Intra-day solar irradiance forecasting for PV power generation utilising machine learning models

Abstract

Accurate PV production forecasting is an important feature that can assist utilities and plant operators in thedirection of energy management and dispatchability planning. In this work, intra-day (1 to 3 hours ahead) solar irradiance forecasting utilising Support Vector Machines (SVM) is derived in order to feed to an Artificial Neural Network (ANN) trained for PV power generation forecasting (1 to 3 hours ahead). The study focuses on improving the accuracy of both the intra-day solar irradiance and power generation forecasting by employing machine learning models that could record the profile of the solar irradiance and the behaviour of the PV system. The performance of the SVM and ANN was assessed against a historical test set exhibiting normalised mean square errors (nRMSE) of 2.93% to 6.52% and 3.52% to 7.84% respectively, indicating that all the behaviour of the local irradiance as well as the behaviour of the system were efficiently recorded by the forecasting models.