Applying artificial neural networks for systematic estimation of degree of fouling in heat exchangers

Abstract

Deposition of undesired materials on the heat transfer surface is one the most challenging problems for application of heat exchangers. Experimental measurements of degree of fouling are both difficult and time-consuming, and often do not provide accurate results. To overcome these problems, artificial neural networks (ANN) is employed for predicting the fouling factor from some easily measured variables of the system. Indeed, fouling factor is estimated as a function of density, velocity and temperature of the fluid, its oxygen content, hydraulic diameter of the fluid passage, surface temperature, and time. Correlation matrix analyses justified that the highest interrelation exists between these independent variables and fourth roots of fouling factor.The ANN model was developed and validated using a huge databank including 11,626 experimental datasets for fouling factor in portable fouling research unit (PFRU) and single tube heat exchangers collecting from six different literatures. The best training algorithm and the optimum numbers of hidden neuron were determined through minimizing the computational effort and maximizing some statistical accuracy indices, respectively. It was concluded that Bayesian regulation backpropagation approach has the best performance among the considered training algorithms. Moreover, the two-layer perceptron neural network with ten hidden neurons was found as the best ANN topology. This ANN model predicts the experimental values of fouling factor with overall AARD%=5.42, MSE=0.0013, RMSE=0.0355, and R2=0.977819. The simplicity of the developed ANN model and its small levels of error for huge experimental databank are some of the key features of our model.