Long-term water demand forecasting: lessons learned with Seattle Public Utilities

The highest priority that all water utilities share is the critical responsibility of providing safe and reliable drinking water to the residents and businesses they serve.

In meeting this charge, it is imperative that every utility has access to source water that will be treated and delivered to their customers. Forecasting the expected quantity of water that will be needed over the long term is an important aspect of managing water resources responsibly, particularly for areas experiencing significant changes in customer growth and behavior. Seattle and the Puget Sound region fit this description as being one of the fastest-growing parts of the U.S. for decades. Seattle Public Utilities (SPU) provides drinking water to 1.5 million people in the Seattle metropolitan area from two Cascade Mountain reservoirs. SPU sells approximately half of its water supply directly to retail customers and half to 24 wholesale water customers (regional cities and other water districts in the region). Raftelis recently partnered with Seattle Public Utilities to develop a new tool that can be used by staff to inform the long-term projection of water use. 

SPU had been using a water demand forecasting model that was developed in 2006 in support of its 2007 Water System Plan. The modeled forecast is reported to state agencies as a regulatory requirement and is used for water supply planning, conservation program evaluation, wholesale customer contract development, and retail rate setting. Though the 2006 model worked well, its forecasting information had been tailored based on institutional knowledge and needed to be revalidated and documented.  

This model used a Variable Flow Factor (VFF) approach, which is more complex than a trend analysis but more simplified than an econometric model. The 2006 model forecasted water demand by sector (single- and multi-family residential, non-residential) for Seattle and each of its individual wholesale customers at an annual timestep, using water demand factors known to drive consumption. SPU wanted to develop a new long-term demand model they could use for the same purposes but using current data and new techniques, and ensuring it is reflective of recent trends in the Puget Sound region.

Modeling Approaches 

At a fundamental level, demand forecast techniques fall into five broad categories: qualitative methods, time series, fixed flow factor/land use, VFF, and econometric models. These approaches vary in their complexity and cost to develop, which are depicted in Figure 1 below, and described relative to how they would be applied to water demand forecasting in the following paragraphs.

Figure 1. Demand Forecasting Methods   

Qualitative methods, as their name implies, rely on minimal mathematical analysis and more so on market research, panel consensus, visionary forecasting, and other approaches. The benefit of qualitative methods is the interpretability and simplicity of the forecast. The disadvantage of this model type is the lack of data-driven decisions and inherent margins for error in the forecast.   

A time series model uses historical values for any data to predict future values, i.e., the past is a predictor of what will happen in the future. These models account well for randomness, trend, and seasonality, and a fundamental benefit to this approach is the minimal data required to develop a forecast (in this case, it only requires historical water use information).  

A fixed flow factor model forecasts demand across different sectors and service areas using a set of consistent levels of per-unit consumption for various sectors (e.g., per capita or per single-family household). These models are generally easy to explain to stakeholders. However, they cannot readily account for the impact of external influences on water demand or allow for sensitivity analysis of these influences on a demand forecast. 

A variable flow factor (VFF) model uses a base water consumption and adjusts the usage established by factors known to influence demand, such as weather, price, etc., recognizing that the quantitative values of these uses influence change over time. This model type is generally easy to understand and allows for the inclusion of variables that have been used in other demand forecasts. SPU has used this model type for the past 15 years and it has worked well to address concerns and needs for the utility.  

Econometric models provide the ability to quantify how factors such as weather, socioeconomic conditions, and conservation measures affect water consumption locally and to generate forecasts for various future scenarios based on these factors. Econometric models rely on historical consumption and a set of quantified external factors and allow statistical software to estimate the influence each factor has on demand. These influences are built into an equation that forecasts water demand.   

Modeled Variables

With any demand forecasting study, it is important to perform a literature review to analyze independent and dependent variables that can be included in model development. These variables frequently include income, weather and seasonal factors, household structure and size, property characteristics, and conservation, among others. Of the influencing factors, some have an intuitive relationship with water demand, while others do not. In the case of SPU, during the rainy season in the Pacific Northwest, there is reduced demand for outdoor water use; conversely, hotter days in the summer tend to increase demand for outdoor water use. Economic prosperity and population growth can add customers and demand to the system. If these customers are moving into new non-irrigating high-rise apartments with low-flow appliances and fixtures, the actual incremental water use may be offset by such efficiencies. Furthermore, the past decades have brought changes in how water is used, along with a general recognition that water conservation and efficiency should be a focus. These recent influences will skew the effect of some variables.  

Results and Discussion 

The appropriate demand forecasting approach will vary for every utility, and Raftelis worked collaboratively with SPU staff to examine multiple methods. An econometric model was selected as an improvement over the variable flow factor approach since it allows for a more data-intensive approach to variable selection. Our process included an evaluation of more than 30 years of customer usage, climate, socioeconomic, and conservation data. Examples of the climate data that was explored as part of this process are presented in Figures 2 and 3, which demonstrate some clear warming trends in the region’s climate while precipitation displays consistent seasonality.   

Figure 2. Historical Temperatures

Figure 3. Historical Precipitation


After a long and careful quality assurance process, the data was used to inform the variable selection process. Selecting variables for inclusion in the demand forecast model is an important task achieved by examining variable combinations and assessing their performance with several statistical measures. We select these variables to measure factors such as their overall explanatory value, the reasonableness of the influence statistically attributed to each variable, and the variables’ relationship to one another (correlations and collinearity).   

The SPU team learned many lessons about identifying and defining variables, data sources, and processing requirements. A few of the key lessons learned include:  

  • It is important to let your data inform the most appropriate model type and structure. Econometric models require a lot of data and may not be appropriate for every utility.  
  • Local issues can inform forecasting variables. For example, if your climate patterns are stable, being able to model climate changes may be less important than understanding the impacts of economic conditions. 
  • Continuous management and improvement will lead to a more robust model as data availability and consistency improves.  

Like its predecessor, the new demand forecasting model will be a dynamic tool that SPU staff will use to inform decision-making and continually improve the results to be as reliable as possible. 

If you want to learn more about this exciting collaboration with Seattle Public Utilities and how Raftelis can help you understand your long-term water demand forecasting needs, please contact Joe Crea at jcrea@raftelis.com 

Raftelis would like to thank Seattle Public Utilities, especially the project team members for their invaluable support and participation in this project. Those SPU team members include Elizabeth Garcia, Ph. D, Kelly O’Rourke, Paul Hanna, Maura Patterson, and John Gibson.  

Get the Latest Insights Delivered