Evaluation of the Optimal Features and Machine Learning Algorithms for Energy Yield Forecasting of a Rural Rooftop PV Installation

The stability and reliability of the electric grid strongly depend on the ability to schedule and forecast the energy output of all sources. Even though the share of photovoltaic installation in the energy mix is continuously increasing, they have one major drawback: their dependence on different environmental parameters, such as solar irradiance, ambient temperature, cloudiness, etc., which have a highly variable nature. Six machine learning algorithms are compared in this study, regarding their ability to forecast the power generation of a rural rooftop photovoltaic installation using different combinations of the input data. The features selected for investigation are solar radiation, ambient temperature, and wind speed, obtained from a meteorological station, as well as two additional time-based variables - the time of the day and the month of the year. During the validation and testing phases, four models performed better - artificial neural network (ANN), kNearest neighbor (kNN), Decision tree (DT), and Random Forest (RF), with ANN achieving the best results in all cases. The optimal combination of input data includes solar radiation, ambient temperature, wind speed, and hour of the day, though the difference with the other scenarios was small. The optimal ANN model achieved R-2, MAE, and RMSE of 0.995, 6.71 Wh, and 13.7 Wh, respectively. The results obtained in this study indicate that the yield of PV installations located in rural areas could be forecasted with high probability using a limited number of meteorological data.