Many indicators are involved in monitoring water quality. For instance, the fecal indicator bacteria are extremely important to detect the water quality. For this purpose, to better predict the total coliforms at the outlet of a Multi-Soil-Layering (MSL) system designed to treat domestic wastewater in rural areas, a neural network model has been developed and compared with linear regression model. The data was collected from the raw and treated wastewater of a three MSL systems during a one-year period in rural village, in Al-Haouz Province, Morocco. Fifteen physicochemical and bacteriological variables have undergone feature selection to select the best ones for predicting the total coliforms concentration in the effluent of MSL system. Furthermore, 80% of the available dataset were used to train and optimize the neural model using repeated cross validation technique. The remaining part (20%) was used to test the developed model. The neural network indicated excellent results compared to the linear regression. The optimal model was a neural network with one hidden layer and 11 neurons, where the R2 was about 97%. The importance analysis of each predictor was established, and it was found that pH and total suspended solids had the greatest influence on the total coliforms removal.
Keywords: Feature selection; Multi-soil-layering system; Multiple linear regression; Neural network; Total coliforms.
Copyright © 2020 Elsevier Inc. All rights reserved.