Keywords

Abstract

Neural network performance;

 Hidden layers;

 Design and training processes;

 Mean squared error; 

Solar power station; 

Investigation;

 Assessment;

 Comparison

Neural networks have a profound impact on many real-life applications. In this paper, the influence of increasing the hidden layers of the neural network on its performance is investigated and presented. For this purpose, three structures of the neural network are designed. The first structure has only one hidden layer, the second structure has two hidden layers, and the third structure has three hidden layers. The inputs and the output of each neural network structure are the same. To design and train these structures, data are collected from a solar power station in Egypt. These data include the temperature of the solar photovoltaic module and the radiation which are the inputs of each neural network structure. In addition, the power of the photovoltaic module, which is the output of each neural network structure. The obtained data is 7200 samples and is divided into three different parts, the largest part for the structure train stage, part for the test stage, and the last part for the validating stage. The main aim of this division is to investigate the efficiency of the structure in different modes. The training of each structure is conducted by Levenberg-Marquardt technique. The mean squared error (MSE) value is the main parameter used to identify the completeness and the effectiveness of the train, test, and validating stages. In addition, the approximated error between the actual output and the predicted outputs by each neural network structure is calculated. Structures with four, five, and six hidden layers are also developed and investigated. The results show that the MSE value is decreasing with the increase of the hidden layers. The MSE values obtained using three, four, five, and six hidden layers are 0.01686, 0.01634, 0.01593, and 0.01586 respectively. Furthermore, the average value of the approximation error is very small and is 0.0396 using the three hidden layers. Therefore, the increase of the hidden layers of the neural network increases its accuracy and performance. The results of the proposed method are compared with previous related works from literature. The result of this comparison shows the superior accuracy of the proposed method.

© 2024 MIM Research Group. All rights reserved.