In predictive mode, you can use TrendMiner to predict future evolution of batch runs, transitions, or continuous processes, based on previously observed historical behaviour.
As predictive mode is based on TrendMiner's search technology, it works in a fundamentally different way to model based predictive technologies. The advantages of which are:
- No extensive data preparation activities required.
- No need to update models. Predictive mode always has access to the latest data and incorporates plant changes automatically.
- Effects of adjusting tag set or history range are immediate (no model retraining).
- No difficult black box predictions to understand or accept. Predictive mode always links back to historical time periods that are used to make predictions.
To switch on predictive mode, click the predictive mode menu icon and toggle on the switch in the panel. When predictive mode is switched on:
- The view is updated so that the first 80% of the picture corresponds to the recent past, and the remaining 20% of the picture refers to the near future.
- Predictive mode automatically adds layers and updates them based on similarity with the actual, visible trends. The base period is indicated with a solid line. The added layers are shown in dotted lines.
- Predictive time periods are identified based on the data in the context time range, with exception of data that is filtered out. The number of layers and the minimally required similarity can be controlled by changing the settings (see below for details).
- The result is shown in the trend below.
To switch off predictive mode, toggle off the switch in the predictive mode menu. Currently, predictive mode is automatically switched off when the user selects a different menu.
The following settings are available:
- 'Number of best matches selected (1-5)': number of predictive signals that should be displayed. Note that fewer signals may be displayed if insufficient matches above the minimum score threshold can be found.
- 'Minimum score (5%-100%)': minimum score for a match to be meaningful. If no matches are found above the desired score threshold, no predictions will be shown. How the similarity scores are calculated and should be interpreted can be found in how should I interpret similarity scores.
- Tag selection: select a subset of tags that will be used to identify historically similar behavior that can be employed to predict. By default, all visible tags are selected.
To predict further evolution of a batch run, add the variables of interest to the batch run and zoom in on the first part of the current run, and then switch on predictive mode. With the default settings, TrendMiner will use all visible tags to identify the three batch runs that were historically most similar based on the current batch behavior, and display these runs as an indication of future behavior.
When the user then overlays the golden batch fingerprint, it is possible to quickly detect whether an ongoing batch is likely to evolve close to the golden batch profile, or whether a transition is likely to remain on track. In the example shown below, 2 out of 3 most similar batches show an end concentration below the golden batch target as we are approaching the end of the batch. Having this information early will trigger operators to closely monitor this batch and undertake the appropriate actions.