Can you tell how probable it is that Adelaide city will be flooded by the River Torrens in 2070 under potential climate change? In this article we look at modelling possible climate conditions, with realistic outcomes.
Why would we bother to do this?
If we can model a probability of destructive flood, we will be in a better position to understand how critical it is that potential risk is mitigated. Let’s look at why modelling past data is useful – and how we might be able to do it.
Different approaches to assessing models
The conventional approach starts with global circulation models (GCMs). These models predict future conditions in many variables such as rainfall and temperature. We then have to use hydrological modelling to translate them to river flow, and then estimate flood potential accordingly.
There exists a large number of GCMs, with emphasis on different physical processes and assumptions. Different GCMs can make contrasting predictions on future climate change conditions. For example, by 2090, different GCMs do not even agree on the direction of changes of annual average rainfall over Australia. (Discover more about the Australian climate projections here).
These differences in GCM predictions can lead us to contrasting future predictions of the flood potential in our opening question. And the problem is – we cannot tell which one is more reliable!
Some studies have reported the performance of individual GCMs in predicting specific climate features (for examples see Johnson et al., 2011). But, we hardly relate these to predicting flood potential. So we still have no idea which GCMs are more suitable for our problem.
Benefits of scenario-neutral approaches
So an opposite way of thinking comes in. Let’s not rely on particular GCM predictions. We can start from as many possible climate change conditions as possible, and model how the flow in Torrens river will respond. Then we can tell which climate change conditions are most likely to cause flood in the Torrens. For example, we may find that changing historical summer rainfall has a big impact on the peak flow, which increases flood potential. We know that to better predict the Torrens flood potential, we should select the GCMs that perform better in summer rainfall.
We call this approach ‘scenario-neutral’, because it does not depend on a ‘climate scenario’ suggested by a particular GCM. As in the above example, the scenario-neutral approach is a useful way of informing the choice of GCMs used to assess the potential climate impact on water resource systems.
The key to this scenario-neutral approach is to:
- Represent many possible climate change conditions. In other words, a ‘climate exposure space’ that a specific water resource system may confront; and then
- Generate continuous climate data for each climate change condition, to enable hydrological modelling.
Currently there is no systematic approach by which to achieve these two objectives.
Below we discuss some key problems in achieving each of them.
We want to include many climate change conditions
As shown in Figure 1, the exposure space should include many possible climate change conditions.
The essential components of the exposure space are:
- Relevant climate variables which are expected to change in the future, such as rainfall and temperature.
- Attributes of the climate variables in (1), that are expected to change our water resource system. The average levels of these climate variables are important. However, often other attributes of them such as variability and peaks are linked to extreme events. (See a list of key climate attributes defined in ETCCDI)
- Possible levels of future change in each climate attribute.
Now we want to include many possible future changes in each of these climate attributes. Some changes may be too far away from historical conditions, and are unlikely to happen. So we first need to define a range in which contain possible future changes in each climate attribute.
For example, Australian annual average temperature is likely to increase between 0.6 and 5.1 °C by 2090, as suggested by many GCMs. We can consider a possible range of changing average temperature as between 0 to +6 °C. As we do not rely on any particular GCMs, we can think of any level between 0 to +6 °C change in temperature as being equally possible. So we can then choose the possible changes in annual average temperature at equi-distant levels within the range.
This is simple if the exposure space only includes one climate attribute. However, when we include many climate attributes, the exposure space becomes high-dimensional. See Figure 2 for this. Then each possible climate change condition is a point in the space, combining a possible change in each attribute. As a result, the selection of equally-possible climate change conditions become more complex.
Essentially, we need a number of points that uniformly span the entire exposure space.
How can we represent possible changes with climate data in a realistic way?
We have considered a large number of possible climate change conditions, so far. The question now is that how these conditions can be represented with actual climate data for hydrological modelling.
This has to be done a realistic way. For example, we want to represent +10% annual average rainfall, with -5 days of annual wet days. How exactly can we change the historical rainfall data to achieve this? Which five raining days should we change to dry days? How can we distribute the +10% average change to each raining days?
Stochastic weather generators are useful tools
Stochastic generators use parameters to produce time-series of climate data in a realistic way. For example, in a generated rainfall time-series, a rainy day is likely to be followed by another rainy day. These stochastic generators are largely used for reproducing historical climate data with specific statistics. To do this, the parameters of stochastic generators are fitted to the statistics of historical climate data.
We can use the advantage of stochastic generators to generate climate data that corresponds to the possible climate conditions in our exposure space. This means that we need to determine individual parameter sets for each climate condition considered.
How do address the problems?
To address the two problems above, we propose a two-step approach:
For problem (1), we need to select a number of combinations of possible changes in each climate attribute. Those attributes uniformly span the multi-dimensional exposure space. To achieve this, we can use an efficient sampling method such as latin-hypercube sampling (LHS).
For problem (2), we need to determine the parameters of stochastic generator corresponding to each climate change condition that is selected. For some climate conditions, this problem can be highly non-linear (Steinschneider and Brown, 2013). Therefore, we can use an optimisation approach to find the most suitable parameter sets.
In this way, we can produce multiple sets of perturbed data from historical data. Each set of perturbed data represents one of the many possible climate change conditions.
How effective is it? A real-life test
We ran a primary test with our approach, in which we aimed to represent possible future changes in rainfall at Adelaide. We used 15 years of historical rainfall data, and aimed to produce 256 sets of perturbed data.
We considered possible change in four rainfall attributes:
- Average rainfall amount (PD)
- Annual average wet days (WD)
- Annual average number of continuous dry days (CDD)
- 99th percentile of rainfall amount on rainy days (Pex99).
We used -50% to +50% as bounds for possible changes in each attribute, for illustration purposes.
Following our approach, we successfully produced 256 sets of perturbed data. These data sets span different levels in each attribute. This means that they are useful to represent various possible climate change conditions.
We are currently working on including more climate attributes to extend the capability of this approach. This will also come with an increase in required computational power. So, we are also looking into options to reduce run-time. We hope to produce an open-sourced R package, to help increase applications of the scenario-neutral approach.
For more details please refer to the original journal paper:
Guo, D., Westra, S., Maier, H.R., An inverse approach to perturb historical rainfall data for scenario-neutral climate impact studies. Journal of Hydrology. DOI:10.1016/j.jhydrol.2016.03.025
Johnson, F., Westra, S., Sharma, A., Pitman, A.J., 2011. An Assessment of GCM Skill in Simulating Persistence across Multiple Time Scales. Journal of Climate, 24(14): 3609-3623. DOI:10.1175/2011JCLI3732.1
Steinschneider, S., Brown, C., 2013. A semiparametric multivariate, multi-site weather generator with low-frequency variability for use in climate risk assessments. Water Resources Research: n/a-n/a. DOI:10.1002/wrcr.20528