Observed record length - implications for analysis using longer time series.

SimCLIM 4.0 for Desktop – outputs/results from SimCLIM 4.0 for Desktop that are generated manipulating the observed data (data browser and extreme value analysis tools) differ if you use a full length of observed records for extreme events analysis rather than the 30 year baseline that is recommended in the training. (e.g. Observatory hill has 155 years of observed data available).

This section outlines an issue I came across doing investor due diligence work and is included below as it is informative on the difficulties of projecting climate using observed daily datasets.

      1. Example for regional BoM sites

Trainee: I am working with these rural sites that often have almost completely continuous precipitation records going back 150+ years.

Some of these records cease in the 70s and 80s.

  1. In your opinion is it better to use a longer run of historical data i.e. 100+ years if it is available
  2. Does it matter if this data ends in the 1990s? i.e. is it preferable to have a data set 1962-2013 or 1860-1996?
  3. When I run the extreme temperature analysis for these sites, once past 2030 (RCP 8.5 high sensitivity) the return period for the baseline returns “errors”. I assume that this is because this projected extreme maximum never occurs in the historical record. Is it ok to explain it in this way to a client?

CLIMsystems response: 1/2. Selection of periods/datasets: There is no simple answer here. The assumption is that no (significant) climate change took place in your historic dataset (no significant trend) and that the fitted distribution function thus properly captures the characteristics of the climate variable at that location. This would argue for a series of about 30 years. However, a 30 year period only has 30 extremes, which is not good for estimating much rarer events (ie. 1 in 100 years). This argues for the longest possible time-series. You have to remember though that the "representative" year of a time-series lies in the middle: for 1860-1996 that would be 1928..., which impacts the application of the proper climate change on that set (you would need to "trick" SimCLIM, by finding the global mean warming between 1928 and 1995, say 0.5 degrees, and then first identify your year/scenario/sensitivity combination in the global table, say 2050:+2.3 degrees, and then find the year that shows 2.3+0.5=2.8 as the year that actually shows the impact from climate change on this dataset by 2050...).

To choose between 1962-2013 and 1860-1996 I would probably a) compare, and b) have a preference for the 1962-2013 set, as the mid year 1986 is relatively close to 1995, the baseline year).

3. Yes. As that temperature is not described by the distribution under the baseline (never occurs), SimCLIM cannot compute how often it happens under the baseline ("never").

Selection of the time period is always a compromise between short enough not to pick up temperature trends that are modelled separately (climate change) and long enough to properly model very rare events. We had many internal discussions as well as discussions with our scientific advisory panel, but have not solved the issue.