Prediction of grid-photovoltaic system output using three-variate ANN models

Abstract

This paper presents the prediction of total AC power output from a grid-photovoltaic system using three- variate artificial neural network (ANN) models. In this study, two-hidden layer feedforward ANN models for the prediction of total AC power output from a grid-connected photovoltaic system have been considered. Three different models were configured based on different sets of ANN inputs. In addition, each model utilizes three types of inputs for the prediction. The first model utilizes solar radiation, wind speed and ambient temperature as its inputs while the second model uses solar radiation, wind speed and module temperature as its inputs. The third model uses solar radiation, ambient temperature and module temperature as its inputs. Nevertheless, all the three models employ similar type of output which is the total AC power produced from the grid-connected system. Data filtering process has been introduced to select the quality data patterns for training process, making only the informative features are available. Thus, the regression analysis and root mean square error (RMSE) performance of each model could be enhanced. After the training process is completed, the testing process is performed to decide whether the training process should be repeated or stopped. Besides selecting the best prediction model, this study also exhibits some of the experimental results which illustrate the effectiveness of the data filtering in predicting the total AC power output from a grid- connected system. Each ANN model was tested with Levenberg-Marquardt training algorithm and scaled conjugate gradient training algorithm to select the best training algorithm for each model. Fully trained ANN model should later be able to predict the AC power output from a set of un-seen data patterns.