Predicting Carbon Dioxide Concentration using Gaussian Processes

Back in 2005 I used a Gaussian Process model to predict the atmospheric concentration of carbon dioxide over a twenty year horizon, as described in Gaussian Processes for Machine Learning (with Chris Williams). Now (2011) that a little more data has become available, I thought it would be fun to see how well the predictions have stood the test of time.

The data used to be available at
http://cdiac.esd.ornl.gov/ftp/trends/co2/maunaloa.co2, but now it seems that more up-to-date data can be found at ftp://ftp.cmdl.noaa.gov/ccg/co2/trends/co2_mm_mlo.txt. The data at these two sources seem to differ slightly - below I use the second source.

I used the matlab/octave gpml toolbox and the following script to make the predictions. The original predictions used an older version of the software, but the model is the same.

Below are the training data (in blue), the predicted 95% confidence region (in grey) and the more recent observations (in red). The actual CO2 concentration grew at close to the fastest rate deemed likely by the model.

mauna1.jpg

Here is another plot, focusing on more recent times and the future.
mauna1.jpg

By Carl Edward Rasmussen, April 20th 2011.