Neural Networks Quick Start

This page provides the quickest description for setting up the neural network predictions. We continue the guide for creating a new prediction model from the Predictions page:

1. Go to Settings.
2. Open Signal Sets settings.
3. Select a signal set for which to make predictions.
4. Click on the Predictions tab.
5. Create a new model using the Add model button and selecting the model class (ARIMA or Neural network).
6. Follow the instructions specific for the model class.

Model settings

The Model name must be set. It is used to identity the model and also to identity the generated signal sets with predictions. Note that the model name can't be changed later.

The Timestamp Signal should be set automatically.

Prediction parameters

These prediction parameters can be used to set the number of generated predictions, that is, how far into the future the predictions will be generated.

The Aggregation interval can be used to aggregate the records before using them for the predictions. We recommend doing that as it ensures that the difference between the time steps will be the same for all the data, even if the source data come into the system in irregular intervals. When the signals are aggregated, the minimum, maximum or average of the values in the aggregated buckets of records is used.

Future predictions specifies the number of generated time steps of the prediction. Use this in combination with the Aggregation interval to set the desired prediction horizon. For example, if one wants to predict one week ahead, one can set the Aggregation interval to 1d and Future predictions to 7 to create one prediction for each of the next seven days. Another option is to set the Aggregation interval to 1h and Future predictions to 168 to create a prediction for every hour of the next week, but the higher the Future predictions, the higher is the error of the predictions, so we recommend the former settings.

Observations is the number of past records used to compute the predictions. We recommend setting them to at least the number of Future predictions.

Overall, if we set the Aggregation interval, the signal set becomes a sequence of evenly spaced records. The Observations sets the number of records used to compute the predictions and Future predictions is the number of records to be generated.

Signals

Here, one can select the signals one wants to make predictions for – the Target signals. The Input signals are used as the inputs of the neural network used to compute the predictions. Often, the Input signals and the Target signals are the same, to facilitate the model setup, this is the default option. To select different input signals, uncheck the Same inputs as targets checkbox.

For each signal, settings can be specified. The Data type can be left with the default Automatic value. The Minimum and Maximum value can be specified if one knows some information about the signal, e.g. if the signal is a percentage, the minimum and maximum can be set to 0 and 100. The Minimum and Maximum value can be left empty to infer the information from the data. Lastly, the Aggregation can be specified if the Aggregation interval is specified. The aggregation is used to aggregate the records from the aggregation interval to produce a value used by the neural network.

Neural Network architecture

We recommend using the LSTM architecture (default). To learn more about the available architectures, see the Neural Networks Architectures.

Neural Network hyperparameters

For a quick start, the default parameters should be fine. More details about the available hyperparameters for each architecture can be found on the Neural Networks Architectures page.

Training parameters

The default settings should be reasonable for a quick start. More details about the settings can be found on the Add Model page.

Save and start training

Keep all the checkboxes checked and click on the Save and leave button to start the training. This will redirect you to the Model overview page where the progress and the results of the training are shown.

When the training completes, the predictions should be available in the created signal sets (which contain the name of the model in their name, so they can be easily found).