Abstract
In an scenario where water resources are limited and with a developing society demanding water with more and more guarantees every day, engineering
is required to develop efficient techniques and methodologies to assure that the vital liquid is delivered in optimum conditions of quality and
quantity for the domestic, commercial and industrial users that integrate the set of subscribers of a city. Every type of users demands water at different time
and quantity scales, but the set of them consuming water jointly generates a city’s global water demand.
Potable water demand and distribution facilities operators are obligated to manage its operations such that all subscribers receive the service at the
moment they demand it. Experience accumulated in operations personnel becomes fundamental to achieve this objective, because they are capable
to predict with great precision future water demands.In the search for forecastings with strong mathematic and statistic foundations,
we have developed this work where the more important methodologies that have been used in the last decades to model and predict urban
water demand in densely populated areas were reviewed. It was found that stochastic models of the type of ARIMA are the bases for the more important
reviewed methodologies. However, we also found that existing models developed and thought for cities where water demand shows a cyclic pattern
with little variability along the annual cycle, and is only affected mainly by climatic and meteorological components. This is not the case for many European,
Spanish and Mediterranean cities, that show great variability derived from sociologic patterns and where climatic components are little relevant.
This variability is generated for punctual events that disturb the water demand process and that when these occur the expected repetitive patterns change.
The ignorance of these events in a short-term water demand forecasting scenario with stochastic models, first result, in wrong predictions at the moment
of occurrence of a punctual disturbing event, and second in, distorted predictions until a determined order after the occurrence of the event. It is expected
that the ignorance of these types of events diminishes the efficiency of the stochastic models.
Along this work, neural networks methodology was also applied to evaluate its efficiency to model and predict water demand. This performance is
compared with the aforementioned models, finding that it is possible to obtain similar results with both methodologies even when their origins are from
different points.
In this thesis work a methodology is proposed to incorporate implicitly in a stochastic prediction model the group of sociologic events that disturb
the water demand process. The methodology is tested in a real case for the city of Valencia, Spain, finding predictions performances that overcome the
results obtained not only by ARIMA models but also by neural networks. The errors obtained are very close to a white noise with a minor residual variance,
which tells us that the proposed methodology captures not only the systematic variability of the series, but also the irregular variability generated by
sociological patterns.
The proposed forecasting prediction outline has shown to be a good tool to predict short-term water demand for the analyzed case, and could easily
be implemented in a system operating in real time.