15. FAQ: Artificial Neural Network

15.1. What is the ANN tool?

With this interface you can create and train an optimised forecast with an Artificial Neural Network (ANN) to predict the future of a variable based on different input curves from teh MSEPS system.

The special feature of the ELFI-ANN is that you do not have to create a ANN from scratch, but use already precomputed enproducts and mix uncertainty estimates in form of minimum, maximum, mean, percentiles and other optimised or weighted combinations of forecasts of the end product parameter together and form a ANN that is optimised on the parameter(s) of your choice and requirements.

Example: The ANN gets data from production, from f1 curve (optimal, RMSE optimised forecast) and a percentile p70. ANN is now searching for patterns to create a new forecast from this database.

ANN optimises also by identifing outliers and verifying whether there is a physical relation betweeen the chosen forecast variables and the measurents or whatever the user chooses to use as the “truth”. ANN makes an independent forecast if the optimal forecast is good.

15.2. How can I create a artificial neural network curve?

1. Select the forecast site and variable
2. By using measurement data for the ANN you have to select where the data comes from

• from a upload (see how to upload a file and in which format this has to be done here)
• from a already registered curve (once a file is uploaded and stored you can use this

funcionality, curve should be stored in u1-u4

3. select the input time frame (you have to key in the time frame from the production file or if uploaded this time frame is given automatically but you have to confirm it)
4. now select the input curves; this curve should give the ann support to find regularity (you should not use more than 3 curves), select the curve by click on the curve name, the curve will be displayed in the selected field
5. select other forecast varaiables if necessary; note: it is not necessary to select the variable already given in section 1
6. Ann can try to detect measurement patterns on time variables which you can select here Note: you can deselect by double click on the variable or curve name at left side.
7. In the data distribution you can select in what percentages the data should be used for what part of the process. At the beginning there is no change necessary. The more experience you get with the ann you can configure new setups.
8. The settings will be listed and can be edited
9. To start the ANN click on „Train Artificial Network“ and you will have the result within a few seconds (the red curve will show you the result for the ann in relation to the optimal as the yellow curve within the uncertainty band
10. result: you can now activate ann which will be confirmed with a window separately

If your activation is successful you will see a success box showing up.

You can also see previous results. To do this, click on the button previous trainings, where parts of the settings are displayed as well. You can also change the activated ANN to a previous ANN in this interface.

Note

Best results can be achieved, if both, the global error and the independent error are low. You can also carry out a verification of your ANN training to compare it to the raw forecast variables min, max, percentiles and optimal (f1).

Note

It takes approximately 5-10min before the ANN is visible in the graphics display and for verification purposes after activation.

Related questions:

. Can I see the verification results for given curves and a new weighted curve ?

15.3. How can I see the ANN curve within my graph?

If the ANN training is done and the activation is confirmed you only have to mark the red ann curve in the legend (Forecast Data) then you can see the red ANN curve within the graph.

What is the difference between the f1-f4 and u1-u4 curves and how can I activate them?