Drinking water utilities rely on robust long-term drinking water demand forecasts for adequate planning of production and storage capacity. Two key metrics for these purposes are prediction of (a) the average daily drinking water demand and (b) the extremes in daily drinking water demand [1
]. For the first metric, a common practice among water utilities is to develop long term (monthly, yearly) projections based on the analysis and extrapolation of autonomous trends that may influence drinking water demand in the future [2
]. Examples of such trends are: Demographic changes, economic development [3
] and in some cases climate change [4
]. A characteristic trait of these trends is that they generally develop slowly over the course of years. Their effect on water demand is therefore generally expressed as an annual growth rate.
Extremes in water demand typically unfold on a short timescale (days) and are stochastic in nature. They are therefore usually quantified with a so called peaking factor, a single probabilistic ratio expressing the peak daily water demand corresponding with a once in N year recurrence period (in the Netherlands and Belgium a once in 10 years period is used). Typically, investments in drinking water infrastructure are based on this factor (which is then multiplied by the average demand, which might in itself also increase or decrease in the future).
Given the long lifespan of infrastructure, failure to accurately estimate the peaking factor in the design phase of new infrastructure can lead to under- or over-estimation of capacity requirements. In practice, the peaking factor is often calculated directly from historical water demand time series, thereby not explicitly including any changes in the demand regime that may occur in the nearby future, such as climate change or evolving socio-economical dynamics. Our hypothesis is that this practice may lead to significant over- or under-estimation of future capacity requirements and therefore unforeseen costs. We argue that it would be better to not calculate the daily peaking factor directly from historical time series, but instead from forecasted water demand time series that are representative for the future period.
Most scientific research on water demand forecasting is focused on either the short-term (essentially predicting the water demand for the upcoming days in order to optimize operational control) or long-term (estimating water demand for the years to come). Whereas short-term forecasting is typically done on daily, hourly or even quarterly time steps, long-term forecasting usually leads to results with monthly or yearly simulation time steps [2
]. For the use case that we have in mind, we essentially need a combination of both approaches: We want to simulate water demand characteristics on a daily time step, but for future periods with a length of decades. Our goal is not to predict the exact water demand on a certain day many years ahead in the future, but rather to simulate water demand time series that are statistically representative for the future period using a probabilistic approach. The simulated time series can then be used to calculate the frequency of occurrence of extreme water demands.
As the peaking factor solely expresses the likelihood of water demand peaks within a year, it is not necessary to include all possible influencing trends in our analysis. Any trend that slowly influences water demand over the course of multiple years does not influence the peaking factor. For example, gradual population growth and economic growth generally increases the cumulative water demand in a region over the course of years. However, the intra-annual fluctuations in demand (relative to the annual mean) are not necessarily affected by such a trend. We therefore focus only on trends that have the potential to influence future intra-annual demand fluctuations: Climate change and changes in vacation absence/presence (tourism). The idea here is that by including these trends, we can arrive at a robust estimate for the peaking factor that water utilities can combine with their regular year-to-year forecasts of average water demand growth.
Although in this context climate change in itself can be considered a gradual process, resulting weather is broadly recognized as an important exogenous factor influencing daily drinking water demand [8
]. Therefore daily weather predictions are generally used as input for short-term water demand forecasting (essentially predicting the water demand for the upcoming days in order to optimize operational control). For example, Bakker and Van Duist [8
] show that change in temperature influences short-term forecast errors. In a study for the city Melbourne, Zhou and McMahon [11
] show how seasonal variation in water demand can be attributed to air temperature, evaporation and rainfall. Surprisingly, little research has been done on impacts of climate change on daily domestic and commercial drinking water demand. Nonetheless, the limited literature available on this topic suggests that climate change induced weather changes are an important factor. Not only for projections of average daily water demand, but also, and probably even more so, for projections of extreme daily drinking water demand [12
Recently, Toth and Bragalli [15
] highlighted the importance of tourism in demand modeling. According to Gossling and Peeters [16
], tourism is only a minor factor in global drinking water use, but a potentially important factor on smaller spatial and temporal scales, as tourism concentrates on traveler flow, and thus water demand, in time and space. This corresponds with the findings of Almutaz and Ajbar [17
], who successfully incorporated tourism fluxes in a water forecasting model. These findings, albeit scarce, indicate that vacation absence/presence patterns might help to explain peaks in water demand, and that ignoring them may result in under- or over-estimation of the effect of weather on peak drinking demand during summer months.
In this study, we present a modeling framework that allows quantification of the impact of climate change and variations in vacation absence on the peaking factor. The modeling framework consists of a machine learning model that predicts daily water demand as a function of meteorological parameters and vacation absence. Water demand time series simulated by this model are subsequently translated to a peaking factor using extreme value analysis. To the authors’ knowledge, no such framework has been proposed before and this is the first time climate change impacts on the daily drinking water demand peaking factor are quantified.
The presented modeling framework allows for simulating water demand on long timescales with a temporal resolution of one day. It enables evaluation of the impact of climate change and variations in vacation absence on both the average daily water demand and the peaking factor. We showed the effectiveness of this model by applying it to eight different supply areas in the Netherlands and Belgium.
We found that the average demand increased somewhere between −0.2% and +3.1%, while the peaking factor increased between −2.9% and 21.3%. Thus we can conclude that variations in climate change and vacation absence affect the peaks in water demand much more than the averages. Even though these numbers are specific to the supply areas that we studied and to the scenarios that were used, they provided an estimate for the change that we might expect in the years to come, and at a minimum pinpoint an order of magnitude for the change.
The model results clearly show how climate change and variations in vacation absence could have surprisingly different impacts on different supply areas. This suggests that the choice of geographic scale is important in such analyses and that, in order for meaningful insights and outcomes to be obtained from such an assessment, it is crucial that relatively small geographic units of analysis are selected.
Results also highlight the importance of accounting for vacation absence (or vacation presence in tourist areas). The modeling framework is generic: It can be applied to any supply area as long as (a) distribution losses are fairly constant and low and (b) multiple years of historic daily water consumption data, vacation absence rates and weather observations are available.