Photovoltaic (PV) Power Prediction Based on Artificial Neural Network with Activation Function Selection and Feature Reduction Method

Abstract

This paper presents an extensive review of the ANN Based PV Power Output Prediction Model and the exploration of the effect of the common meteorological variables that is used in some research work. This study will discuss the result of the two simulated Models, wherein Model A uses all the parameter as the input variables and Model B applied the feature reduction method that explores all the possible reduced parameter combinations.
A data set consisting of 755 variables (PV power output model) were used to trained and test a 2- layer (1 hidden layer) neural network model. The study simulated two models A and B. The model A used the conventional method of modelling, training and testing using the six input variables such as solar irradiance (G), maximum temperature (Tmax), minimum temperature (Tmin), rainfall (Rf), wind speed (Ws), and relative humidity (Rh). After thorough simulation, the final neural network for model A with six input variables, with 8 hidden neurons, using tan sigmoid activation function, 1 layer and 1 output node. The coefficient value of the PV power model was R (All) = 0.89264, R(Test) = 0.89071, R(Training) = 0.88527, R(Validation) = 0.92738, MSE = 0.025118. For the final Model B, the best parameter combination is consisting of four variables the G, Tmax, Tmin and Rh with 10 hidden neurons, using tan sigmoid activation function, 1 layer and 1 output node. For model B, the results are R(All)= 0.9034, R(Test) = 0.87312, R(Training) = 0.8969, R(Validation) = 0.95613, MSE= 0.024645. Based on testing and validation of Model A and Model B, the MAPE are 44.06% and 19.88% respectively.
The study shows that the Model B using four input variables with solar irradiance, maximum and minimum temperature and relative humidity provides good forecasting results predicting solar pv power output, as justified by the result of its MAPE obtained from the validation and testing of data.