Desiccant wheel system modeling improvement using multiple population genetic algorithm training of neural network

Since the physical modeling method for the desiccant wheel system (DWS) is complex and costly for calculations, the modeling method based on neural network (NN) gains attention for its simplicity and effectiveness. The previous NN models of DWS are mostly based on backpropagation (BP) NN and adopt the gradient searching method to obtain the weights and thresholds. However, the gradient searching method results in “overfitting” easily. In this paper, a novel neural network training algorithm, trainmpga, is proposed. The algorithm searches the optimal weights and thresholds of NN by making use of the multiple population genetic algorithm, thereby conquering the “overfitting” of the gradient searching method and the “prematurity” of the genetic algorithm. Meanwhile, related configurations of NN, such as parameters and framework, are studied. Finally, the proposed modeling method trainmpga proves to have high training and prediction accuracy in comparison to the training algorithms in the MatLab toolbox and has good application prospects.