Understand prediction parameters in detail

Important parameters for the Forecast App

In the Forecast App in Eliona, three parameters determine the quality and expressiveness of your time series forecasts. Below you will find a compact description of the most important settings in each case:

---

1. Forecast length

Definition: Number of time steps that the model should forecast into the future.

• Calculation of the time distance: The app automatically determines the average time interval between your most recently saved data points in Eliona.

• Examples:

  • With 5-minute intervals and [forecast_length] = 3 the forecast corresponds to a horizon of 15 minutes.

  • With daily measurements and [forecast_length] = 5 five days are forecast.

Make sure the data intervals are regular. If data is stored irregularly, the time axis cannot be derived correctly. It is recommended to activate the Trending option: "always" in the attribute.

---

2. Context length

Definition: Number of past data points that the model uses as historical context.

• Short-term patterns: Small values (e.g. 24) are suitable for daily cycles or hourly trends.

• Long-term effects: Larger values (e.g. 168 or more) are useful when seasonal or weekly patterns are to be recognized.

• LSTM advantage: Thanks to the Long Short-Term Memory-architecture, the model also remembers relevant information beyond the selected window, which usually means a moderate window size is sufficient.

Longer context windows greatly increase the computational effort and memory requirements.

---

3. Feature Attributes

Definition: Feature Attributes are additional input variables that the model uses to calculate the forecast. They extend the context and thus improve the accuracy of the forecast.

Automatic time features:

If no other feature is selected, the Forecast App automatically adds the following time-based attributes:

• minute_of_hour

• hour_of_day

• day_of_week

• month_of_year

These fields appear automatically in the form when a forecast item is created and provide the model with cyclical information about the current time structure.

The actual model processing takes place internally via the transformed variants of these values (_sin/_cos), e.g.:

• hour_of_day_sin / hour_of_day_cos

• day_of_week_sin / day_of_week_cos

• month_of_year_sin / month_of_year_cos

This transformation ensures that cyclical time patterns (e.g. the transition from hour 23 to 0) are mathematically represented correctly.

Optional: asset-related features

You can add further attributes of the same asset, e.g.:

• Temperature

• Humidity

• Energy consumption

These help the model recognize relationships between different measured variables.

Best practices:

• Do not use meter readings directly. Instead, insert differenced or relative values.

• Create your own features: Use the Eliona Calculatorto generate combined or derived values (e.g. area calculation, growth rates).

• Ensure data quality: Use filters to remove outliers or smooth noisy signals.