Predict anomalies with a daily resolution


Model

The model consists of a feed forward neural network with one input layer, one hidden layer, and one output layer. The network was defined to make predictions whether there was any anomalies the next day or not. The input to the network are daily sums of the planetary magnetic index Kp extending over the last 6 days.

A database of about 700 anomalies from the Meteosat-3 satellite was used for this purpose. The 700 anomalies were distributed over almost 500 days. The database was extended to include the same number of days with no anomalies. The total set, with about 1000 examples, were divided into to sets with equal number of examples to be used for training and testing respectively. After optimization the final network has 6 input neurons, 3 hidden neurons, and 1 output neuron.

Performance

The network correctly predicts 65% of the examples in the test set. It should be noted that the test set is balanced so that there are equal numbers of anomaly and no-anomaly days.

The applet

The model is implemented in an applet that can be run in a Java enabled web browser. The model is loaded to the client so that it runs locally. The data for a selected period is requested from the SAAPS server and sent to the client after which the model processes the data and produces the plot. The applet have been tested on the following platform and browser combinations: It does not run with Microsoft Internet Explorer on Windows 95 or Windows 98 and MacOS with Netscape. The reason for this will be examined.

Run model.
 



Peter Wintoft, 2000-02-17.