Short-term forecasting of solar photovoltaic output power for tropical climate using ground-based measurement data

Abstract

This paper highlights a new approach using high-quality ground measured data to forecast the hourly power output values for grid-connected photovoltaic (PV) systems located in the tropics. A case study using the 1-year database consisting of PV power output, global irradiance, module temperature, and other relevant variables obtained from Universiti Teknikal Malaysia Melaka is used to develop forecast models for three typical weather conditions - clear, cloudy, and overcast sky conditions. A machine learning method (Support Vector Regression - SVR) and an Artificial Neural Network method (nonlinear autoregressive) are used to produce the models and the results are compared with a benchmark model using the persistence method. Comparison with all the variables suggests that tilted global horizontal irradiance (GHItilt) and module temperature (Tmod) are the essential input variables to forecast the PV power output. It has also been observed that SVR performs well across all types of sky conditions, with the forecasting skill values between 0.65 and 0.79 when compared to the benchmark persistence method.