagging Ensemble and a grid search-optimized Histogram Gradient Boosting Ensemble model. All predictor models generally performed well under strong El-Nino-affected data but were sensitive to very strong El-Nino during 2015-2016. The method used and insights gained from this study also benefit other jurisdictions with similar contexts.