Forecasting Solar Power Generation: A Comparative Analysis of Machine Learning Models

This study aims to point out accurate machine learning (ML) prediction methods to forecast solar energy generation. We analyze a dataset with 8,760 rows of data and 6 variables: Wind Speed (i), Sunshine (ii), Air Pressure (iii), Air Temperature (iv), Relative Air Humidity (v), and System Production (vi). A year of hourly data (01-01-2017 - 31-12-2017) is used. We compare the accuracy of the linear regression, the decision tree, the random forest, and the XGBoost model. We measure the accuracy using MSE, RMSE, MAE, and the coefficient of determination. We expect the XGBoost and the random forest to provide more accurate results because these models have fewer issues with over-fitting. Our findings indicate that the random forest and the XGBoost provide the highest accuracy levels. We decide to conduct two forecasts using the XGBoost model. The first model predicts the most accurate output for power generation for the next 31 days. We find that 1310.03 kW is the highest possible value, while 628.50 kW is the lowest possible value. Our model regarding power plant production forecasting finds that hour 12 and 13 show the highest production.