Comparison methodology of short-term techniques for forecasting the production of electrical energy from photovoltaic solar plants

Abstract

There is increasing awareness about the need to use renewable energy sources to produce electricity. Among these sources, photovoltaic solar energy stands out due to its great growth in recent years, and the possibility that photovoltaic solar plants have of operating in on-grid and off-grid modes. In either case, it is important to have a forecast of the plant's electrical energy production. In this research, a methodology is proposed to compare the short-term forecasts of the electrical energy produced by a photovoltaic solar plant, using machine learning algorithms, time series analysis, and making use of the historical data of the production of the plant and climatic variables. Forecast models were obtained with data in both hourly temporal resolution and fifteen-minute temporal resolution. The performance metrics used in the evaluation of the models were: R2, RMSE, MAE, and MAPE. For the resolution of fifteen minutes, both models, artificial neural network, and K nearest neighbors, fulfilled the statistical assumptions, but the network model had the lowest values of the metrics. For hourly resolution, again the artificial neural network model had the best performance over the ARIMA model of time series analysis.