Artificial Intelligence for an Hour-Ahead Photovoltaic Output Power Forecasting

Photovoltaic (PV) energy integration in the power grid has increased continuously in recent years. However, this type of energy is highly dependent on climatic conditions. This problem complicates the power-dispatching plans of power generation plants in the case of large-scale grid-connected PV systems or the management of distributed energy resources for small-scale microgrids. To overcome this problem, PV power forecasting is crucial to ensure the power grid's stability and reduce energy costs for small microgrids end-users. In this regard, this paper presents six algorithms used to predict the output power of a photovoltaic generator (PVG) one hour ahead. Five algorithms are based on machine learning methodologies (Simple Regression Tree, Polynomial Regression, Gradient Boosted Trees, Linear Regression, and Random Forest), and the last one is based on a deep learning model (Multilayer Perceptron (MLP)). In order to optimize the predictive models, Grid Search K-fold Cross-Validation technique was used for hyperparameter tuning. In this paper, forcasters' input parameters are the PVG output power for the previous 24 hours and the mean daily of solar irradiance, ambient temperature, and wind speed. Obtained results show the superiority of MLP and random forest over the other methods to forecast PV output power an hour ahead with an accuracy of more than 0.91 R-squared and mean squared error of less than 0.7%.