What Explainable Nutrient Early Warning Should Show an Operator

Beyond a single predicted number

An early-warning tool that shows a single predicted ammonium value is giving an operator very little to work with. The number may be right or wrong, but on its own it does not say how far ahead it looks, how confident it is, why it has changed, or what should happen next.

Explainable early warning is less about a particular algorithm and more about what the interface makes visible. Waste Water Care organises this around six elements.

1. Forecast horizon

Operators need to know when, not only what. Showing short horizons side by side — for example +1h, +2h and +4h — lets an operator see whether a change is expected to build gradually or arrive suddenly, and how much time there may be to investigate.

2. Confidence and abstention

Every forward-looking view should communicate how much it should be trusted. In practice this means a visible uncertainty band and an explicit confidence state: normal, reduced, or unable to assess.

Abstention matters. If a key input has failed a data-quality check, the honest response is to say so rather than to show a precise-looking trajectory. In the current MVP, uncertainty is shown as a curated illustrative band; calibrated uncertainty requires validated model work on real plant data.

3. Driver context

Operators reason about cause. A view that indicates which signals are associated with a change — dissolved oxygen falling, influent flow rising, temperature dropping, aeration output declining — connects the forecast to things the operator can check.

Driver views must be labelled carefully. In the MVP they are illustrative driver analysis for sample scenarios, not attribution values from a validated model.

4. Alert

An alert turns a risk view into a unit of work. It should have an identifier, a source, a clear status and an owner. Without that, early warnings become background noise.

5. Operator action

The alert lifecycle should mirror how operators actually work: acknowledge, assign, investigate, record an action, close — and reopen if needed. Each step should leave an audit event with a timestamp and a note.

This is also where the boundary of the product is drawn. Waste Water Care records what the operator decided; it does not take operational action itself.

6. Evidence

Finally, everything should be traceable. Which data was used, which checks passed or failed, what the operator observed, what was done. That evidence feeds shift handovers, incident summaries and future validation work.

An evidence-grounded assistant can help, but only if it answers from an approved source set, shows its sources, and says plainly when the available evidence is insufficient.

Putting it together

Horizon, confidence, drivers, alert, action and evidence together form a workflow, not a widget. That workflow is what the Waste Water Care MVP frontend demonstrates on synthetic sample data, with every predictive element clearly marked as curated demo output.