Comparison of Nonlinear Autoregressive Neural Networks Without and With External Inputs for PV Output Power Prediction

Abstract

Accurate and precise prediction of power output plays a great significance in power system industry, as it provides the basic outlooks and views for making future decisions in power system planning and operation. This paper used monthly and annual dataset of output power and solar irradiance from a stand-alone solar system to compare nonlinear autoregressive neural network (NARNET) and nonlinear autoregressive with External (Exogenous) input neural network algorithms in predicting (short term and long term prediction) the output power from photovoltaic (PV) system. During the execution of prediction, The Levenberg-Marquardt method was adopted and used for training. Comparison results reveals that NARNET and NARX neural network are both suitable for performing PV output power prediction, but NARX was found to be more accurate. NARNET best prediction result was achieved with MSE (Mean Square Errors) equal to 1.6153 and coefficient R equals to 0.47643 while for annual NARNET prediction model, the MSE and R were found to be equals to 0.84891 and 0.82244. For NARX monthly prediction model, the MSE and R were equal to 0.99206 and 0.53094 while for the case of NARX annual prediction model, the MSE and R were 0.8008 and 0.85248 respectively. The lower the MSE indicates few errors and the more R is close to unity indicates the better correlation and great relation between actual and predicted values.