Accounting for climate change in water infrastructure design: evaluating approaches and recommending a hybrid framework

The practice of basing the hydrologic design of water infrastructure on principles where climate is stationary and future conditions can be represented by variances in historical trends is no longer appropriate given the preponderance of climate studies predicting a varied and uncertain climate future (IPCC 2001, 2014; Milly et al. 2008; Brown 2010). This varied climate future is expected to increase the severity of extreme hydrologic events in many parts of the world (Christensen & Busuioc 2007). Climate change is generally expected to result in temperature increases of 1.5 to 2.5 degrees C (IPCC 2021). The extents of these changes may, however, vary from region to region. Rising temperatures are likely to result in more intense storm surges, requiring more substantial flood control structures (IPCC 2014, 2021). Static approaches for water infrastructure design may result in failed infrastructure investments that could lead to loss of lives, property and stormwater pollution, and sanitation issues that present adverse conditions for public health (IPCC 2014, 2021). Therefore, a shift to a dynamic approach that accounts for the changing climate is urgently needed and frequently argued (Garrote 2017).

In response, research on designing and managing water resources infrastructure to account for climate change has focused on improving climate forecast data by downscaling climate change projections from a general circulation model (GCM) to a local level and translating the climate projections into precipitation projections at a design input level (Hoar & Nychka 2008; Islam et al. 2018). It is argued that downscaling refines spatially coarse and biased GCM output and makes them usable for local and regional climate analysis (Salathé 2003). The climate change-informed data is then used in design and management decisions. However, there are two limitations to this approach.

First, downscaling techniques routinely rely upon developing statistical relationships between local climate variables (such as precipitation and temperature) and large-scale (global and regional) predictors (Lanzante et al. 2018). The statistical relationship is then applied to the output of GCMs to predict local climate characteristics in the future (Hoar & Nychka 2008; Lanzante et al. 2018). Uncertainties in the input variables are carried through and the resulting downscaled climate change forecasts may lead to further uncertainty in the design variables due to the wide range and uncertainty of the input, for instance, carbon emission projections (Cook et al. 2020).

Second, downscaling approaches assume that the relationship between the GCM outputs and the regional climate data derived based on historical data is valid for future predictions (Dixon et al. 2016). The assumption that the relationship between the GCM outputs and regional climate data is constant (statistical stationarity) is not always supported as changes in the environment such as urbanization may alter the relationship between the GCM outputs and the regional climate data.

To address the limitations of climate change projections and GCM downscaling methods, this paper systematically reviews the literature and identifies five approaches accounting for climate change in the design and management of water infrastructure, outlining the strengths and weaknesses of each approach and assessing the applicability of each approach to water infrastructure design. Each approach proposes methods to reduce the errors in resulting infrastructure design, recognizing the inherent uncertainty in climate data, and GCM outputs available for decision-making.

This paper is structured as follows: the first section describes the research methodology, including the selection of literature, approaches, and case studies. The section Review of Approaches gives a summary of approaches accounting for climate change in water infrastructure design in literature and selected case studies. This paper then presents a Discussion of the approaches and implications for future design, followed by a Conclusion.

The contribution of this study is holistic. First, in meeting its aim, this study offers a summary of approaches to designing more climate-resilient water infrastructure. Second, it extends the literature and knowledge of applicable approaches, allowing water managers to consider a broader suite of management options with stakeholders. Third, to our knowledge, this is the first comparative review of uncertainty approaches for addressing non-stationarity in water infrastructure design.

To select the approaches accounting for climate change in water infrastructure, and subsequently select case studies that demonstrate the applicability of such methods, a literature review was conducted between April and June 2022. Relevant literature was retrieved using Primo Search and Google Scholar. The keywords/phrases used to retrieve publications within these search engines/libraries were climate change on precipitation, non-stationarity, and water infrastructure design. We selected literature that discussed the application of methods for addressing non-stationary as opposed to literature that only utilized a method.

Following stakeholder consultation with water industry experts on current approaches and challenges for addressing climate change, literature was examined against four criteria for addressing climate change uncertainty in water infrastructure design. (1) The applicability of the approach to water infrastructure design; (2) the reduction in uncertainty achieved by implementing the approach; (3) the adaptability of the approach to various design conditions and basins; and (4) the computational cost-effectiveness of implementing the approach.

The criterion, applicability, refers to whether the approach can be applied in the design and management of water infrastructure. In review we considered whether the approach yielded quantifiable hydrological estimates that can serve as the basis for designing and managing water infrastructure, supports water industry application, and whether it is reproducible (should produce similar results irrespective of user).

The second criterion, reduction in uncertainty, addresses how the approach reduces the challenge of including uncertainty in the decision-making process. Water resources practitioners and managers face management issues ranging from insufficient hydrologic record lengths, natural variability blended with anthropogenic induced changes, inconclusive results from climate change studies, and the effect of future climatic changes on the hydrologic design (Stakhiv 2011). It is therefore important that water practitioners have accurate hydrological estimates of the future climate to make reliable decisions. We reviewed approaches for how well they eliminate or accurately quantify the uncertainty regarding climate change projections and downscaling methods. Having a quantified estimate of uncertainty aids decision-making.

Our third criterion, adaptability, refers to how well the approach can be applied to basins of different sizes, uses, and in various geographic regions. Approaches that are specific to water infrastructure types, regions, were viewed as lacking adaptability whereas those approaches that could be applied broadly to all basins would have more universal application and be of use to the global water community.

Our final criterion is cost. Cost is an often-cited factor in the acceptability and usefulness of an approach intended to be used by water practitioners to design and manage water infrastructure. We considered cost as the relative computational cost-effectiveness of an approach in terms of effort and funds expended in implementing the approach. Cost was not calculated but qualitatively assessed in the review based on industry stakeholder consultation.

Case studies were selected from published and grey literature (in academia, government agencies, and the engineering industry) and demonstrate how the approaches described in this paper have been applied to watersheds in arid and temperate parts of the United States. A descriptive review of the case studies was conducted to gather insights into the strengths and weaknesses of the methods as they relate to the four criteria above.

Literature and case studies were limited to publications after 2001 to correspond with the third Intergovernmental Panel on Climate Change report detailing the scientific basis for adapting to the effects of global climate change (IPCC 2001). The third IPCC report presents observations from complex physics-based climate models that indicate confidence in the projections was on the rise and create a sound footing for the need to address non-stationarity effects. The 20+ year period provided an extensive review timeframe to suggest insights about how the knowledge and perspective on the approaches for indoctrinating climate change in water infrastructure design have evolved.

Conrad et al. (2013) define adaptive management in a water context as an iterative process for improving and revisiting management policies and practices by learning through monitoring and from outcomes of rigorously designed investigations. The generalized sequence in adaptive management involves analysis of the problem that needs to be solved, developing a management plan that includes treatments or responses to the problem, implementing the plan, monitoring, and evaluating and adjusting the plan based on observation. The evaluation results are then fed back into the decision-making process to aid future planning and management. Proponents of adaptive management argue the framework provides a flexible decision-making process for adapting to uncertainty and allows for evolving engineering and planning paradigms to increase flexibility in the infrastructure design (Pahl-Wostl 2008; Allan et al. 2013; Conrad et al. 2013; Kundzewicz et al. 2018).

Adaptive management approaches mitigate against the uncertainty brought about by the non-stationary nature of climate by continually adjusting the water infrastructure design and management decisions based on improvements in climate change data and methods. Uncertainty exists in water infrastructure design parameters due to climate change. Information about future climate needed for design of water infrastructure will always be incomplete. Adaptive management provides a flexible approach to managing the uncertainty and allows water practitioners to respond as situations change (Pahl-Wostl et al. 2007). Adaptive management supports water managers adjusting future actions based on answers to questions such as: To what extent were management goals met? How accurate were past predictions? How valid are the assumptions about the impacts from climate change? Did the actions have the desired effect? And what additional data and information is required to evaluate future actions? Adaptive management is a rigorous tool for learning through intentionally designed management actions and requires constant monitoring and learning to formulate adaptive adjustments as added information becomes available (Conrad et al. 2013).

Adaptive management is widely used in the fields of forestry, ecosystem restoration, and others (see Fernandez-Gimenez et al. 2008; Allen & Garmestani 2015). However, it has not yet been widely accepted in the water infrastructure design and management and may face institutional and behavioral challenges during implementation (Varady et al. 2016). Adaptive management may also require changes to already built infrastructure and such actions are undesirable for some water practitioners. Additional challenges and barriers in the implementation of adaptive management include the following: institutional concerns in admitting uncertainty cannot be overcome; challenges securing the additional resources required to design robust experimental, management activities, and monitoring schemes; difficulties maintaining long-term continuity of funding; additional costs for model development, assessment and monitoring; difficulty in detecting and documenting success; political pressure to alter management plans; institutional inertia (‘path dependency’) and inflexibility (‘lock-in’); and institutional arrangements that do not foster activities with long-term time frames (Conrad et al. 2013).

Inverse climate change impact approaches assess the influences of climate change in terms of hydrological extremes on infrastructure design by identifying the critical hydrologic exposures, such as floods and droughts, that may lead to failures of the hydrological system and transforming the hydrological exposure into corresponding meteorological conditions by means of hydrological models which are linked with future climate scenarios generated by a weather generating algorithm coupled with GCM outputs (Cunderlik & Simonovic 2007). These approaches allow for the evaluation of the performance of water infrastructure independently of climate change projections (Guo et al. 2018).

These approaches aim to mitigate the uncertainty brought about by the non-stationary nature of climate by providing conservative estimates of the extreme events that result in failure of the infrastructure. For instance, Cunderlik & Simonovic (2007) note that inverse climate change impact approaches are advantageous because they focus on specific existing and anticipated future water resource problems, have a direct link with the end-user, and allow for easy updating, when new and improved GCM outputs become available.

The structure behind inverse climate change impact approaches involves identifying critical local hydrologic exposures to existing or future water resource systems. Followed by a hydrologic model to transform hydrologic exposures through extreme precipitation events and dry spells over an extended period. The critical meteorological conditions are then simulated under existing and future climatic scenarios using a weather generator that is linked with GCMs and an ensemble of simulations that reflect possible future climatic conditions. Weather generators produce ensembles of climate scenarios that can be sampled at yearly, monthly, daily, or hourly timescales. The result from this process is information about the frequency of the critical meteorological events that are likely to result in the failure of the hydrological system (Cunderlik & Simonovic 2007; Guo et al. 2018).

In summary, this approach addresses uncertainty through a methodological process by (1) identifying the critical flooding depths that will result in failure of a hydrologic system, (2) estimating the rainfall depths and durations that will result in the critical flooding depths using a hydrologic model, (3) estimating the combinations of meteorological conditions (present and future) that can produce the critical rainfall depths and durations using weather generators and GCM outputs, and (4) estimating the likelihood and frequency of such meteorological conditions.

Physical and statistical models, which are the most common forms of modeling climate, atmospheric, and hydrological systems, have shortcomings such as accuracy, weakness in uncertainty analysis, high computation cost, and the need for a comprehensive amount of data (Mosavi et al. 2018; Ardabili et al. 2020). The data requirements for physical models make them unsuitable for use in short-term prediction (Mosavi et al. 2018).

Surrogate modeling techniques such as interpolation, regression analysis, and artificial intelligence (AI) may provide a means for utilizing complex climate models in decision-making environments. Machine learning and deep learning methods have shown to be able to overcome high computational requirements and uncertainties in modeling through their efficient computation and intelligence (Mosavi et al. 2018; Ardabili et al. 2020; Farzin & Valikhan Anaraki 2021).

Mosavi et al. (2018) define machine learning as a field of AI used to induce regularities and patterns, providing easier implementation with low computation cost, as well as fast training, validation, testing, and evaluation, with high performance compared to physical models, and relatively less complexity. Machine learning methods can numerically formulate the flood nonlinearity, solely based on historical data without requiring knowledge about the underlying physical processes therefore providing an appropriate surrogate for physical models. Machine learning methods are also quicker to develop with minimal inputs (Mosavi et al. 2018). Machine learning and deep learning approaches mitigate against the uncertainty brought about by the non-stationary nature of climate by better characterizing the level of uncertainty thereby making it possible to account for the level of uncertainty in decision-making.

Machine learning methods can be classified as tree-based, support vector machine (SVM), neural network-based, and hybrids and ensembles. Tree-based ensemble methods include Random Forest (RF), M5tree, gradient boosting decision tree, and extreme gradient boosting. RF has been shown to have higher accuracy when compared to other methods (Ardabili et al. 2020; Kadkhodazadeh et al. 2022). Deep learning techniques are a significant part of machine learning methods (Ardabili et al. 2020).

The main shortcoming of machine learning and deep learning techniques noted in the literature is the extensive amount of data needed to accurately train them. Machine learning and deep learning, a subset of machine learning, techniques may also not accurately account for future variations in climate due to their reliance on historical data. As noted by Mosavi et al. (2018), machine learning methods learn to predict the target task based on past data. These methods are only as good as their training, so it is essential that the training data be as robust as possible. If the data are scarce or does not cover varieties of the task, their learning falls short, and hence, they cannot perform well when they are put into work.

Despite the data limitations, applications of machine learning methods for predicting future flood under the uncertainty of climate change have proved to be cost-effective and to increase overall predictive capacity (Mosavi et al. 2018). Applications include modeling the effects of climate and land use changes on flood susceptibility areas in the Tajan watershed (Avand & Moradi 2021), developing flood prediction for urban management (Motta et al. 2021), as a surrogate for high resolution output from MODFLOW, a modular hydrologic model (Miro et al. 2021), runoff prediction (Farzin & Valikhan Anaraki 2021), flood frequency analysis (Anaraki et al. 2020), and for predicting evapotranspiration (Kadkhodazadeh et al. 2022).

Gilroy & McCuen (2012) note that climate change is expected to affect flood magnitudes by changing precipitation patterns and urbanization affects flood magnitudes by changing the runoff for each precipitation event and in response, developed a method to adjust a flood record for future non-stationary conditions based on two factors: climate change and urbanization. The goal of their study was to apply a non-stationary peak discharge record to one that is stationary for a selected design year, with the added advantage that the adjusted stationary record can reflect future climate conditions and urbanization (Gilroy & McCuen 2012). Adjustment factors are developed for both the climate change and urbanization components in response to precipitation projections from selected GCMs. The goal is to estimate the expected temporal change in heavy rainfall intensity over time for a specified climate change scenario and then apply the estimated change to the observed precipitation record (Gilroy & McCuen 2012). The process is repeated until each peak discharge event within the measured flood series is adjusted to the urbanization and climate change conditions of the design year (Gilroy & McCuen 2012).

The method developed by Gilroy & McCuen (2012) allows for the development of a flood frequency analysis using data adjusted to reflect future climate conditions and urbanization. The resulting flood estimates reflect stationarity of the underlying distribution. The adjustment factor method reflects the expected change in peak discharge based on both urbanization and climate change and has the advantage of adaptability in that the adjustment factor can be updated as new climate change data become more available and more accurate or more complex hydrologic models are selected (Gilroy & McCuen 2012).

Similarly, Solaiman & Simonovic (2011) describe an approach for updating rainfall intensity–duration–frequency (IDF) curves to account for climate change. The approach involves determining possible realizations of future climate using downscaled outputs from GCMs; fitting annual maximum series of rainfall to Gumbel distribution to develop IDF curves for 1, 2-, 6-, 12- and 24-h durations for 2, 5, 10, 25, 50, and 100 years of return periods; estimating the associated uncertainties using nonparametric kernel estimation approach; and developing an IDF curve based on a probabilistic approach (Solaiman & Simonovic 2011). The overall method involves adjusting rainfall records to reflect long-term effects of climate change, transforming rainfall to runoff, and adjusting the peak discharge to account for future urbanization.

Flood frequency approaches may mitigate against the uncertainty brought about by the non-stationary nature of climate by employing an ensemble of climate change projections in determining the expected change in the frequency of a given storm event. Solaiman & Simonovic (2011) noted that using a multi-model approach, rather than a single scenario, is encouraged due to the inherent uncertainties associated with different GCMs. GCMs represent physical processes in the atmosphere, ocean, cryosphere, and land surface and are the most advanced tools available for simulating the response of the global climate system to increase greenhouse gas concentrations. GCMs offer only possibilities of future climate pattern in differing socioeconomic conditions depending on the continual growth of population, increased carbon dioxide emission, rate of urbanization, etc. Outputs from GCMs, thus, should not be considered as the forecasts of future climate conditions. The authors argue that for a comprehensive assessment of the future changes, it is important to use collective information by utilizing all available GCMs and synthesizing the projections and uncertainties in a probabilistic manner.

Soft computing approaches mitigate against the uncertainty brought about by the non-stationary nature of climate by providing a means to characterize and quantify the level of uncertainty thereby making it possible to account for the level of uncertainty in decision-making. For instance, Steinschneider et al. (2012) noted that Bayesian methods provide a formal mechanism to characterize the error in hydrologic model predictions, along with uncertainties surrounding parameterization.

Mondal & Mujumdar (2015) noted that the low resolution of GCMs and their inability to simulate localized processes necessitate the use of statistical downscaling models which are often used together with physically based hydrologic models for the generation of streamflow projections. This approach usually leads to projections of hydrological extremes with low to medium confidence because of the limitations of climate models and lack of sufficient observations at smaller scales. The extreme value theory (EVT) provides the theoretical basis for modeling extreme events. Non-stationary extensions to the traditional stationary extreme value models enable incorporation of effects of one or more physically based covariates in the parameters of the statistical extreme value model. Non-stationary extreme value distributions can be applied to study climate change extremes. Mondal & Mujumdar (2015) used a two-component non-stationary peak-over threshold (POT) approach of EVT statistical model based on Poisson-Generalized Pareto (Poisson-GP) distribution to model changes in the frequency and intensity of droughts.

Steinschneider et al. (2012) proposed a Bayesian formulation for rainfall–runoff modeling that can be used to help quantify uncertainties in a climate change impacts analysis. Steinschneider et al. (2012) noted that the direct use of downscaled, multi-model GCM output as forcing data for climate change assessments can only generate a lower bound on the maximum range of future climate uncertainty. Since GCM simulations over the historic record do not fully explore the multiple sources of uncertainty at play, it may prove difficult, if not impossible, to develop a satisfying error model and bracket the true uncertainty of future climate projections. The objective of the formulation was to demonstrate how future climate uncertainties can be nested in a framework aimed at quantifying the total uncertainty in future hydrologic projections. Steinschneider et al. (2012) accounted for different sources of hydrologic model uncertainty using Bayesian modeling. They formally characterized the distribution of model residuals to quantify predictive skill and used Markov chain Monte Carlo sampling to infer the posterior distributions of both hydrologic and error model parameters. They then integrated parameter and residual error uncertainties to develop reliable prediction intervals for streamflow estimates. They then extended the Bayesian hydrologic modeling framework to a climate change impact assessment. They downscaled ensembles of baseline and future climate from GCMs and used them to drive simulations of streamflow over parameters drawn from the posterior space. They calculated time series of streamflow statistics from baseline and future ensembles of simulated flows and integrated uncertainties in hydrologic model response, sampling error, and the range of future climate projections to help determine the level of confidence associated with hydrologic alteration between baseline and future climate regimes. Their formulation can be used to emphasize the importance of prediction error in uncertainty analyses under climate change and highlight the complex interactions between different sources of modeling uncertainty.

Teegavarapu (2010) observed that results from climate change models for hydrologic analysis are highly uncertain due to shortcomings and capabilities of GCMs and the different coupling methods used at various levels of scales in integration of GCMs and hydrologic models. Although a lot of effort has been put into improving climate change predictions, substantial uncertainty remains in the trends of hydrological variables because of large regional differences, gaps in spatial coverage, and temporal limitations of data (Teegavarapu 2010). Teegavarapu (2010) highlights several issues related to climate change and its impact on water resource management and addressed a few by providing methodologies that can help to deal with climate change in water resources management models. Teegavarapu (2010) proposed a fuzzy mathematical programming framework to aid decision-makers include their degree of uncertainty in their climate change predictions and processes. Teegavarapu (2010) noted that membership functions are generally defined on fuzzy sets to describe the degrees of truth on a scale [0–1] for a physical quantity or a linguistic variable. The scale is used to characterize the uncertainty or vagueness associated with the definition of the variable as perceived by humans. Fuzzy membership functions are used to model the decision maker's preferences attached to possible variations in hydrologic inputs due to climate change.

Teegavarapu (2013) noted that soft computing approaches composed of fuzzy rule-based inference systems and neuro-fuzzy methods and their derivatives are ideally applicable for situations in which system parameters are imprecise and vague and are primarily useful for adaptive control under uncertain decision-making environments. Teegavarapu (2013) noted that precipitation, as an important component of hydrologic cycle, bears a significant influence on hydrologic design and water resources management. The uncertainties associated with future climate change coupled with limitations of climate change models and uncertainties in projections, and also our inability to quantify these introduced additional complexities in the hydrologic design using future precipitation extremes. Teegavarapu (2013) proposed a new optimal compromise hydrologic design of a storm sewer system using a fuzzy mixed integer nonlinear mathematical programming model with discrete, binary variables, and logical constraints.

The model proposed by Teegavarapu (2013) incorporates preferences of hydrologist toward possible changes in future precipitation extremes within an optimization formulation for a hypothetical storm sewer system. Linear fuzzy membership functions are used to describe decreasing preferences toward cost. Also, preferences toward changes in future precipitation extremes are modeled using linear, nonlinear (exponential, Gaussian) and triangular membership functions. The fuzzy nonlinear optimization model provides climate change-sensitive hydrologic design that considers compromise between current and future precipitation extremes by handling the preferences. Discrete values of optimal decision variables provided by the model aid in the implementation of optimal design solutions in field applications.

This section describes three case studies of the application of climate change forecast in the design of water infrastructure. Selection of case studies was limited to those completed from 2001 to present to focus on the application of the most recent advances in climate change modeling that rely on relatively reliable GCM outputs.

Forsee & Ahmad (2011) analyzed the projected changes in design-storm depths and the resulting effects on stormwater infrastructure design for an arid watershed in Las Vegas delta using the updated flood frequency approach. They used multipliers that represent the percentage change in climate to transpose projected future changes in climate onto point precipitation data. Forsee & Ahmed (2011) initially based the projected changes in design-storm depths on five North American Regional Climate Change Assessment Program (NARCCAP) data sets. The NARCCAP data sets have a 50-km spatial resolution and provide precipitation on a 3-h temporal resolution. The next step was to use climate change projections from GCMs and Regional Climate Models (RCMs) that best reproduce historic precipitation statistics to calculate delta-change factors. The delta-change factors were then applied to design-storm depths, enabling the analysis of stormwater infrastructure under different scenarios. The time span used was 30 years for both historic (1971–2000) and future (2041–2070) scenarios.

Forsee & Ahmad (2011) utilized the existing Pittman HEC-HMS stormwater model for their analysis. In the Pittman HEC-HMS stormwater model, each sub-basin has a specified 6-h 100-year point depth which were obtained from NOAA Atlas II estimates of the 6-h 100-year depth. The Pittman HEC-HMS model was run with baseline precipitation data. The precipitation data for each of the sub basins were then adjusted based on the delta-change factors and then rerun to estimate the impact of climate change on the hydraulic design on stormwater infrastructure within the basin.

Our case study exploration in Washington State demonstrates the application of the updated flood frequency analysis approach. Rosenberg et al. (2010) examined both historical precipitation records and simulations of future rainfall to evaluate past and prospective changes in the probability distributions of precipitation extremes in the Thornton Creek and Juanita Creek watersheds. Rosenberg et al. (2010) based their historical analyses on hourly precipitation records for the period 1949–2007 from weather stations surrounding the Puget Sound region (including Seattle, Tacoma, and Olympia), the Vancouver (WA) region (including Portland, OR), and the Spokane region. Rosenberg et al. (2010) simulated changes in future precipitation changes using two runs of the Weather Research and Forecast RCM for the time periods 1970–2000 and 2020–2050, statistically downscaled from the ECHAM5 and CCSM3 Global Climate Models and bias-corrected against the SeaTac Airport rainfall record. Rosenberg et al. (2010) used the bias-corrected and statistically downscaled (‘BCSD’) hourly precipitation sequences as input to the Hydrological Simulation Program – FORTRAN (HSPF) model to simulate streamflow in two urban watersheds in central Puget Sound. HSPF is a FORTRAN-based hydrologic simulation program which was developed under contract and is maintained by the U.S. Environmental Protection Agency. HSPF is a lumped-parameter model that simulates discharge at user-selected points along a channel network from a time series of meteorological variables (notably, rainfall, temperature, and solar radiation) and a characterization of hydrologic variables (such as infiltration capacity and soil water-holding capacity) that are typically averaged over many hectares or square kilometers (Rosenberg et al. 2010).

Rosenberg et al. (2010) used the BCSD precipitation data for the periods 1970–2000 and 2020–2050, as input to HSPF to reconstitute historical streamflows and predict future streamflows in the Thornton Creek and Juanita Creek watersheds. Additionally, Rosenberg et al. (2010) used HSPF to evaluate differences in 1970–2000 simulated flows as forced by both the historical rainfall record and the BCSD rainfall. Their results from both the Thornton Creek and Juanita Creek modeling runs did not indicate significant differences in streamflow biases from the direct comparison of observed and simulated rainfall records.

Rosenberg et al. (2010) then simulated streamflows for both watersheds and each of the two RCM runs using the BCSD rainfall for the periods 1970–2000 and 2020–2050. They fitted Log-Pearson Type 3 distributions to the resulting annual maxima and tested for statistical significance using the Komolgorov-Smirnov, Wilcoxon rank-sum, and Mann–Kendall tests at a two-sided α of 0.05. Their results at the mouths of both watersheds indicate increases in streamflows for both RCM runs and all recurrence intervals.

Thakali et al. (2016) conducted a study to assess the potential effects of climate change on existing infrastructure systems for urban stormwater infrastructure in Las Vegas Valley using the updated flood frequency approach. Thakali et al. (2016) used historical and projected future precipitation from all the available GCM-driven and RCM-projected precipitation data from NARCCAP to calculate the deviation between historical and future design-storm and used the data that best represents the historical precipitation to calculate delta-change factors, which were applied to the design-storm depths.

In their study, Thakali et al. used regional frequency analysis (RFA) based on L-moments (a linear combination of probability-weighted moments) to calculate the design-storm depth (i.e., 100 year-6 h) from the all the datasets of NARCCAP and North American Regional Reanalysis (NARR) and accounted for projected changes in precipitation data using delta-change factors. Thakali et al. (2016) used data from four encompassing grids of the Flamingo and Tropicana watershed for the RFA. First, they aggregated 3-h NARCCAP precipitation data to 6-h precipitation data by shifting a 6-h window through each 3-h value. Using an algorithm, they calculated series of 6-h annual maximum data for each grid and model combinations. For regionalization, they pooled the data by dividing each maximum by a median of the same data series for each grid. They fitted the generalized extreme value (GEV) distribution to each standardized pooled-annual extreme dataset of all the four grids of each climate model combination. The GEV distribution method is a commonly accepted approach in the non-stationarity study of extreme flows owing to the skewed nature of annual flow maxima and the ability to include covariates in the parameters of distribution (Thakali et al. 2016). Thakali et al. (2016) performed this process for all sets of the data from each climate model. They then calculated the delta-change factors by dividing the projected future 100 year – 6-h depth by historic 100 year – 6 h depths.

Thakali et al. (2016) analyzed stormwater systems under the projected climate change conditions using an existing Hydrologic Engineering Center's Hydrological Modeling System (HEC-HMS) model of Las Vegas Valley watershed developed by the Clark County Regional Flood Control District (CCRFCD). The existing HEC-HMS model of the Flamingo and Tropicana watershed was used to assess the effects of climate change on stormwater facilities. The output from HEC-HMS was analyzed further according to guidelines from the CCRFCD design manual, and comparisons were made among the various climate projection scenarios.

Thakali et al. (2016) first performed a baseline HEC-HMS simulation where the existing HEC-HMS model was run with no changes. To assess the effects of climate change, they applied the delta-change factor (DCF) calculated from the NARCCAP models to all existing design depths in the HEC-HMS model.

Based on the investigations in this study, we find that adaptive management approaches are most likely to incorporate changes in climate data in future changes in design, as it becomes available, with the disadvantage of requiring constant monitoring, management, and in certain cases changes to the built environment. Furthermore, as noted by Allan et al. (2013), changing from conventional to adaptive approaches requires leadership and a reframing of the management perspective.

Inverse climate change impact approaches identify critical hydrological exposures that lead to system failures and determine future climate change conditions that likely result in a critical event. Inverse climate change impact approaches address the uncertainty in GCM projections by using the most conservative projections. As indicated in the discussed studies, inverse climate change impact approaches are advantageous because they are scenario independent (Guo et al. 2018), focus on specific existing and anticipated future water resource problems, have a direct link with the end-user, and allow for easy updating, when new and improved GCM outputs become available. We, however, find inverse climate change impact approaches to be limited in their application since they only focus on one structure at a time.

Machine learning and deep learning techniques attempt to reduce the uncertainty with climate projections by reducing prediction error in climate change forecasts and better quantifying the residual uncertainty. Machine learning approaches can numerically formulate the flood nonlinearity, solely based on historical data without requiring knowledge about the underlying physical processes. This is a major cost-saving attribute. However, we find that the ability of machine learning and deep learning techniques to reduce uncertainty is often reliant on having extensive amount of data to accurately train them. This is likely a major shortcoming of using machine learning approaches due to the lack of extensive historical climate change data within most basins. Machine learning and deep learning techniques may also not accurately account for future variations in climate due to their reliance on historical data.

Flood frequency and intensity duration curve approaches attempt to reduce the uncertainty in climate projections and downscaling techniques by using an ensemble of climate change forecasts in the development of the future flood frequency curves. As demonstrated by the reviewed case studies, these approaches are more widely accepted by water practitioners and have been extensively applied in basins across the United States (Stakhiv 2011). Some of these approaches account for future changes in climate as an adjustment factor. This allows for future changes in the flood frequency curve once better data becomes available in the future. It is worth noting that the resulting flood estimates from these approaches reflect stationarity of the underlying data and distribution and hence the uncertainty of the resulting flood estimates could be higher in a rapidly changing climate.

Soft computing approaches such as variations of EVT that account for nonlinearity due to climate change and fuzzy set theory are mathematical programming formulations that predict future hydrological behavior based on probable future changes in precipitation extremes. As discussed in the literature review, soft computing approaches have been portrayed as having the capability of accounting for the uncertainties associated with future climate change projections. Since soft computing approaches are mathematical formulations, preferences of hydrologists toward expected uncertain future changes in precipitation extremes can be included in the mathematical formulations. We find that the inclusion of preferences of hydrologists toward expected uncertain future changes in precipitation extremes may lead to biases in the predictions.

In terms of applicability to water infrastructure design, we find that the flood frequency analysis approach is likely the overall preferred option because it is most compatible with standard hydrological analysis methods. For this reason, the flood frequency analysis approach is widely accepted among water practitioners. Adaptive management approaches are best suited for reductions in uncertainty. In terms of adaptability, both the flood frequency and the inverse climate change impact approaches appear to be viable approaches for infrastructure design as they can be adapted to many basins. Machine learning and soft computing approaches are likely the least expensive to implement as they rely upon existing data and the approaches and benefits for using sophisticated computational methods are ever-growing (Richards et al. 2023).

The authors wish to note that the approaches reviewed in this study are not inclusive of all studied approaches for addressing non-stationarity but rather those approaches that may provide an intermediate step that accounts for the inherent uncertainty in climate data that is used in designing and managing water infrastructure. Other approaches for addressing climate change include ecosystem modeling approaches (John et al. 2020), the dynamic adaptive policy pathways approach that utilize simulation gaming to introduce uncertainty in decision-makers decision processes (Lawrence & Haasnoot 2017), weighted comprehensive assessment method (Xu et al. 2022), and other more classical Bayesian methods for classifying uncertainty (Hobbs 1997).

As noted by one reviewer, datasets, as well as the methods are important for managing climate effects. This paper did not review climate datasets; however, the datasets used in the reviewed papers include precipitation predictions from Global Climate Models. The most used model was the coupled model intercomparison project (CMIP). Other models used include Coupled Global Climate Model (CGCM), Hadley Centre Coupled Model (HadCM), and the Canadian Earth System Model (CanESM). Some also used data from the NARCCAP which contain outputs from RCMs.

The following five approaches accounting for climate change in the design of water infrastructure are presented in this paper: adaptive management, inverse climate change impact method, machine learning, updated flood frequency analysis, and soft computing approaches. Except for the soft computing approaches, all these methods rely on climate projections from GCMs and downscaling techniques. All five approaches recognize that there is uncertainty in climate data available for decision-making due to the uncertainty in the GCM outputs, the downscaling methods, among others.

The flood frequency analysis approach may likely be the most applicable to water infrastructure design as it may be, among those reviewed, the least disruptive in terms of standard hydrological analysis methods, likely cost-effective, and is likely adaptable to most basins. However, adaptive management approaches may be best suited for reductions in uncertainty since they provide for opportunities to constantly adjust decisions based on improved climate change data. A combination of flood frequency and adaptive management approaches could provide an optimal way of accounting for climate change in the design and management of water infrastructure.

All relevant data are included in the paper or its Supplementary Information.

The authors declare there is no conflict.