Prediction tags enable you to generate a new tag based on a multi-variable linear regression analysis. This is the ideal feature to create a soft sensor by combining different influencing tags.
Note: The prediction tags tool has replaced the "influencing factor" analysis tool, found in older TrendMiner versions.
The prediction tags can be found in the tag builder menu.
The general overview of tag builder's tag workflows can be found here.
Note: Prediction tags deviate from the usual tag builder workflow, as this tool has its own build-in workflow, described below.
Prediction tag submenu
Some starting parameters need to be set before the prediction tag flow can be initiated.
Tag/attribute to predict: In this field you select the tag or attribute you want to use as the basis of your prediction analysis. A common practice is to first upload offline measurements such as lab quality measurements via the data import functionality.
The prediction chart: A chart will be loaded as soon as you have selected your tag of interest. The same period is chosen as selected in the focus chart. When creating a prediction tag, an isolated workflow is initiated with its own chart. This chart should not to be confused with the focus chart, and not all options (such as filtering,…) will be available.
The timespan chosen on the chart will be used for the prediction analysis, so be sure to zoom in or out on a specific period.
The prediction chart is persistent if the prediction tag flow is active. You can switch freely between this chart and the actual focus chart by switching menus in the sidebar. If you want to include a filter after starting a prediction flow, you first need to switch menus and go back to the tag builder menu.
Tags/attributes to search for: You can choose with the radio-button, which tags the analysis needs to be performed.
- Only the tag and attributes present in the active tag list.
- Choose this option if you already have an idea which other tags are good candidates to explain the behaviour.
- All indexed tags and attributes present in TrendMiner
- Choose this option if you don't have an idea which others could explain the behaviour.
- Tag filter expression: When performing an analysis on all indexed tags you can limit which tags TrendMiner will look through. E.g., Filling out "*temp*" will look through all indexed tags that contain "temp" in their name. Using the tag-filter results with faster computation time and more specific candidates to choose from.
Upstream shift (max): Other process parameters could be influencing your current tag at the same time or with a lag in time. TrendMiner can take this lag into account for its analysis using the automatic shift detection. The maximum upstream shift parameter sets out how far in the past, alignments of candidate factors are checked.
The maximum upstream shift that can be set is 2 years.
Note: The specific shifts for which the correlation analysis is performed depends on the specified maximum shift. The specified maximum shift is divided into 100 equidistant intervals and the analysis is performed for each interval. The smallest possible shift is equal to the index resolution.
You'll notice that your screen is horizontally split in two after starting the prediction flow.
The upper part is divided in four columns which you can follow from left to right to complete the analysis flow:
- Summary-column: An overview of the Prediction settings as entered in the previous menu.
- Candidates-column: Lists all indexed tags/attributes that influence the current behaviour as shown in the prediction chart. Each candidate shows a percentage approximation score and the optimal time shift resulting in the highest score.
The candidates are sorted based on their score, only the top 300 results are shown.
From here candidates can be selected and added to the "selected candidates" column by clicking the arrow. The scores of the remaining candidates will be recalculated after adding one candidate.
Note: A total of 3 to 10 candidates can be added depending on your system resources.
- Selected candidates: List all the candidates you have selected from the candidates-column.
Selected candidates can be removed again after which a recalculation of the candidates and result column is triggered.
- Prediction results: A linear combination between the selected candidate factors and tag of interest is shown here as a formula. This formula approximates the behaviour of our tag of interest based on the input of the selected candidates. The combined score is shown as the total accuracy as well.
From here you can save the approximated model as prediction tag. The model that comes out of the selected candidates could then be used as a soft-sensor for offline/less-frequent measurements or even predict what a certain value is going to be if the candidates were early indicators. When saving a prediction tag, the whole workflow is saved. This allows you to revisit the analysis if the model is no longer accurately predicting the real values.
The bottom part shows the prediction chart. The full plotted line shows the original situation of your tag of interest, while the dotted line plots an approximated model based on the selected candidates.
Tip: After saving the prediction tag it can be useful to validate the model by checking out other periods used in the prediction flow.
Percentual approximation score
The percentual approximation score indicates how much a specific candidate could explain the behaviour of the tag of interest as shown in the prediction chart. A score of 100% means that the selected candidate can accurately describe how the behaviour of the tag of interest would be.
In order to prevent memory issues, the prediction tag functionality is restricted to:
- 3 (up to 10) selected 'candidate'-factors (the number of candidates you can add is dependent on the resources of the system).
- A selected time period of 2 years in the Focus Chart, including the maximum shift.