Artificial neural network based water network state estimation tool for bangalore inflow system

The main aim of this work is to develop a software tool for water distribution system modeling by coupling Artificial Neural Network (ANN) with real time flow and pressure data from the system. This tool helps in predicting the future state of the system, i.e. the flow in every pipe, pressure at every node and reservoir levels, for a given set of sensor readings at the current time step. The study helps to perform an ANN based sensitivity analysis of the network, and it can be extended to sensor placement optimization and demand prediction. Here, we have utilized feed forward artificial neural network with three hidden layers to predict water level in Bangalore inflow model. To compensate for practical sensor error, random noises were added in the training data set. The objective of this work is to create a collection of ANN if there is a well know question (state) we can instantaneously answer with help of the model. Genetic algorithm was used to optimize the network architecture. Gradient descent, and resilient back-propagation were used as training algorithms. In this research work, it was observed that the computational cost of the ANN based model is less than that of classical modeling approaches and hence can be used to replace hydraulic based tools for system state estimation. In addition, the normalized root mean squared error of our best model is around 0.05, meaning that little information would be lost by replacing a classical model with the neural network model.