Autoencoder networks for water demand predictive modelling

Following a number of studies that have interrogated the usability of an autoencoder neural network in various classification and regression approximation problems, this manuscript focuses on its usability in water demand predictive modelling, with the Gauteng Province of the Republic of South Africa being chosen as a case study. Water demand predictive modelling is a regression approximation problem. This autoencoder network is constructed from a simple multi-layer network, with a total of 6 parameters in both the input and output units, and 5 nodes in the hidden unit. These 6 parameters include a figure that represents population size and water demand values of 5 consecutive days. The water demand value of the fifth day is the variable of interest, that is, the variable that is being predicted. The optimum number of nodes in the hidden unit is determined through the use of a simple, less computationally expensive technique. The performance of this network is measured against prediction accuracy, average prediction error, and the time it takes the network to generate a single output. The dimensionality of the network is also taken into consideration. In order to benchmark the performance of this autoencoder network, a conventional neural network is also implemented and evaluated using the same measures of performance. The conventional network is slightly outperformed by the autoencoder network.