This page talks about short- to long-term trend types for the forecasted data Working_with_trend_predictions - Help
Is the length of those fixed?
Hi Bruno,
The lengths of the predictions shown in trend graphs are indeed fixed. On each trend data aggregation level, we fixed the forecasting horizon to be at most 200 predicted points into the future. Also, we do not allow predictions to be longer than 10% of the length of the data history which was fed to the prediction model training.
Reason for having a maximum length is that the further ahead, i.e. the more points ahead, we forecast, the more prediction errors accumulate and the more uncertainty is typically associated with the forecast.
- The most detailed prediction is generated by a prediction model built on the history of real-time trending data. Typically, one day of real-time data is stored in the database. The forecasting horizon will depend on the (polling) frequency of this real-time data. If the time step between two real-time values is for instance 30secs, the forecast length on the high precision prediction will be 200 times this time step, so 100 minutes into the future.
To have a longer forecasting horizon, it makes more sense to look at predictions generated on higher aggregation levels which have longer data history and where more long-term trends or patterns can be captured by the modelling:
- The short-term prediction is generated by a prediction model trained on the history of 5-minute averaged trend data. The forecasting horizon is limited to about 16hours (=200*5min).
- The mid-term prediction has a time step of one hour between each predicted point, as it is generated by a model trained on the history of hourly averaged data. The horizon for this prediction level goes up to about 8 days.
- When a sufficiently long history of hourly data is available, a long-term prediction is created where the time step between consecutive predicted points is equal to one day. If one year of hourly averaged trend data is available for the predicted parameter, the forecasting horizon will be a bit more than one month (+-10% of one year). If the history is shorter and no year of data is available yet, the prediction will also be shorter.