This thesis concentrates on short-term solar Photovoltaic (PV) power output forecasting at array and system levels. The analysis was conducted on three arrays around the world and one system level system in California. Array level power was found to have higher power fluctuation than system level power. Hence, the proposed array level forecasting involves a similar day-based data-preprocessing to deal with this fluctuation. The processed array level data was fed into a forecasting engine. A persistence model, an Auto Regressive Integrated Moving Average (ARIMA), a Radial Basis Function Neural Network (RBFNN) and a Least Squares Support Vector Machine (LS-SVM) model was used as a forecasting engine. This thesis also investigates the applicability of a number of established forecasting methods for system level solar power output forecasting. In particular, ARIMA, RBFNN and LS-SVM are examined and simulation results are provided.
Through simulation, the best array level forecasting accuracy is achieved by a forecasting tool which combines the proposed similar day method and persistence model. The proposed similar day method works better than similar day methods in the literature and the best array level forecasting tool generated a more accurate forecast compared to a autoregressive with exogenous input (ARX) model in the literature. Due to the lower fluctuation of system level power data, system level forecasting has a better forecasting accuracy. The best performance is achieved through ARIMA model.