Forecasting hourly short-term solar photovoltaic power using machine learning models

Abstract

Forecasting solar photovoltaic power ensures a stable and dependable power grid. Given its dependence on stochastic weather conditions, predicting solar photovoltaic power accurately demands applying intelligent and sophisticated techniques capable of handling its inherent nonlinearity and volatility. Controlling electrical energy sources is an important strategy for reaching this energy balance because grid operators often have no control over use patterns. Accurately forecasting photovoltaic (PV) power generation from highly integrated solar plants to the grid is essential for grid stability. This study aims to improve forecasting accuracy and make accurate predictions of solar power output from the selected grid-connected PV system. In this study, the weather data was collected on-site and recorded PV power from a 20 kW on-grid system for one year, and different machine learning techniques like deep neural networks, random forests, and artificial neural networks were evaluated and benchmarked against reference support vector regression model. With improvements in forecasting accuracy of 2 to 37% over the reference model at study location (22.780 N, 73.650 E), College of Agricultural Engineering and Technology, Anand Agricultural University, Godhra, India, simulation results showed that the random forest technique is effective for the forecasting horizons of 1 to 4 hours.