Sunday, 5 February 2017

Minimum temperatures

We have recently had some very warm nights, including the warmest recorded in our home location.  However this wasn't reflected in the daily minimum temperatures.  This appeared to be because the temperature dropped later in the day and my weather station records on the basis of a "calendar day".  The issue is shown by this graph.
The early morning of 24 January was very warm, with a recorded minimum of 21.1oC at 3am, marked with a red star.  However the next evening was a lot cooler and by  11pm had dropped to 17oC.  As indicated above my "calendar day" approach records the day's minimum as 17oC.

Cutting to the chase, I have concluded that for the purposes of recording the weather over a longer term it is not worth the extra effort to recalculate minima based on a different time of starting a day.  Where a different purpose is in mind - eg assessing the comfort levels over a shorter period such as a month - a specific analysis could well be appropriate.  There follows some detail.

The BoM records temperatures based on a day starting at 9am against the day on which the recording was made.  However at 9am on 23 January the temperature was (again, by coincidence) 17oC which would be the recorded minimum for 24 January 2017.  This led me to contemplate how much difference there would be between minima recorded on a day beginning at 9am and one beginning at midnight.

My first approach was to compare the minimum temperatures recorded between midnight and 9am (on the basis that the temperature usually dropped over night) and the calendar day minima recorded for January 2017.  This correctly showed the 21.1oC and 17oC and a few other similar, but less dramatic, differences.

I then extended this analysis to cover all 1148 days for which I have records which showed that 80% of days didn't change.

The second major approach was to adjust the date for times greater than 9am (which in the BoM methodology get recorded the following day).  With a little consequential fiddling about I ended with a series of minima based on a day ending at 9am.  The expected consequence of this is that because a relatively cold morning on day n would be colder than the early morning temperatures of day n+1.  Thus my calendar day minimum would be higher than the BoM timed series.  And so it turned out to be..

In the following chart the series labelled "night" represents the total set of midnight -9am minima compared with the calendar date series while that labelled "full" is the full switch to a day ended 9am.
As expected the full shift produces some shifts where the calendar approach overstates the minimum based on a BoM approach.  In summary 72% showed no difference, 10% had a higher minimum for the 9am base and 18% were lower.

An important overall statistic is that the two series show a correlation coefficient of 97%: in other words the pattern is VERY similar.  Given that the re-calibration to get a 9am based series involves ...
  • a modest amount of extra effort; and 
  • a reasonable possibility of introducing error greater than the potential improvement
.. . I conclude that the cost-benefit ratio of trying to change to a 9am base for general analysis of minimum temperatures is not positive. I think the same issues apply to other measures (eg maximum temperature, and thus  conclude that I will stick with what I have been doing, supplementing this with ad hoc analysis as required.

No comments: