Application of Neural Networks for Estimating Nodal Outflows as a Function of Pressures in Water Distribution Systems

Abstract

Water distribution systems are susceptible to failures. It is common for a failure in the distribution system to cause a reduction in pressures resulting in reduced nodal flows to consumers. In order to predict the reduction in the levels of service as a result of the reduced flows, it is important to relate pressure changes with nodal outflows during failure events. Conventional network analysis models are generally demand driven and do not allow the nodal outflow to be adjusted due to reduction in pressure. Modified network analysis is required where pressure dependent outflow functions are used. However many shortcomings associated with the pressure dependent functions have been reported in the literature. In this paper a modified network analysis program is presented where nodal outflows are developed as functions of pressure and secondary network characteristics. Outflow is estimated by means of an Artificial Neural Network (ANN) that has been trained with extensive data on pressure, flows and secondary network characteristics for a selection of secondary networks. A multi layer perceptron network has been used in predicting the pressure dependent nodal flows. The neural network is incorporated into a network analysis model and the network is solved using numerical differentiation. The developed model has been tested on several networks and found to be performing well. The changes in flows in secondary network as a result of the changes in the network conditions are predicted and compared with micro level models.