Municipal water demand forecasting under peculiar fluctuations in population: a case study of Mashhad, a tourist city

摘要

Forecasting future water demands has always been of great complexity, especially in the case of tourist cities which are subject to population fluctuations. In addition to the usual uncertainties related to climate and weather variables, daily water consumption in Mashhad, a tourist city is affected by a significant different fluctuation. Mashhad is the second most populous city in Iran. The number of tourists visiting the city is subject to national and religious events, which are respectively based on the Iranian formal calendar (secular calendar) and the Arabic Hijri calendar (Islamic religious calendar). Since religious events move relative to the secular calendar, the coincidence of the two calendars results in peculiar wild fluctuations in population. Artificial neural networks (ANNs) are chosen to predict water demand under such conditions. Three types of ANNs, feedforward back-propagation, cascade-forward and radial basis functions, are developed. In order to track how population fluctuation propagates in the model and affects the outputs, two sets of inputs are considered. For the first set, based on evaluating several repetitions, a typical combination of variables is selected as inputs, whereas for the second set, new calendar-based variables are included to decrease the effect of population fluctuations; the results are then compared using some performance criteria. A large number of runs are also conducted to assess the impact of random initialization of the weights and biases of networks and also the effect of calendar-based inputs on improvement of network performance. It is shown that, from the points of view of performance measures and unchanging outputs through numerous runs, the radial basis network that is trained by patterns including calendar-based inputs can provide the best domestic water demand forecasting under population fluctuations.