Machine Learning Models for Electricity Generation Forecasting from a PV Farm

Accurate forecasting of the electricity generation from photovoltaic farms plays a significant role in their proper technical and financial management. Reliable forecasts enable management of inertia and frequency response during contingency events and proper planning of the spinning reserve of PV farms. In this work, six machine learning models applying Convolutional Neural Network, Extreme Learning Machine, Random Forest Regression, Gradient Boosted Regression, AdaBoosted Regression, and K-Nearests Neighbors Regression were proposed to forecast electricity generation from a 700 kW photovoltaic farm located in Poland. The models were developed based on four widely available meteorological parameters: ambient air temperature, cloudopacity, and relative humidity, and hour, day, month and year. The comparative performance of the model revealed that the gradient-boosted regression was the most reliable with the determination coefficient \(R^2\)  = 95.503%, the mean absolute error MAE = 15.003 kWh, and the root mean square error RMSE = 29.975 kWh.