Application of artificial neural networks to forecasting water quality in a chloraminated water distribution system

Abstract

Accurate forecasting of disinfection residuals in a water distribution system (WDS) is essential for optimal control of disinfectant dosing in order to maintain good water quality within the system. The most commonly used disinfectant in the drinking water treatment process throughout the world is chlorine. Previously, artificial neural networks (ANNs) have been used successfully to predict chlorine residuals in chlorinated WDSs. Recently, an increasing number of water utilities are using chloramine as a secondary disinfectant in WDSs. Chloramine has several advantages compared with chlorine including fewer issues with disinfection by-products (DBP) and longer persistence. However, limited research has been conducted into the modelling of disinfection residuals (chlorine and free ammonia) in a chloraminated WDS. In this research, a general regression neural network (GRNN) has been used to forecast chlorine and free ammonia residuals in one section of the Goldfield and Agricultural Water System (G&AWS) east of Perth. In this study, the system under investigation spans from the original chlorine and ammonia dosing location at Mundaring pump stations to Goomalling pump station. Ten water quality variables and six flow variables have been used for the development of both chlorine and free ammonia forecasting models. The data were collected at 10 minutes intervals over a one-year period from February 2009 to January 2010. They have been converted into hourly data for this study. A maximum of 94 lags are used, which result in a total of 1,536 potential inputs for each model. In order to select appropriate inputs for each of the models within a reasonable time, a three-step input selection procedure using both mutual information (MI) and partial mutual information (PMI) has been used, which results in 12 selected inputs for the chlorine forecasting model and 10 selected inputs for the free ammonia forecasting model. A deterministic data splitting method called DUPLEX has been used to divide the data into the training (60%), testing (20%) and validation (20%) sets. The modelling results indicate that ANNs have the ability to forecast chlorine residuals in a chloraminated WDS. This provides a useful tool that can be used to assist in optimal control of disinfectant levels in WDSs. However, further information on both the input variables and the hydraulics of the system is required in order to improve the performance of the chlorine forecasting model. In contrast, the free ammonia forecasting model performed poorly and cannot be used to provide accurate forecasts of free ammonia levels in the system. This is mainly because the free ammonia data collected were estimated based on the difference between the total ammonia nitrogen and the monochloramine nitrogen, and therefore are not accurate. As a result, accurate free ammonia analysers are required in order to obtain precise free ammonia data for the development of an accurate ANN model for the forecasting of the free ammonia levels in a WDS.