PV generation forecasting utilizing a classification-only approach

Abstract

The increasing use of photovoltaic (PV) systems in electricity infrastructure poses new reliability challenges, as the supply of solar energy is primarily dependent on weather conditions. Consequently, to mitigate the issue, enhanced day-ahead PV production forecasts can be obtained by employing advanced machine learning techniques and reducing the uncertainty of solar irradiance predictions through statistical processing. The objective of this study was to present a methodology for accurately forecasting day-ahead PV production using novel machine learning techniques and a classification-only forecasting approach. Specifically, the central component of the proposed method is a classifier model based on an Extreme Gradient Boosting (XGBoost) ensemble algorithm that classifies the respective daily 30-min profiles of the forecasted global horizontal irradiance (GHI), the measured incident irradiance (Gi), and the AC power (PAC) into a predetermined number of classes. The formed classifier model was used as a dictionary to designate the newly arrived forecasted GHI to a particular class and ultimately identify the corresponding forecasted PAC. The results demonstrated that the proposed forecasting solution provided forecasts with a daily normalised root mean square error (nRMSE) of 8.20% and a mean absolute percentage error (MAPE) of 6.91% over the test set period of one year, while the model's reproducibility was also evaluated and confirmed. Additionally, a comprehensive evaluation based on clear-sky index categories revealed that the model's performance was notably accurate on clear-sky days, while maintaining acceptable accuracy levels on moderate and overcast days. These findings underscore the versatility and robustness of the proposed methodology in handling diverse weather conditions and hold promise for improved PV production forecasts.