This is part three of a series on Global Warming using
records from the Global Historical Climatology Network. In my first post of
this series I discussed my data source, my means of selecting which records to
use, and the time frame of study. I used only station records which were
complete for the time frame of study. I had determined not to impute or
estimate any data.
The previous post of this series described the results from
that data in terms of number of days above 85°F,
90°F, 95°F, and 100°F as well as the number of days where the daily highs did
not exceed 32°F, 20°F, and 10°F. As we saw, for the locations involved, the
number of warmer days has consistently fluctuated up and down at apparently
regular intervals, but have generally decreased since the 1930’s. The number of
colder days has consistently fluctuated up and down in a repeated pattern. For
the number of days at or below freezing the data indicate the years of 1900 to
1930 and 1980 to 2003 are nearly identical. There is evidence to suggest that
pattern began to repeat again in 2005 to 2010.
As I explained in the first post of the series there are
certain limitations to this study which bear repeating here. The long term data
from the GHCN daily max min tables is very limited. Most of the coverage is in
the lower 48 states of the US. Canada, Europe, Central Asia, and Australia are all
represented but to significantly lesser degrees. Africa, Central America, and
South America are not covered.
With those limitations in mind, let me describe in general
terms the process I used to develop the data into useful information.
Goal defined
When performing an analysis
for this type of data the goal is to develop a model which accurately describes
the data and then determine a means of applying that model to other, similar
situations which fit the definitions of the model. This is a process which involves creating a model, testing the model
against known data, evaluating the results, and adjusting the model
accordingly. An accurate model will be able to perform accurate predictions for
known data within acceptable levels of data variation. A model which cannot
make accurate predictions against known data is flawed and therefore would not
be useful.
Creating the Model
My first pass approximation for such a model was the
simplest model available, which is a raw average of all the data. I chose to
test this model by comparison to selected samples of individual station data. Without
going into details, let me just say this initial model failed. For example, the
model failed to describe the general time series trend of individual stations. Meaning
where things started off and where they ended up. That failure informed my
method for refining the model.
I refined my data model to address that failure by creating
a new data set consisting of beginning and ending temperatures for each
station. I analyzed that data by calculating a temperature change delta for
each station and performing statistical analysis on that data set.
I found quantifying stations by the overall start to finish
temperature change, the temperature delta, produced a near normal data set. Using
this as a starting point I refined my model in to three separate models. One model
covers the -1° to 1° range which contains 75% of all stations. The second model
covers the 1° to 4° range which contains 17% of the stations. The last model
covers the -1° to -2.5° range which contains the last 8% of the stations.
As before, I tested the data models against actual station
data with model selection based upon the temperature delta parameter. Again,
the models failed to accurately reflect all the data. Quantifying that failure
was easy as all failures occurred in Australia. Separating Australia as a
separate data set and performing the same analysis as above, I created two
additional models. The number of models is now three for the northern
hemisphere and two for Australia. A total of 5 models.
These five models, based
upon two selection parameters, are accurate within ± 1° for over 75% of the
individual stations and within ±1.5° for the remainder. This is an acceptable degree
of error in my opinion.
Refining and Utilizing the Model
One of the primary reasons for creating a model, beyond
defining what has gone before, is to act as a predictive tool. Having defined
models which describe what is known to have happened it now becomes necessary
to try and define and quantify those factors which affected what happened.
Therefore, it is helpful to have five different models and a wide range of
results. These five models can be further reduced into three essential models
based upon the over all results: Temperatures rose, temperatures fell, or
temperatures did not change appreciably. Those are distinctly different
outcomes.
Each station in this data set has been affected by certain
factors. I will define those factors as local, regional, or global. Local factors
contribute to unique site results. Regional factors contribute to site results
over a certain area. The scope of regional factors may vary quite a bit. Global
factors would contribute to outcomes all over the world.
The process going forward is defining and quantifying local
factors, regional factors, and then global factors. In the process of doing so
the various models become refined to include those parameters. Ideally, they
will be combined into one or two models. The process of model redefinition
would as always include testing the models for descriptive and predictive
accuracy.
The Example of Australia
Australia is an interesting case study for this method.
There are only 10 stations with usable data. However, this data extends back to
1895. Australia not only has a distinctive regional difference, there are
distinctive local differences. Australia is also a mostly sparsely populated
place. One factor which became apparent immediately was population density. Given
the time frame involved, it is reasonable to assume the initial population
density is essentially zero. Predicting which model applies to a site by
current population density proved 100% accurate. There are sites which are
geographically close but far apart by population totals. The magnitude of
temperature increase over the period 1950 to 2005 between these proximate sites
was as much as 3.5° higher for the more heavily populated site.
When you consider the generally accepted figure for
temperature rise over the past century due to CO2 is 1.5°, a 3.5° temperature
increase differential over 55 years due to a population increase differential
would be significant. Accepting those numbers as reasonably accurate and
assuming 1.5° as the result of Global factors, and assuming linearity, the inference is
the localized factor of population growth has a greater influence on local
temperatures than global factors by several orders of magnitude.
Moving Forward
Understand, this process is far from complete. I am
presenting this information on what I am working on essentially in real time as
the process progresses. I am letting the data lead me and not the other way
around. I may end up somewhere totally unexpected. Even so, I think it is
worthwhile to make these posts. I would certainly welcome constructive input.
Next post: The five models.
No comments:
Post a Comment