Since January of 1991, a research team at the National Oceanic and Atmospheric Administration in Boulder, Colorado has been training a neural network to predict El Nino. According to head researcher, Dr. Vernon Derr, the purpose of the study was to determine if a neural network could accurately predict warm and cold events in the Pacific ocean, and to compare the prediction capabilities of the neural network to other methods, particularly the Persistence method. According to Dr. Derr, the neural network did surprisingly well.
Researchers defined an El Nino or warm event as a departure of more than 1 standard deviation larger than the long term mean in various regions of the Eastern Pacific ocean. If the standard deviation was 1 standard deviation below the long term mean, it was a cold event. While the Persistence method is often used to make weather and climate predictions, it is unable to forecast change or predict the onset of a new situation. The neural network on the other hand was able to show a correlation between the prediction of El Nino and the actual occurrences of warm and cold events in the Pacific. The neural net proved to be a useful device for predicting out to about six months, and depending on the input data, could possibly be useful to the fishing industry.
Researchers used input data found in the Comprehensive Ocean Atmospheric Data Set (COADS). COADS is world wide oceans data giving the sea surface on a monthly basis since 1884. Because warm events occur every five to seven years, and because each event is unique in terms of duration, onset and decay, the statistical character of each even is quite varried. As a result, an event is difficult to predict by any means.
For input, Dr. Derr's team primarily used ship's data from various part of the Pacific Ocean dating back to Mathew Fauntainmaury who was the original oceanographer in the Navy. It includes wind, air temperatures, surface temperatures and southern oscillations which is a comparison of sea surface and pressure between Darwin, Australia and Tahiti. It is a know fact that this difference in pressure occurs during El Ninos, but not before, so it is therefore not useful in predicting them.
One part of the research study was to determine the best set of data. According to Dr. Derr, "the set of data we used to predict things over the last year is probably not the ideal set, and we will be using a different set in the future." Because the team used most of the available data for training, only 10% of the data was left for validation and this remaining 10% may not even encompass a period in which El Nino occurred. The network was trained using using the standard sigmoid transfer function. Using the genetic algorithm method, the team studied learning rates and tolerances to determine the best set for the data set they were using. They also varried the number of hidden neurons to determine the optimum number, but have not yet gone to more than 1 layer;. although according to Dr. Derr that is something they want to do in the future.
Testing went as follows: In January of 1991, the team started predicting skill scores -- actually the RMS differences between the actual ocean temperature and the predicted temperature-- for up to six months ahead Then in February of 1991, they again predicted (on the basis of current data) for 1 to 6 months ahead. They continued in this manner up until the present time. According to Dr. Derr, "Those were quite good in the sense that the RMS skill differences were in the order of less than a degree averaged over a long period." However one of the problems was that neither the Persistence method nor the neural net did a thorough job of predicting the onset of the warm or cold event. This fact leads Dr. Derr to speculate that the data was not sufficient for the purpose and that it should include not only at the sea surface temperature in Region 4 of the Pacific, but also at least a nine or ten year the history of it.
Dr. Derr plans on concluding his studies at the end of 1993. In the means time he plan to employ a rather unusual validation process. The team will train using all the data they have and then they're going to find skill scores for the same period of time again using all the data they have.