PREDICTIVE MODELS OF RESIDUAL CHLLORINE CONCENTRATION IN WATER DISTRIBUTION SYSTEMS BY LONG-TERM MEMORY NETWORKS

Abstract

 Chlorine injection is strictly controlled so that the residual chlorine level of tap water in the supply and distribution system should be more than 0.1 mg/L. Currently, skilled operators decide the injection volume based on thier experience. However, due to retirement of a large number of veteran operator and a decrease in the number of staff, new methods for managing the injection volume of chlorine that do not depend on experience are required.

 This study focused on the time-series data of residual chlorine concentration possessed by water purification plants, and aimed to construct a predictive model that can infer residual chlorine concentration several hours ahead by learning the trend of change in concentration in time series. Specifically, we attempted to construct predictive models of residual chlorine concentrations at 3 hours, 6 hours, 12 hours, and 24 hours ahead using long-term short-term storage networks (LSTM) algorithms.

 We have found that the optimal block for building the model is 24 hours, and the prediction time needs to be 6 hours or less in order to keep the error target less than ±0.025. It was also found that the minimum amount of data required to construct the model was the amount of reduction in residual chlorine concentration for 4 months from April to July.

 From the above, it was found that the residual chlorine concentration 6 hours ahead can be predicted by LSTM if the time series data of 1 hour interval of residual chlorine concentration can be prepared for 4 months.