3.1. Downscaling for hydroclimatological assessment
Global-scale modelling endeavours are useful to drive climate change policy and give overviews of large-scale hydrological changes (see Sanderson et al., 2011; Todd et al., 2011). However, there is a spatial disparity between what GCMs can offer and what water resource managers require to make decisions on water infrastructure and policy (Buytaert et al., 2010); therefore downscaling coarse GCM information to a higher spatial resolution is necessary for most hydroclimatological assessments (Varis et al., 2004; Hashmi et al., 2009). Qian et al. (2010) showed that stochastically downscaled information better reproduced extreme hydrological events than data taken directly from GCMs.
There are many different approaches to downscaling coarse resolution GCM information for use in hydrological impact studies, and various review papers have shown the strengths and weaknesses of each (Xu, 1999; Hanssen-Bauer et al., 2005; Xu et al., 2005; Fowler et al., 2007; Maraun et al., 2010).
Most hydrological impact assessments require time series of weather variables (chiefly precipitation and potential evapotranspiration (PET)) on a daily time-step (Kilsby et al., 2007). The most readily available source of this information is the instrumental record, so hydroclimatological studies have often been based around ‘scaling up’ previous flood and drought events using average monthly change factors from GCMs (Scibek and Allen, 2006; Leander et al., 2007; Boukhris et al., 2008) a technique often referred to as the change factor method (CFM—Jackson et al., 2011) or an ‘implicit’ approach (Zhang, 2011). This process does not allow for changes to climatic variability and often uses short instrumental records, leading to underestimations of future hydrological extremes (Semenov and Barrow, 1997; Holman et al., 2009; Zhang, 2011).
The CFM assumes that the climate of the past is analogous to the climate of the future (or even present), which in terms of variability and seasonality it is not, as shown by large-scale climate modelling (Solomon et al., 2007). An example relevant to the water industry is that using change factors gives an equal number of precipitation occurrence days in the future as the past for no other reason than that was the number in the particular baseline period in the relatively short instrumental record (Diaz-Nieto and Wilby, 2005). The inadequacies of not accounting for climate variability when downscaling for hydrological purposes has long been known (Srikanthan and McMahon, 2001), yet the technique remains in use as a result of its simplicity and inexpensive computational demands (Diaz-Nieto and Wilby, 2005).
The detail and spatial resolution that is suitable when assessing the impact of climate change on water resources will vary from catchment to catchment based on some perceived risk (Todd et al., 2011; Hall et al., 2012). Greater depth of analysis should be afforded to areas with high proposed investment in adaptation of the water resource system than to those where no investment is planned (Hall et al., 2012).
Given the myriad of available climate model downscaling techniques, each with their own particular strengths and limitations, selecting the correct method to use depends on the application (Wilby et al., 2009). WGs have particular attributes that render them a distinctly useful approach for detailed assessments of the impacts of climate change on vulnerable water resources at a high spatial resolution (Diaz-Nieto and Wilby, 2005; Kilsby et al., 2007). Primary amongst these attributes lies the allowance of changes to climate variability and the creation of potentially endless synthetic sequences of temporally-consistent weather information that permit the projection of meteorological (and thus hydrological) extreme events (Wilks and Wilby, 1999; Hulme et al., 2002; Kilsby et al., 2007; Jones et al., 2009) at a suitable temporal resolution for inputting to biophysical models.
In a study on groundwater recharge under climate change forcings, Holman et al. (2009) recommended stochastic modelling is used to assess vulnerable or sensitive groundwater systems, thus enabling improved understanding of future risks of drought severity and persistence as well as high recharge years causing groundwater flooding. However, this level of detail would not always be required: using dynamical downscaling approaches with no assessment of extreme events (such as Cloke et al., 2010) would suffice in areas with less risk (Hall et al., 2012; Todd et al., 2011).
3.2. The stochastic weather generator
WGs are a form of statistical downscaling of coarse climatic data from GCMs, where statistical relationships between large-scale climatic variables and small-scale hydrometeorological variables are searched for. Essentially a collection of stochastic models, WGs create a distribution of plausible estimates of a particular weather climatic parameter (Boukhris et al., 2008). The basics of stochastic modelling have long been available (Matalas, 1967; Richardson, 1981), and have spawned a huge array of WGs, notably WGEN (Richardson and White, 1984), LARS-WG (Rackso et al., 1991) and CLIMA (Donatelli et al., 2005). For a technical review of different stochastic modelling approaches see Wilks and Wilby (1999) and Maraun et al. (2010).
WGs have historically been used as a method for infilling missing or erroneous weather records (Wilks and Wilby, 1999), and so are designed to recreate an array of observed weather variables as accurately as possible. The skill of the WG is determined by validating this baseline synthetic weather sequence against the instrumental record (e.g. Min et al., 2011).
The basic premise of adapting WGs for future climate projection is the assumption that statistical relationships between climatic parameters in the present (or past) will remain constant in the future. Therefore, it stands to reason that by forcing a WG with the fundamentals of future climates garnered from climate model information, weather sequences typical of future climate scenarios can be produced. The effect of these sequences on hydrology and water resources can then be explored through hydrological models then compared to a baseline, thus constituting a hydroclimatological impact assessment (Wilks, 1992).
WGs enable climate change impact assessments to be conducted at greater resolution in space and time than regional climate models (RCMs) allow, and are particularly relevant to studies in which the sequence of events is important, such as water resource provision (Wilks and Wilby, 1999; Jones et al., 2009). Studies comparing the ability to determine climate change impacts on hydrology of statistical downscaling techniques (such as using a WG) with other methods have been carried out. Diaz-Nieto and Wilby (2005) suggested that there is a place in research for both, with the coarser-resolution dynamical downscaling approach used for ‘broad-brush’ high level assessments of vulnerability (Sun et al., 2007; Bates et al., 2008; von Christierson et al., 2011; Dai, 2010; Todd et al., 2011), and statistical downscaling techniques delving deeper to explore detailed impacts deriving from sequencing and persistence of daily events, normally once vulnerable water resources have been identified (Diaz-Nieto and Wilby, 2005).
Combining RCM ensembles with stochastic WGs to create daily weather parameters for future climates has become an increasingly-used method for performing hydroclimatic impact assessments. For example, Herrera-Pantoja and Hiscock (2008) used the CRU WG (Jones and Salmon, 1995) to assess the impact of climate change on groundwater recharge at three sites in the UK, finding significantly increased dry periods leading to a reduction in recharge at each site as the century progresses. Each site presents increased climatic variability in the future, with the dry season found to be particularly affected. They conclude that sites already under groundwater supply pressure will come under increased stress as the century progresses.
Single-site WGs, such as EARWIG, CRU WG and UKCP09WG, are the most commonly used and least complex form of WG and therefore have the advantage of being computationally inexpensive (Semenov, 2008; Wilby et al., 2009). Multi-site WGs are more complicated and not part of the suite of tools provided by UKCP09. As a result of this commercial unavailability multi-site WGs are not currently useful for estimation of future deployable outputs in the UK water sector. For a review of multi-site and full-field WGs see Maraun et al. (2010).
After Fowler et al. (2007) and Bates et al. (2008) there has been a move within the hydroclimatic research community towards providing decision-making tools for future planning and management rather than focussing on more in-depth comparison of downscaling methods. At the same time, bespoke single-site future WGs with science-hidden interfaces such as EARWIG (Kilsby et al., 2007) and UKCP09WG (Jones et al., 2009) have become available, greatly simplifying the process for carrying out a WG-based hydroclimatological impact methodology in the UK and overcoming the issues of low awareness and user-friendliness that held back the take up of WGs (and other forms of statistical downscaling) in the past (Diaz-Nieto and Wilby, 2005; Groves et al., 2008).
The process of creating daily future weather sequences using a WG now requires no manual data input, prior knowledge of climate modelling or the need to develop local-scale WGs from scratch as was previously necessary (Varis et al., 2004). Such ‘science-hidden’ tools (Fowler et al., 2007) allow non-specialist end users to use the WG approach effectively, facilitating more widespread uptake in industry (e.g. Severn Trent Water Ltd, 2011). This approach does, however, make a WG less flexible: without the ability to take the model apart for further development by third parties, end users can be hamstrung by the omission of a particular variable. In the case of UKCP09WG, a lack of wind speed information reduces its effectiveness in many sectors, particularly railways.
Limitations of WGs remain. UKCP09WG, for example, is inhibited by an inability to produce the most extreme meteorological events, and in particular is not set up to recreate blocking regimes that create heatwaves/droughts and exceptionally cold winters (Jones et al., 2009). The February 2011 upgrade of UKCP09WG has substantially reduced the impact of this problem, however extreme high return periods of any given meteorological event should still be treated with caution. Furthermore, the single-station nature of most commercially available WGs creates a problem in that a weather sequence produced at one site will not correspond in time with another station nearby, so an extreme event at station A will not occur on the same day as it does at station B, even if in reality those stations would be subject to the same large-scale weather system (Jones et al., 2009). The size of the site can be increased (in the case of UKCP09WG, from 5 km2 to 10 000 km2), but this involves spatially-averaging the area, thus reducing accuracy.
These issues, despite reducing the ability to produce realistic projections of future weather sequences, should not deter those in industry from using WGs as the fundamental advantages of the approach over simpler CFMs are substantial. It is important to strike a balance between continually improving the skill and complexity of WGs and actually using them to make real-world decisions.
3.3. Case study: Weir Wood Reservoir
In a direct comparison of WG-based and CFM-based appro- aches, Harris et al. (2009) use an ensemble of RCMs to drive a weather generator (EARWIG) to assess climate change impacts on hydrological multi-seasonal drought events at Weir Wood Reservoir in North Sussex, UK. Drought periods are identified by precipitation totals over three consecutive winter half-years (October to March). Using the weather generator approach, it is found that inflows to the reservoir during future drought events are substantially below levels found in the 102 year instrumental record, and a regular period in the 2080s would constitute extreme multi-seasonal drought today (Figure 2). It is found that the GCM used to drive the weather generator is more important than the emissions scenario used, showing the need for the move towards large ensembles of GCMs or probabilistic information that has been seen in recent years.
Figure 2. Simulated cumulative inflows at Weir Wood Reservoir over three consecutive winters for the 13th ranked drought periods. Two GCM/RCM combinations and two emissions scenarios (A1FI and A1B) are used. The central line denotes the average of the models, with the upper and lower bounds representing the model range
Download figure to PowerPoint
A further analysis of the climate change impact on reservoir yields at Weir Wood shows substantially increased pressure on the reservoir in the 2050s and 2080s during drought episodes than in the baseline period (that is, yields for a particular drought rank within a dataset are lower in the future than in the baseline period) (Figure 3). All modelled simulations of the 2050s and 2080s project that yields during equivalent-ranked drought events will be much lower than in the baseline simulation (1960–1990). The yield during the worst drought in the instrumental record of 8.9 MI day−1 is surpassed regularly in all scenarios.
Figure 3. Simulated yields during droughts at Weir Wood Reservoir during the 21st century. (a) 1st ranked 3 winter droughts; (b) 7th ranked 3 winter drought; (c) 13th ranked 3 winter droughts. Ranks were chosen according to their close relation in the baseline simulation with periods in the instrumental record (the baseline simulations shown at 1975 in this figure correspond to the precipitation totals during real drought events in the period 1918–2006). Central lines denote the average of the models, with the outer line showing the range of output
Download figure to PowerPoint
With the CFM applied to the instrumental dataset, less severe hydrological drought events in the future are indicated and the sign of change is not always certain (Table 1). This occurs primarily as a result of less substantial PET increases in the future compared to the WG approach, and the inability of the CFM technique to account for changes to climate variability. The number of RCM/GCM combinations used to drive the weather generator or to obtain the change factor values is four, except the 13th ranked simulated droughts, where two combinations were analysed. This represents only a portion of the climate modelling uncertainty, and a larger amount of simulations would be needed to obtain robust statistics against which decision-making could be based. However, the difference between the two approaches in terms future drought severity projection is clear.
Table 1. Changes to total inflow at Weir Wood Reservoir in the 2080s compared to baseline and instrumental conditions
| ||1919–1922/1st ranked simulation drought||1970–1973/13th ranked simulation drought|
|Change factor method||− 5.4% (−38.38 to + 24.3%)||− 1.86% (−28.93 to + 21%)|
|Weather generator method (A1B emission scenario)||− 74.47% (−83.05% to − 65.7%)||− 69.07% (−79.13 to − 59.01%)|
Harris et al. (2009) show that the WG is able to capture variability and change in droughts in the latter twenty-first century better than the change factor approach. Crucially, the periods of high evapotranspiration within the synthetic dataset that is the stimulus for the major multi-seasonal droughts of the 2080s are not apparent in the perturbed data. The increases in evapotranspiration need to be further investigated to determine exactly why they are occurring at such a greater rate in the WG approach than the perturbation approach. It may be the case that differences in the methods of PET calculation account for some of the disparity.
As this work does not use probabilistic climate information it would be inappropriate for use as the basis of a water resource decision-making process in the sub-catchment. However, the project does show that there is significant scope for underestimation of hydroclimatological impacts in the future when CFM methods are used.