WATER QUALITY MODELING USING ARTIFICIAL NEURAL NETWORK AT THE DEAD-END SECTIONS OF DRINKING WATER DISTRIBUTION SYSTEM

Abstract

Despite efforts in treating water in water distribution systems, water borne disease outbreaks persist as a result of 
undergoing factors within the treatment tank, one of which is the water condition at the dead-end section. Existing 
models have addressed these problems in the mains and distribution, but that of dead end is still lagging. Therefore, 
this research seeks to address issue of water deterioration at the dead-end section of water distribution system of Abuja, 
Federal Capital territory of Nigeria. The Artificial Neural Network (ANN) model was deployed, for its ability to 
accurately model data of both short and long duration so as to establish dependability of the governing water quality 
parameters. The ANN models developed for training, testing, validation, and overall analysis indicated no significant 
variation between the predicted outputs and the original data values, demonstrating strong predictive capability and 
reliability. Among the different RANDIV distribution ratios examined, the 70:15:15 ratio produced better results for 
the Bayesian Regularization (BR), Levenberg-Marquardt (LM), and Scaled Conjugate Gradient (SCG) algorithms 
compared to the 80:10:10 and 60:20:20 ratios, with the BR algorithm showing superior overall performance. The 
ANN model equation, Output = 0.68(Target) + 0.02, showed close agreement between predicted and observed residual 
chlorine values, with only about 10 out of 259 data points exhibiting deviations up to 0.45%, while the majority differed by approximately 0.01%. The model achieved high reliability with R² values of 0.81 for training and between 0.62 and 0.75 for testing, validation, and overall datasets. Furthermore, the model-generated initial residual chlorine value of 0.02 mg/l complied with the USEPA recommended standard of less than 0.1 mg/l. The BR algorithm also demonstrated the best computational efficiency with a performance value of 0.0092 achieved within only 10 epochs, while LM and SCG required over 1000 epochs. The Nash-Sutcliffe Coefficient of Efficiency of 0.9236 (92.36%) further confirmed the accuracy and effectiveness of the developed model.EPANET simulations of the water distribution network revealed no significant chlorine loss around the dead-end sections, thereby maintaining acceptable residual chlorine concentrations and minimizing water quality deterioration. Simulations conducted over 
critical periods of 24 hours, 144 hours, and 240 hours demonstrated that the distribution system consistently maintained recommended chlorine residual levels under varying flow velocity, demand, and head conditions. Residual chlorine concentrations at selected dead-end nodes remained within safe operational limits, with a maximum value of 0.44 mg/l, which is below the recommended limit of 0.5 mg/l. The findings indicate that the distribution system is capable of preventing the growth and regrowth of disease-causing microorganisms while ensuring safe water quality throughout the network.