Establishing a Range of Extreme Precipitation Estimates in California for Planning in the Face of Climate Change

In California, as in much of the world, infrastructure and water resource planners and managers at state and local agencies are attempting to estimate their vulnerability to future changes in hydrologic conditions. Planning for a future in the face of climate change requires unbiased projections of precipitation and temperature at relevant time and space scales. Climate models are a foundational basis for these projections, because they contain, with varying levels of skill, representations of the large-scale physical drivers of climate change and of the Earth system model responses to it (Flato et al. 2013). These physically based models are needed to project events that are without historical analogues, but existing climate models are blunt instruments for projecting infrastructure risk, and can exhibit extreme levels of disagreement with each other at the local level (Hallegatte 2009; Gersonius et al. 2013). Although there are a number of potential sources of model error that can be related qualitatively to problematic cloud-process representations or large-scale earth system variability representation (Eden et al. 2012), the water resource planning community cannot wait for the identification of the cause(s) of these errors and their resolution; precipitation projections are needed to develop plans now that can address vulnerabilities for decades to come.

The principle challenge for developing these projections is that global models often are run at resolutions on the order of 100 km, which is far too coarse for understanding precipitation and the terrestrial hydrological response (Barros and Lettenmaier 1994; Leung and Ghan 1995). This is especially problematic for hydrologic projections for California, because at such coarse resolutions, climate projections offer little direct information about the spatial details of climate differences and variability that drive the output from most of California’s watersheds (Maurer et al. 2007; Wang et al. 2012; Hamann et al. 2013). At the same time, recent studies have indicated that there could be severe changes in the global- and regional-scale spatial patterns of the precipitation amount and thermodynamic phase that also directly impact the future of California water resources (Cayan et al. 2008; Pierce et al. 2013; Ullrich et al. 2018; Rhoades et al. 2018). These changes could adversely impact water resource availability, and global models are likely needed to estimate such large-scale shifts.

There are tentative efforts to include both large- and fine-scale information, either through computationally expensive global high-resolution calculations, such as the High Resolution Model Intercomparison Project (HighResMIP) (Haarsma et al. 2016), or through regionally refined simulations (e.g., Rhoades et al. 2016; Tang et al. 2019). These higher-resolution simulations include realistic topography and land-type contrast, but questions remain regarding both the prohibitive computational expense of global high-resolution simulations and unresolved issues with high-resolution and variable-resolution simulations regarding the representation of processes at the fine-scale that remain heavily parameterized in these models (Barsugli et al. 2009; Pierce et al. 2013). Alternatively, we can combine the global or regional-scale climate-change signals from climate models with methods that capture local effects at relevant scales. Nevertheless, we must recognize that for hydrologic projections, temperature and precipitation from climate or downscaling models are used as inputs to hydrological process models, and caution must be exercised in chaining together multiple types of models (Hagemann et al. 2011; Muerth et al. 2013). Early biases in the model-chaining exercise can propagate to produce implausible hydrological model output (Maraun 2016), and there can be a numerical challenges with efforts to scale precipitation change, for example, through quantile mapping (QM) (Maraun et al. 2018).

Previous efforts that evaluated global climate model (GCM) performance for Northern California found that, in general, an ensemble of GCMs performed better than the individual models when a broad range of historical climate metrics were considered (Brekke et al. 2008). Although an ensemble of climate model simulations may be preferred, the range in emissions scenarios (van Vuuren et al. 2011) and climate model responses (Andrews et al. 2012) provide far too many simulations for the purposes of water planning exercises; it is simply infeasible to develop water plans for more than a handful of data sets. This problem is particularly acute for local water districts, which often lack the technical or financial capacity to analyze and consider such complex modeling outputs. This challenge was recognized recently, and preliminary guidelines were developed to use a limited set of tailored scenarios to represent the spectrum of climate scenarios optimally (Ntegeka et al. 2014).

This paper describes the process used to develop local projections, including balancing regulatory workflows that operate based on historical data with the need to consider how to plan for future hydrometeorological events, either individually or as a group, that are without historical analogues. The paper begins by describing the regulatory workflows, discusses the process of developing downscaled projections, and highlights where practical issues and scientific drivers thereof must meet to deliver information to local planning agencies.

California presents a particular challenge for water resource planning because the state’s approximately $\mathrm{164,000}\text{-}{\mathrm{mi}}^2$ span a vast range of hydroclimates, from temperate forests in the north to arid deserts in the southeast, and a large area with a Mediterranean-type climate in between. A wide range of processes are responsible for this range, including large-scale dynamics and small-scale topographic effects. Resource planners must consider how precipitation and temperature will change across these scales, especially where water resources are the result of a dozen or fewer synoptic events each year (Dettinger et al. 2011). Additionally, there is a particularly time-sensitive aspect to water-planning in California. As of this writing, the regulatory framework for water resources in California is in the midst of responding to the Sustainable Groundwater Management Act (SGMA) of 2014. This act gave the California Department of Water Resources a number of responsibilities and authorities, and robust hydrological information is a prerequisite for developing plans in the face of climate change.

At the agency level, practical limitations in resources and technical capacity often restrict the ability of resource management agencies to evaluate multiple projections of future conditions. To complicate matters further, climate may be just one of several future uncertainties that an agency may need to address (e.g., population growth, land-use patterns, and potential regulatory changes), and exploration of future conditions across multiple dimensions of uncertainty quickly can exceed an agency’s capacity or resources. To simplify these analyses, it often is useful to provide a more limited set of future climate conditions upon which to bound and anchor the analysis.

A recent example of this bound and anchor approach is the California Water Commission’s implementation of the $2.7 billion Water Storage Investment Program (WSIP) initiative. This process required applicants for funding to evaluate proposed projects under future climate conditions as indicated by an ensemble-informed climate future representing a median-like future condition. Applicants were also required to evaluate the climate sensitivity of proposed projects under two more-extreme bookend-like conditions representing wetting and moderate warming of California’s climate and drying and extreme warming of California’s climate. These analyses provided the decision makers with key information on how projects would be expected to perform under future different climate conditions.

For both the WSIP and SGMA programs, state agencies were required to develop regulations and technical resources to support analyses conducted by local agencies. Many of the local agencies receive at least a portion of their water supplies from the state owned and operated State Water Project (SWP), the federally owned and operated Central Valley Project (CVP), or both. Both of these major water supply projects draw water from mountainous watersheds in northern and central California, store water in large reservoirs, use natural river channels to convey water to the Sacramento–San Joaquin Delta, and then pump water out of the Delta and convey it through aqueducts to service areas throughout central and southern California. As part of the development of plans to manage these projects, since 2004, respective state and federal operators have collaboratively built, managed, and run a general-purpose reservoir–river basin simulation model known as CalSim (Draper et al. 2004). To be relevant for planning and management of water resources in these projects, hydrological simulations need to conform to the unique requirements of CalSim and its successor versions (currently CalSim 3.0).

CalSim simulates system operations for a multiyear period at a monthly time interval. The model assumes that facilities, land use, water supply contracts, and regulatory requirements are constant over this period, representing a fixed level of development. In addition, boundary forcing conditions such as sea level and inflow hydrology also must be stationary; although they may vary from month to month and from year to year, their statistical distribution should not migrate over the course of the simulation. CalSim also uses empirical data as input. Because these empirical values are based on historical observed conditions, CalSim generally is run with the historical streamflow record, adjusted for the influence of land-use change, upstream flow regulation, and climate change. The simulation output thus represents the possible range of water supply conditions for the Central Valley at a given level of development and timeframe (e.g., 2015 or 2050). Rim watershed inflows, stream accretions and depletions, water diversion requirements (demands), and return flows are the primary components of the input hydrology (California Natural Resources Agency 2017).

In developing climate change scenarios to be used across California, CalSim’s requirement for stationary streamflow traces constrain the range of approaches available. Because the direct use of downscaled GCM projections run through a rainfall–runoff model or other traditional approaches fail to produce stationary hydrologic traces, alternative approaches are needed that can map the important climate change trend information from GCMs onto the historical meteorological and hydrological signal.

In response to these water resource planning needs and following the approach indicated in Ntegeka et al. (2014), we developed a process for producing the temporal distribution of precipitation and temperature in California at hydroclimatically relevant spatial scales for less extreme, wetter conditions and for more extreme, drier conditions, and discussed inherent numerical challenges in developing these projections.

The Coupled Model Intercomparison Project—Phase 5 (Taylor et al. 2012) is a widely utilized data set for climate model studies which contains nearly 3 dozen different climate model projections for each emissions scenario. For water resource planning in the state of California, it is essential that models exhibit a level of skill with respect to regional metrics (Rupp et al. 2013), including correlation and variance of mean seasonal spatial patterns, amplitude of seasonal cycle, diurnal temperature range, annual- to decadal-scale variance, long-term persistence, and Western US regional precipitation teleconnections to the El Niño–Southern Oscillation (ENSO).

Different GCMs exhibit greater skill for different metrics, and when multiple metrics were considered, no individual model emerged as the best model for California. Recognizing the need for multiple GCMs, as well as the requirement for a smaller set of simulations, CCTAG (2015) performed a model evaluation effort to identify a smaller set of GCMs by removing or culling the models that did not perform as well for a set of different evaluation metrics. This approach downselected the number of California-relevant climate models to 10: ACCESS-1.0, CCSM4, CESM1-BGC, CMCC-CMS, CNRM-CM5, CanESM2, GFDL-CM3, HadGEM2-CC, HadGEM2-ES, and MIROC5. The output from these models from both the Representative Concentration Pathway (RCP) 4.5 and RCP8.5 scenarios has been used to construct a multimodel ensemble, but specific scenarios that provide a range also are needed.

For efforts such as WSIP, described previously, and the implementation of regulations for the SGMA, which has similar constraints and requirements to WSIP, there is a need to explore the bounds of expected temperature and precipitation change throughout the 21st century in California. These bounds then can form the basis for water resource projections (Maurer et al. 2010) using a well-established hydrological model such as the Variable Infiltration Capacity (VIC) model (Liang et al. 1994; Gao et al. 2010), with results that can be used in CalSim.

Two bookend or bounding cases were developed for use in these programs (California Code of Regulations, Title 23, Division 7, Section 6000-60015). The first, known as wetter moderate warming (WMW), describes a future for California that exhibits moderate warming while also generally receiving more precipitation. The CNRM-CM5 model, run with the RCP4.5 emissions scenario, was chosen as the GCM/RCP projection pair because this projection generally exhibits the combination of least warming and largest increase in precipitation among the $20\mathrm{GCM}/\mathrm{RCP}$ projection pairs in the ensemble chosen by CCTAG (2015). This model exhibits a midrange value of equilibrium climate sensitivity (ECS) ($3.25\mathrm{K}/2\times {\mathrm{CO}}_2$) (Andrews et al. 2012), and although it generally exhibits regional biases, the model does not exhibit precipitation biases in California relative to Global Precipitation Climatology Project (GPCP) observations (Voldoire et al. 2013).

The second case, known as drier extreme warming (DEW), exhibits extreme warming while also generally receiving less precipitation. The HadGEM2-ES model (Jones et al. 2011a) run with the RCP8.5 scenario was chosen as the GCM/RCP projection pair because this projection generally exhibits the combination of most warming and largest decrease in precipitation among the $20\mathrm{GCM}/\mathrm{RCP}$ projection pairs in the ensemble chosen by CCTAG (2015). This model exhibits a high value of diagnosed ECS ($4.69\mathrm{K}/2\times {\mathrm{CO}}_2$) (Andrews et al. 2012) and has superior performance with respect to precipitation relative to historical observations over North America (Sheffield et al. 2013). Together, these cases help establish a nominal range of conditions that state and local planning agencies across California would be likely to face (CCTAG 2015).

Fig. 1 shows the relative change in annual upper-quantile daily precipitation for downscaled (methods discussed subsequently) DEW and WMW cases. At the 90th quantile, the DEW and WMW cases lead to decreased and increased precipitation, respectively, whereas for the 99th and 99.9th quantiles, both cases show significant increases even while the DEW case indicates more extreme coastal precipitation and the WMW case indicates increases in extreme precipitation in southern California and the Sierra Nevada.

In Fig. 1, a significant amount of localized information is available for local management and planning purposes. Although not exhaustive, the change in the North American Monsoon is driving changes in precipitation in Southeastern California, which has significant implications for infrastructure resilience to handle flooding, but minimal implications for water resources which are derived from elsewhere. The changes in precipitation in the coastal ranges, the Sierra, and the Siskiyous are of greatest interest for planning and management perspectives. In that light, the southern Sierra and the coastal ranges near the Central Coast are impacted most heavily, and although the results of these changes are discussed subsequently from a state-wide perspective, we note here that agencies in charge of those resources may need to consider how to manage and plan for changing extreme precipitation in those areas.

The patterns of changing temperature are quite distinct from those of changing precipitation (Fig. 2). The spatial patterns of daily minimum temperature change are much less variable than those of precipitation for both the DEW and WMW cases, and can be summarized as 6°C and 3°C, respectively. However, the impacts of temperature on water resources are indirect: they propagate into VIC by impacting the phase-partitioning of surface precipitation and also influence other water flux terms such as evapotranspiration.

Many models are unable to reproduce both low- and high-order moments of observed historical distribution of precipitation at regional spatial scales (Christensen et al. 2008; Mearns et al. 2012; Sillmann et al. 2013). Fig. 3 specifically indicates the need for bias correction. We compared an observationally derived data product with present-day runs from the DEW and WMW models. Both models are biased low in northwestern California for the 90th, 99th, and 99.9th quantiles, and the models are biased low across the Sierra for the 99th and 99.9th quantiles (Fig. 3). In particular, in Northern California, biases in downscaled model output can exceed 30% of the total precipitation value. For planning purposes, these biases must be addressed, and although a univariate bias-correction does not address the source(s) of model bias or how they propagate into future projections, it does recognize that a basic adjustment to account for this bias is needed from a practical perspective.

This motivates a hybridized approach wherein statistical information derived from present-day observations is merged with climate-model projections. Historical observations can be used to bias-correct historical climate-model projections, and, although the bias-correction approach is controversial (Ehret et al. 2012), it has been shown to be tenable (Teutschbein and Seibert 2012).

Because the spatiotemporal distribution of precipitation across any given area often exhibits zero absolute values, bias-correction techniques that are used for interval values (e.g., temperature) necessarily lead to negative, and therefore unphysical, values. To address this issue, a ratio approach to bias correction can be used, which obviates this apparent problem. Using a ratio to scale bias-corrected precipitation avoids the creation of an artificially continuous spatiotemporal precipitation field when it should be discontinuous. Additionally, a ratio technique may be able to preserve the relative change contained within a model or a multimodel ensemble between future and present-day precipitation.

The approach taken here uses a combination of bias correction and downscaling to create future projections. The process chain by which these projections are created is indicated in Fig. 4. Historical model results are downscaled using the localized constructed analogs method (LOCA) (Pierce et al. 2013) to a resolution of 6-km and are then bias-corrected based on a gridded product that is derived from historical observations (Livneh et al. 2013). The relative change in precipitation is determined from LOCA-downscaled future model results, and then a quantile-mapping procedure is used to create a set of bias-corrected downscaled projections that attempt to preserve the parent model’s relative change in precipitation at regional scales.

However, if appropriate caution is not exercised in this process chain, the bias-corrected downscaled projections would be problematic. Many downscaling techniques themselves incorporate bias correction, but additional bias correction is needed here. This is because for water planning purposes in California, local agencies first focus on how downscaled temperature and precipitation fields impact water resources over the historical period before altering their future water plans, and these fields need to force the version of VIC calibrated for California and complementary simulations from CalSim to produce historically accurate flows.

This is especially the case for extreme precipitation, for which it has been recognized that traditional quantile mapping, wherein the effects of climate change are realized by scaling the historical precipitation cumulative distribution function by the ratio of the mean future modeled distribution to the mean historical modeled distribution, can lead to artificial corruption if the mapped precipitation quantities do not reflect the parent model’s relative change in precipitation across quantiles (Hagemann et al. 2011; Cannon et al. 2015). An example of the artificial corruption that can arise from traditional quantile mapping is given in Table 1, which shows that QM, by not preserving the relative change across quantiles, produces mean seasonal precipitation changes that are overestimated substantially, including by hundreds or even thousands of percent, from the parent model.

A recently developed approach called quantile delta mapping (QDM) is designed in principle to preserve the relative model change in bias-corrected precipitation across all quantiles (Cannon et al. 2015). QDM also seeks to avoid the systematic overestimation of precipitation that arises when a univariate scaling is applied to perform quantile mapping (Cannon et al. 2015).

QDM represents a transfer function, and it can be used to map model-projected precipitation at a given time to a bias-corrected, relative-change preserving value. The approach that it takes, therefore, is first to determine the quantile to which the original model-projected value corresponds. In the next step, the relative change is determined from the ratio of modeled projected precipitation for that quantile value in the model-projected cumulative distribution function (CDF) to the modeled historical precipitation for that quantile value in the modeled historical CDF. Finally, the historical precipitation from the same quantile value but from the observed historical CDF is scaled by the model ratio. The functional form of this procedure operates on an individual grid box as follows:where ${\hat{x}}_{m,p}(t)$ = bias-corrected future projection of precipitation at a given time; ${\tau }_{m,p}(t)$ = modeled quantile value of model projected value from model-projected CDF; $F_{m,h}^{-1}$ = inverse CDF function of modeled historical CDF; $x_{m,p}(t)$ = original model-projected precipitation value; and ${\hat{x}}_{o:m,h:p}(t)$ = observed precipitation value in historical record corresponding to historical observed CDF at quantile ${\tau }_{m,p}(t)$.

We created a centennial-length data record of daily precipitation and minimum and maximum temperature that approximates the distribution of these variables under climate change conditions. In order to be relevant for hydrological modeling and planning purposes, we produced these variables at 6-km spatial resolution across the state of California, and, to a limited extent, over the areas of neighboring states that drain into rivers in California. A total or 11,367 grid boxes were analyzed, over a total of 35,429 days spanning a historical record ranging from January 1, 1915 to December 31, 2011, or a future record covering a period of comparable length. These were designed for direct compatibility with CalSim 3.

Although the QDM approach initially appeared straightforward, a number of numerical stability issues arose in the application of QDM to develop California hydroclimate projections.

First, the use of ratio scaling to obtain QDM precipitation can lead to a numerical instability if the denominator in Eq. (1) scales differently from the numerator. QDM requires the construction of empirical CDFs from observationally gridded products (Livneh et al. 2013) and downscaled climate model data (Pierce et al. 2014). There are thresholds associated with each of these products below which precipitation is set to zero. By simply calculating a ratio of the two CDFs, the mapped precipitation quantities can be distorted severely, but the distortion arises due to numerical instabilities in the small but nonzero values of precipitation. This numerical instability can be managed by setting a threshold for when the QDM ratio for precipitation mapping should be set to zero. We used a threshold of $0.1\mathrm{mm}/\mathrm{day}$, which is below the rain-gauge detection limit.

The second numerical challenge arises when precipitation values in the downscaled historical model simulations exceed the maximum observed precipitation in the historical record. Under these conditions, QDM scales the maximum observed precipitation value in the historical record by the ratio of future to historical precipitation for the maximum quantile. There is no clear scaling limit (i.e., Clausius–Clapeyron) to extreme precipitation (Lenderink et al. 2017), and this remains an open issue that could create distortions in projections of extreme precipitation.

The third numerical challenge faced in the creation of a QDM mapping approach pertains to a Mediterranean climate, in which long-tailed precipitation distributions lead to large differences in precipitation values at the highest quantiles. As a practical matter, the implementation of a QDM requires a numerical discretization of quantiles, but the choice of this discretization can matter. QDM for precipitation mapping across California is extremely computationally and memory intensive; an implementation with 1,000 evenly spaced quantiles requires 64 GB of memory to be run without disk-swapping. Therefore, there is a trade-off between discretization error and memory in terms of resolving the structure of the upper-precipitation quantiles. There may be a role for using extreme-value distributions for these quantiles (e.g., Hertig et al. 2019).

After addressing these numerical challenges, we used QDM to produce daily minimum and maximum temperatures and daily precipitation fields for the DEW and WMW cases. The 96 years of mapping, based on climate model projections centered between 2056 and 2085, provided two new sets of hydroclimatology that can be used to bound and anchor water resource projections. Fig. 5 summarizes extreme precipitation for both cases in terms of daily precipitation return values, showing that extreme precipitation becomes more extreme. Whereas Fig. 1 shows relative changes in the precipitation distribution at lower quantiles, Fig. 5 shows that the changes in precipitation lead to generally higher 5-, 20-, and 50-year daily precipitation return values in the DEW case than in the WMW case. This suggests that extreme precipitation is more of a function of the emissions scenario (RCP8.5 driving the DEW case versus RCP4.5 driving the WMW case) than of a particular model, and is broadly consistent with findings from the Fourth California Climate Assessment (AghaKouchak et al. 2018).

With the QDM temperature precipitation fields across the State of California for the DEW and WMW cases, we can understand the water resource implications for changes in hydroclimatology by using the mapped values as inputs into a calibrated version of the VIC model, and then use those results as inputs to the CalSim-II model. A version of the VIC model was configured at $1/16°$ [approximately 6 km (3.75 mi)] spatial resolution throughout California. The data from Livneh et al. (2013) were used as a preliminary data set in the VIC Model setup.

Daily VIC model simulations were performed from 1915 to 2011. The daily runoff and base flow simulated from each grid cell was routed to various river flow locations. For the simulations performed here, streamflow was routed throughout the Sacramento and San Joaquin River basins. VIC model routed flows are considered to be naturalized because they do not include effects of diversions, imports, storage, or other human management of the water resource.

The results of the calibrated VIC model forced by QDM-derived precipitation for the multimodel ensemble, and the DEW, and WMW cases are shown in Fig. 6. These plots indicate approximately a 50% range in water resource availability between the DEW and WMW cases over the Sacramento and San Joaquin River basins. The precipitation received by the WMW and DEW cases provides a bound for the multimodel ensemble [Figs. 6(c and d)], and the multimodel mean snow-water equivalent is also bounded by the DEW and WMW cases [Figs. 6(c and d)].

Using the outputs from the VIC model as inputs into CalSim-II indicated numerous implications for water resource allocations for the state. At a high level, because the ratio of precipitation falling as rain to that falling as snow is expected to increase, there will be a greater need for flood releases early in the year, which will reduce available storage later in the year. For example, during wet years under the WMW case, the Shasta, Folsom, and Oroville dams will operate under flood conditions 90%, 70%, and 40% of the time, respectively. Under the DEW case, the corresponding numbers are 80%, 70%, and 70%, respectively.

In drought years, a minimal number of dead storage months under WMW is indicated, but a much larger number under DEW indicates that the State Water Project and Central Valley Project will struggle to meet operational, regulatory, and contractual commitments under DEW (Table 2).

This work focused on the creation of a range of precipitation and associated water resource conditions for the development of climate change–aware plans for the State of California and at the local level. The approach taken here used a hybrid approach that combines climate models with statistical downscaling and quantile delta mapping to quantify precipitation and water-resource availability, both for moderate warming with California generally becoming wetter, and for more extreme warming with California generally becoming drier. This differences between these two climate realizations resulted in approximately a 50% range in water resource availability, although there were pronounced spatial patterns to the precipitation, especially for extreme events.

There are caveats to this analysis. First, statistical downscaling efforts face a number of well-known issues, and their impact for water resource projections for California needs to be evaluated. The LOCA downscaling was trained with an observationally derived data set (Livneh et al. 2013) and has a number of assumptions that need to be tested. LOCA assumes stationarity in downscaling, but this should be tested to the extent that is practical. Recent work showed that crude and inaccurate representations of topography and, by extension, the snow–albedo feedback, led to the growth of errors in any statistical downscaling approach under climate change, because the downscaling is unable to compensate for insufficient process representations in the parent model (Walton et al. 2020). Process-aware downscaling may be especially apparent in complex terrain in California, which is precisely where most water is derived (Walton et al. 2020). Future work is needed to understand both the processes responsible for, and the temporal and spatial scales of, the decay in skill in statistical downscaling techniques. Information contained in recent results from the North American Coordinated Regional Downscaling Experiment (NA-CORDEX) (Giorgi et al. 2009; Jones et al. 2011b), along with the High Resolution Model Intercomparison Project (Haarsma et al. 2016), may provide insights into these questions.

Second, this work touched on the numerical challenges of using quantile mapping methods, especially for very extreme precipitation ($\tau >0.99$). Fundamentally, changes in precipitation extremes result from processes that differ from changes in precipitation at lower quantiles. Recent focused efforts in Europe through the VALUES project highlighted these challenges and tentatively offered novel approaches for managing these challenges (Hertig et al. 2019; Maraun et al. 2018; Lanzante et al. 2019).

Third, the choice of single model results for providing bookending needs a strong justification. Recent work has supported the model evaluation and severe downselection that was employed here (Sanderson et al. 2017; Herger et al. 2018), although further work is needed to demonstrate how the processes, which lead to acceptable model performance in terms of their response to historical perturbations, are impacted by model-predicted climate changes to determine if further observational analogues can be used to test models’ climate change performance. However, there is a strong intersection between science and practicality for bookending: local agencies need, at most, a minimal number of models to analyze in order to develop plans. Using the average of a multimodel ensemble is a reasonable place to start, in principle, but in practice, model-averaging approaches need to be cognizant of model independence to avoid biases from shared model genealogy (Knutti et al. 2013). Additionally, averaging especially tends to mute precipitation extremes (Risser et al. 2019). Together, these ensemble-navigation challenges motivate the use of a limited subset of models.

Finally, recent work has recognized that the procedure of gridding observations alone can be problematic for extreme precipitation estimates (Sylla et al. 2013). Alternative approaches that recognize the fractal nature of precipitation (Maskey et al. 2016) are needed preserve extremes in gridding; by not doing so, biases in 20-year return value estimates can exceed 30 mm in California (Risser et al. 2019). A careful consideration of historical observations is needed for future projections, especially for precipitation extremes.

Despite these caveats, there is reason to have some level of confidence in the range presented here; the models utilized here have been diagnosed at the highest level with a wide range of ECS. The range incorporates a reasonable range of emissions estimates. Furthermore, the models exhibit skill in capturing the contributions to the modes of variability that are relevant to California, and because their representation of precipitation-controlling factors for California varies widely, model uncertainty in precipitation process representation is being captured. Although additional research is needed to refine model downselection, model downscaling, bias correction, and the use of historical data sets, it is first necessary to address these numerical artifacts in downscaling projections to ensure that (1) the downscaled projections are plausible, and (2) they are consistent with parent GCM projections. Because plans cannot wait for the resolution of the cause(s) of GCM projection biases in California, we present here a bias-corrected range of downscaled projections that, at least by addressing the source(s) of numerical instability in the process chain of producing these projections, enables a path forward for planning agencies in the State of California to understand a range of hydrologic conditions that they might face in the 21st century.

These types of bounding conditions are a crucially important source of information for local planning that considers the hydrological implications of climate forcing. Bounding-condition scenarios provide a range of conditions far more stressful than the mean or median conditions that are often used, and can help planners and decision makers explore how sensitive their systems are to assumptions about future climate conditions. Consistent conditions applied over different watersheds and water districts allow for comparative analysis of vulnerability and risk.

The analyses presented here were derived from several data sets available elsewhere. The data set from Livneh et al. (2013) can be downloaded from ftp://ftp.cdc.noaa.gov/Datasets/livneh/metvars/. The DEW and WMW data sets can be downloaded from ftp://gdo-dcp.ucllnl.org/pub/dcp/archive/cmip5/loca/LOCA_2016-04-02/ by navigating to the HadGEM2-ES and CNRM-CM5 model results, respectively. The VIC model can be downloaded from https://github.com/UW-Hydro/VIC. The DEW and WMW projections and code used to produce those projections can be downloaded from https://bit.ly/2WQBjaE.

This research was supported by funding from the California Department of Water Resources Climate Change Program, with supplemental funding from the Strategic Environmental Research and Development Program under Project RC18-1577. The discussions with Alan Rhoades of LBNL and Minxue (Kevin) He, Elissa Lynn, and Dr. Mike Anderson of DWR also contributed to this work.