Many water utility companies are beginning to amass large volumes of data by means of remote sensing of flow, pressure and other variables. For district meter area monitoring there has been increasing interest in using this sensor data for abnormality detection, such as the real-time detection of bursts. Research pilots have explored systems for generating 'smart alarms' and a key requirement is usually a prediction of future time series values. Artificial neural networks have been employed in this capacity, however built in temporal memory in the network architecture (tap delays, feedback etc.) has not been widely explored. In this comparative study, a number of artificial neural network architectures are evaluated for water distribution flow time series prediction, in particular by exploring using temporal memory. These models included multilayer perceptron, mixture density network, time delay network and recurrent network. In addition the mean diurnal cycle (calculated from the data set) was utilised as a baseline prediction. Genetic algorithm optimisation was utilised in some cases to optimise the number of hidden processing elements and the learning rates parameters for the neural network. Two reference data sets are used as a case study originating from typical real world distribution systems and the performance assessed by means of mean absolute error. The results of the study show that of the static networks, the mixture density network is superior for repeatability and insensitivity to parameter settings. Similarly, the recurrent network is generally superior to the time delay network in this capacity. However, the use of either time delays or feedback results in approximately 50% less error than a static network for the best performing networks.
CORE (COnnecting REpositories)
CORE (COnnecting REpositories)