Prediction Models Interface

Prediction models in the left menu is where you create, open, and manage prediction models. One model predicts one telemetry key for the items of one device or asset profile, and keeps its forecasts, events, and alarms up to date.

Prediction models opens on two tabs. The Summary tab shows the state of every model at a glance.

#	Area	Contents
1	Create model	Opens the wizard, described in Getting Started
2	Summary cards	Refresh, accuracy, intents, and alarms across all models
3	Profile coverage	Every profile with the models built on it
4	System pulse	The latest training and refresh runs

The Models tab lists every model. Click a model to see its details beside the table.

#	Area	Contents
1	Search and profile filter	Find a model by name, or show the models of one profile
2	Intent chips	Show the models of one intent
3	Create model	Opens the wizard
4	Models	Every model with its status, intent, and accuracy. The row menu renames or deletes a model
5	Detail panel	The selected model at a glance, with Open model and Refresh now

The model page shows one model. Its tabs show the results and settings, and its actions work on the whole model.

#	Area	Contents
1	Tabs	Result, Properties, Events, Accuracy metrics, and Tasks, each showing one side of the model
2	Actions	Copy and Monitoring, which work on the whole model

The Summary tab opens by default. It shows every prediction model of the tenant: how they refresh, how accurate they are, which profiles they cover, and how their latest runs went.

Summary cards. Four cards at the top show the main figures across all models.

#	Card	Shows
1	Auto refresh	How many models refresh on a schedule. Below: Total created, Ready (trained for every item), Partially failed, and Failed
2	Avg accuracy	The average accuracy of the trained models. Below: Best model, Degraded (≥5%↓), the models that lost 5% or more against the previous 30 days, and vs last month
3	Models by intent	How many models there are of each intent
4	Alarms triggered	The alarms raised by prediction models, by severity

Profile coverage. Every device and asset profile, with the models built on it.

Column	Description
Profile	The profile name and type
Models	How many models are built on the profile
Monitored items	Items in a model, out of all items of the profile
Intent distribution	Models per intent: Forecast / Period Total / Time to Threshold / Hit Target
Auto-refresh	On when at least one model refreshes on a schedule
Save telemetry	On when at least one model writes its predictions to ThingsBoard
Alarms	Alarms raised by the models of the profile

+ Create starts the wizard with this profile already selected. A profile with models shows View > instead, which opens the Models tab filtered to that profile.

All models → in the table header opens the Models tab with no filter.

System pulse. The latest runs of all models, newest first, so you can see whether training and refreshes succeed.

Each entry shows the model, how long ago it ran, the kind of run, and its item counts:

Tag	Run	Counts
Train	Training	Items trained and failed
Refresh	A refresh	Items refreshed, skipped, and failed
Accuracy	An accuracy check	Items checked and failed

The dot shows how the run ended: Completed, Partial, Failed, or Canceled.

The Models tab lists every model of the tenant, newest first.

Each row shows:

Column	Description
Name	The model name and its profile
Created	When the model was created
Status	Ready, Training, Partially failed, Failed, or Cancelled
Intent	Forecast, Period Total, Time to Threshold, or Hit Target
Horizon	How far ahead the model predicts, for example Next 7 days or This Month
Auto-Refresh	Whether the model refreshes on a schedule
Saving Telemetry	Whether the model writes its predictions to ThingsBoard
Accuracy	The accuracy of the model
Entities	Items in the model, out of all items of the profile

By profile. The left select switches between Device profiles and Asset profiles, the right one picks a single profile.

By intent. The chips under the toolbar show the models of one intent, or All.

Search matches the model name as you type, and combines with both filters.

Rename from the panel. Click the pencil beside the model name in the detail panel. The name becomes editable: type the new name and press Enter to save, or Escape to cancel.

Rename or delete from the table. Click the ⋯ at the end of a row:

• Rename makes the name editable in the table. Type the new name and press Enter.
• Delete removes the model after a confirmation. The dialog names the business entities that use it.

Click a model row to see it beside the table, without opening it.

Field	Description
Status and Intent	As in the table
Intent figure	Prediction horizon for Forecast, Period for Period Total, Current / Upcoming events for Time to Threshold, On track / Off track for Hit Target
Trained / Failed items	Items the model trained for, and items it failed on
Accuracy	The accuracy of the model
In model / In profile	Items in the model, out of all items of the profile
Auto-refresh	Whether the model refreshes on a schedule, how often, and for how many items
Alarms	Whether the model raises alarms. Time to Threshold and Hit Target only
Save telemetry	Whether the model writes its predictions to ThingsBoard
Profile, Predicted field, Prediction horizon	What the model predicts, and how far ahead
Rules or Targets	The number of event rules or targets. Time to Threshold and Hit Target only
Last refresh and Created	When the model last refreshed, and when it was created

The panel has two actions.

Open model opens the model page on its Result tab.

Refresh now refreshes the model at once, without waiting for its schedule. Last refresh reads Refreshing… until the run ends.

The model page has up to five tabs. Click a tab name to switch to it.

• Result shows what the model predicts, in a layout that depends on its intent.
• Properties shows how the model is set up, and lets you change it.
• Events lists the predicted events. Time to Threshold models only.
• Accuracy metrics shows how close past forecasts came to what really happened.
• Tasks shows the training and refresh runs of the model.

The first four are described in detail on their own pages, linked below. Tasks is described here.

The predictions of every item. The chart shows the selected item, and the table lists all items.

Its layout depends on the intent, and each one is described on its intent page:

Intent	Result tab shows	Described in
Forecast	The predicted values over the horizon	Forecast
Period Total	The predicted total of the current period	Period Total
Time to Threshold	The time left until the value crosses a limit	Time to Threshold
Hit Target	Whether each item reaches its goal by the end of the period	Hit Target

Two parts are the same for every intent: the item actions, and the Data status card.

The ⋯ at the end of an item row opens the actions for that one item.

Action	What it does
Refresh	Learns from the readings that arrived since the last refresh, and predicts again
Retrain	Trains the item again from its whole training period, for example after its behavior changed or its training failed
Delete	Takes the item out of the model. The device or asset in ThingsBoard stays as it is

The selected item and how long ago it was refreshed, with buttons that run the same work for more items at once:

• Refresh item refreshes the selected item.
• Refresh all refreshes every item of the model.
• Retrain all failed trains again every item whose training failed.

Every setting of the model: the data it learns from, how far ahead it predicts, and how it learns.

See Properties tab.

Every predicted event of a Time to Threshold model, across all items and rules. Click an event to see it on the chart.

See Time to Threshold.

How close past forecasts came to what really happened, for the whole model and for each item.

See Accuracy.

Trendz runs the work of a model as tasks: one that trains it, and one that refreshes it on a schedule while periodic refresh is on. The Tasks tab lists them, so you can check when each run happened and whether it finished.

Column	Description
Name	The task and the model, for example Prediction model, refresh: Time to Threshold: Backup Generators (fuelLevel)
Created	When the task was created
Updated	When the task last changed
Status	RUNNING while a run is in progress, NOT ACTIVE between runs
Task type	Run Once for training, Scheduled for the periodic refresh
Job Type	Train prediction model or Save prediction model result to TB

The ⋯ menu at the end of a row holds two actions:

• Details opens the runs of the task: when each started and finished, how long it took, and its status. Clicking the row does the same.
• Go to Task opens the task on the global Tasks page, in a new browser tab. The page is described in Tasks service.

The top right of the model page holds the actions for the whole model.

#	Action	What it does
1	Copy	Creates a copy of the model
2	Monitoring	Turns on refreshing on a schedule, retraining, saving to ThingsBoard, and alarms

Create a copy of the model, for example to try another setting without touching the model in use.

1. Click Copy in the page header.
2. The copy opens in a new browser tab, named with (copy) at the end.

Set how the model keeps working on its own: refreshing on a schedule, retraining, saving predictions to ThingsBoard, and raising alarms. Click Monitoring in the page header, switch on what you need, and click Save changes. Each switch is described in Monitoring & Alarms.

• Forecast

  Predict the next values of a metric, like the energy of the next 7 days.

• Time to Threshold

  Predict when a metric crosses a limit, and raise an alarm before it happens.

• Period Total

  Predict the total, peak, or average of a period, such as the energy of this month.

• Hit Target

  Predict whether a goal is reached by the end of a period, and how far off it is.

• Models & accuracy

  How a model learns, its settings, and how accurate its forecasts are.

• Monitoring & alarms

  Refresh and retrain on a schedule, save forecasts to ThingsBoard, and raise alarms.

• Show on dashboards

  Chart history and forecast together in Trendz views and ThingsBoard dashboards.