Predicting the performance of multi-media filters using artificial neural networks

Abstract

The impact of flow rate and turbidity on the performance of multi-media filtration has been studied using an artificial neural network (ANN) based model. The ANN model was developed and tested based on experimental data collected from a pilot scale multi-media filter system. Several ANN models were tested, and the best results with the lowest errors were achieved with two hidden layers and five neurons per layer. To examine the significance and efficiency of the developed ANN model it was compared with a linear regression model. The R2 values for the actual versus predicted results were 0.9736 and 0.9617 for the ANN model and the linear regression model, respectively. The ANN model showed an R-squared value increase of 1.22% when compared to the linear regression model. In addition, the ANN model gave a significant reduction of 91.5% and 97.9% in the mean absolute error and the root mean square error, respectively when compared to the linear regression model. The proposed model has proven to give plausible results to model complex relationships that can be used in real life water treatment plants.