Anomaly Models Help


This tool contains a set of models for the prediction of anomalies for specific satellites. The models are trained neural networks that use the daily sums of Kp as input. Each model has been optimized so as to best predict the anomalies for a specific satellite.

In the Model List a model can be selected. Information about the model can be retrieved by pushing the "Inspect model" button. The default model is the "Anom 002".

In the year, month, and day fields the date for a specific prediction can be entered. The default date is the current date. The today's date can also be selected by pushing the "Today" button.

The plot displays the risk of anomaly over a 14 day period. The risk goes from 0% to 100%.

About the models

The models consist of standard feed-forward neural networks with one input layer, one hidden layer, and one output layer. The input is the daily sum Kp value, which is the sum of eight 3-hour Kp values. As the time history of Kp is important the input consist of several sum Kp values delayed in time.

The output is the prediction whether there will be an anomaly or not. Both one day forecast and nowcast models have been developed. Generally to one day forecast only work for internal charging type of anomalies.

The network has been trained to predict the anomalies for a specific satellite. During the optimization procedure the length of the time history of sum Kp and the number of hidden neurons have been varied.

When the optimal network has been found the prediction accuracy can be examined. The prediction of an anomaly is a binary yes/no problem. However, the network output is a real value which is related to how accurate the prediction is. By examining the network output on the satellite anomaly set probabilities can be assigned to the network output. The figure below illustrates this for one model. The solid lines are the probabilities for correct predictions while the dashed line shows the fraction of events in each bin. The bins are marked with the vertical dotted lines. There are five bins and we see that the events are approximately evenly distributed with 20% in each bin. We see that there is a clear correlation between the network output (|y| in the figure) and the probability. Based on the output value from the network we can thus give the probability that the prediction is correct in the plots.


Peter Wintoft, Wednesday, August 29, 2001