Identifying critical climate conditions for use in scenario-neutral climate impact assessments

Key points

• It is critical that scenario-neutral assessments consider climate conditions to which system performance is most sensitive.

• A generic approach is presented to enable such critical conditions to be identified and applied to a reservoir case study.

• Results demonstrate the utility of the approach as different attributes are critical for different objectives.

Abstract

Scenario-neutral climate impact assessments are being used increasingly to assess water resource system responses to possible climate changes. The purpose of such assessments is to identify system sensitivity to a range of plausible climate conditions, often including an evaluation of the joint effect of multiple climate stressors. Given the large number of climate variables and associated statistics (averages, seasonality, intermittency, extremes, etc.) that could plausibly change and impact on system performance, it is essential that scenario-neutral assessments focus on those to which the system is most sensitive. A generic approach to identifying these ‘critical’ climate conditions is presented and tested on the Lake Como reservoir in Northern Italy considering two system objectives: irrigation deficit and flood reliability. Results indicate that different climate conditions are critical for the two objectives, and that the choice of which climate conditions to include has a significant effect on the climate impact assessment outcome.

Introduction

Climate change can impact water resource systems through stresses to both supply and demand (IPCC, 2014). These stresses are the result of changes to atmospheric variables such as precipitation, temperature and evapotranspiration, which influence water resource systems via a complex set of catchment-scale and system-level processes that in turn are dependent on the system's geography, configuration, operation and performance measures. Understanding the possible impacts of climate change on water resource systems therefore requires mapping changes in large-scale climate processes to changes in system performance, accounting for the unique features of each system (Mastrandrea et al., 2010).

To this end, scenario-neutral climate impact assessments are being used increasingly to assess and convey the sensitivities of water resource systems to changes in climate (Brown and Wilby, 2012; Prudhomme et al., 2010). These assessments work by ‘stress testing’ a system against a set of hydrometeorological time series that represent potential future climate conditions. These time series are then run through a system model, providing information on how the system responds to changes in climate conditions, and identifying critical performance thresholds and other decision-relevant information (Brown et al., 2012; Prudhomme et al., 2010). This creates a scenario-neutral space (also referred to as an exposure space or a response surface) that maps system performance to changes in a range of statistics (e.g. averages, seasonality, intermittency, variability, extremes and low-frequency autocorrelations) of climate variables (e.g. precipitation, temperature, potential evapotranspiration)—the combination of which is hereafter referred to as climate ‘attributes’. The scenario-neutral space can be coupled with climate projections to understand the likelihood and possible timing of these changes (Taner et al., 2017; Turner et al., 2014); moreover scenario-neutral analyses can be used to explore the robustness of different management strategies to the set of plausible future changes (Brown et al., 2012; Culley et al., 2016; Whateley et al., 2014).

The primary stages in the construction of scenario-neutral spaces consist of:

• Selection of the climate attributes against which to stress-test the system (this usually corresponds to the selection of the axes of the scenario-neutral space).

• Development of a set of perturbed attribute values, representing the plausible changes to be used for stress testing (this is equivalent to selecting the locations in the scenario-neutral space at which to assess system performance).

• Generation of climate perturbed hydrometeorological time series that reflect the perturbed attribute values (i.e. generation of climate perturbed time series that represent each selected location in the scenario-neutral space).

• Assessment of system performance at each location in the scenario-neutral space by passing the climate perturbed hydrometeorological time series generated in (iii) through a system model and calculating the corresponding system performance.

A number of methods have been developed to support the above stages. For stage (ii), different approaches have been used to determine the combinations of changes in climate attributes for which to assess system performance. When the system is sensitive to only a small number of climate attributes—such as only changes to annual average rainfall and temperature but not to changes in seasonality, intermittency, variability, extremes or low-frequency oscillations—then it becomes possible to have a low-dimensional (e.g. 2- or 3-dimensional) scenario-neutral space. In these cases, system performance can be evaluated over a regular grid of changes to each attribute (Brown et al., 2012; Culley et al., 2016; Prudhomme et al., 2010; Steinschneider et al., 2015; Turner et al., 2014). In contrast, when the dimensionality of the scenario-neutral space is high (e.g. larger than 10), a variety of sampling techniques have been used to cover representative regions (e.g. Latin hypercube, improved distributed hypercube, eFAST), as the use of a regular grid would be too computationally expensive (Beachkofski and Grandhi, 2002; Gao et al., 2016; Kasprzyk et al., 2013; Stein, 1987). In practice, such high-dimensional scenario-neutral spaces thus far have only been used in cases where changes in climate attributes were combined with other factors affecting system performance, such as demands and costs (Herman et al., 2014; Kasprzyk et al., 2013; Ray et al., 2018; Shortridge and Guikema, 2016), with the number of climate attribute dimensions considered still being small (i.e. 2 to 3).

For stage (iii), different approaches also have been developed and/or used to generate hydroclimatic time series that represent the desired attribute changes. As discussed in Guo et al. (2018), these time series traditionally have been obtained by applying scaling factors to historical data (Culley et al., 2016; Kasprzyk et al., 2013; Kay et al., 2014; Weiβ, 2011; Wetterhall et al., 2011). However, this approach cannot account for changes in attributes such as intermittency, autocorrelation and extremes, or for complex combinations of changes such as an increase in extremes coupled with a decrease in mean. To address this limitation, stochastic weather generators have received increased attention as they have more flexibility to simulate complex combinations of future changes (Borgomeo et al., 2015a; Guo et al., 2018; Ray et al., 2018; Steinschneider and Brown, 2013). However estimation of stochastic weather generator parameters to simulate time series with pre-determined attributes values can be a challenging task (Culley et al., 2019; Guo et al., 2017b, 2018). To address this, Guo et al. (2018) introduced an inverse approach that optimizes the weather generator parameters in order to generate hydrometeorological time series with desired attribute values; this approach was further formalized and refined by Culley et al. (2019).

Despite these advances, a formal approach to identifying the attributes in stage (i) is still missing. For reasons likely to include some combination of a priori judgement, analytical tractability, ease of interpretation and capacity to generate combinations of perturbed attributes, most studies consider changes in two attributes (Bussi et al., 2016; Culley et al., 2016; Singh et al., 2014; Turner et al., 2014; Weiβ, 2011; Wetterhall et al., 2011; Whateley et al., 2014). The selected attributes usually correspond to changes in mean precipitation and mean temperature, or in some cases changes to precipitation seasonality (Kay et al., 2014; Prudhomme et al., 2013a; Prudhomme et al., 2013b) and shifts in peak flows (Borgomeo et al., 2015b; Nazemi et al., 2013; Quinn et al., 2018). However, it is well known that water resources systems can be affected by a much wider range climate attributes (e.g. averages, seasonality, intermittency, variability, extremes and low-frequency autocorrelation), which can change both individually and in combination. For example, groundwater storages are less affected by evaporation than reservoirs but also subject to lower frequency variations in climate (e.g. multi-year droughts), while temperature can be critical for water storages driven by snowmelt (Culley et al., 2016; Eckhardt and Ulbrich, 2003; Ray et al., 2018). Importantly, even for the same system, the climate attributes that have the biggest influence on its performance can vary depending on the performance metrics (Culley et al., 2016; Kasprzyk et al., 2013). This highlights a need to rigorously test the assumptions of stage (i), with omissions of important attributes at this stage potentially leading to the misinterpretation of a system's climate sensitivity, the inability to identify important system failure modes, and/or the incorrect assessment of the relative performance of different decision alternatives.

Given that identification of key system sensitivities is a fundamental principle that underpins scenario-neutral climate impact assessments (Brown and Wilby, 2012; Nazemi and Wheater, 2014), a formal approach for selecting the most appropriate axes of scenario-neutral spaces is warranted. This is because all subsequent stages in scenario-neutral assessments are conditioned on this stage, making it impossible to identify key sensitivities in subsequent analyses if key climate attributes are not considered in the first place. So while stochastic weather generators are opening up the opportunity to simulate a much larger set of climate attributes as part of stage (iii) (Borgomeo et al., 2015b; Guo et al., 2017b; 2018; Herman and Giuliani, 2018; Quinn et al., 2018), for this to be effective it must be accompanied by the appropriate selection of the attributes to be perturbed in the first place as part of stage (i).

The overarching aim of this paper is to develop an approach for identifying the most important climate attributes, so that scenario-neutral climate impact assessments provide accurate information on system sensitivity and the conditions that could lead to system failure (Broderick et al., 2019; Bryant and Lempert, 2010; Culley et al., 2016; Lempert et al., 2008; Prudhomme et al., 2010). This will not only help improve the accuracy of system ‘stress tests’ when applied to current system configurations, but can also help support assessments on the climate conditions for which an alternative system configuration (sometimes referred to as an ‘option’, ‘solution’ or ‘decision’) is preferable to another (Brown et al., 2012; Brown and Wilby, 2012; Giudici et al., 2020; Groves and Lempert, 2007; Hadka et al., 2015; Herman et al., 2015; Kasprzyk et al., 2013; Lempert and Collins, 2007), as well as the relative or absolute robustness of alternative system configurations (Herman et al., 2015; McPhail et al., 2018; McPhail et al., 2020).

To this end, the specific objectives of this paper are (i) to present a general approach for identifying which climate attributes to include in scenario-neutral studies for a given system and performance objectives, and (ii) to evaluate the benefits of using the most appropriate climate attributes in a scenario-neutral impact assessment. The approach is demonstrated and tested using the Lake Como system, a regulated lake in Northern Italy. Two performance criteria are considered—flood reliability and irrigation deficit—to highlight the importance of tailoring the approach to the system under consideration, including different system objectives (Kasprzyk et al., 2013).

Details of the proposed approach can be found in Section 2, with a description of the case study and implementation of the proposed approach in Section 3. A description of further analysis designed to test how well the proposed approach performed is also presented in Section 3, as well as a demonstration of a tipping point type impact assessment (Brown et al., 2012; Frey and Patil, 2002; Guillaume et al., 2016; Haasnoot et al., 2013; Hyde et al., 2005; Raso et al., 2019; Ravalico et al., 2010) using the critical climate attributes. The results of implementing and testing the proposed approach are presented in Section 4, along with a discussion of its advantages and limitations. Conclusions are presented in Section 5.