Why Data Quality Comes Before AI in Wastewater Forecasting

A forecast is only as reliable as its inputs

Discussions about artificial intelligence in wastewater treatment often start with the model: which architecture, how many layers, which horizon. In practice, the more decisive question comes earlier. Before any model output reaches an operator, someone has to establish whether the underlying plant data is fit to interpret at all.

Biological treatment plants generate continuous streams of measurements from analyzers, probes, flow meters and blowers. These instruments drift, foul, lose communication, get recalibrated and are occasionally replaced with units that report in a different scale. A model trained or run on that data inherits every one of those issues, and it usually does so silently.

That is why Waste Water Care places data-quality assessment as a dedicated step in the operator workflow, directly after plant context and before any nutrient-risk view.

What should be checked

The checks that matter most are not exotic. They are deterministic rules that any operations engineer would recognise. Making them explicit and visible is the point.

• Missing timestamps — gaps in the expected sampling sequence.
• Duplicate records — the same timestamp reported more than once.
• Missing values — readings that are absent even though the timestamp exists.
• Invalid ranges — values outside a physically plausible range for that signal.
• Sudden jumps — step changes larger than the process can plausibly produce between samples.
• Flatlines — a dynamic signal such as dissolved oxygen reporting an identical value for an extended period.
• Stale data — the most recent value for a signal is older than expected.
• Unit mismatch — values that only make sense in a different unit, such as a temperature reported in Kelvin.
• Coverage — the share of expected values actually present over the window.

Why these checks should be rules, not AI

It is tempting to describe every automated function in a software product as AI. For data quality that would be a mistake. Operators and engineers need to know exactly why a reading has been flagged, and they need the same answer every time the same data is evaluated.

Deterministic rules with explicit thresholds provide that. A flatline is defined, for example, as four or more identical consecutive readings on a dynamic signal. A coverage below a stated percentage moves the status from Pass to Review or Fail. Anyone reviewing the result can reproduce it.

In the Waste Water Care interface these checks are labelled plainly as rule-based. The product does not present them as machine learning, because they are not.

Pass, Review, Fail — and what that means downstream

A three-state result is easier to act on than a single score. Pass means the check found nothing that should reduce confidence. Review means something should be looked at, but the data may still be usable with caution. Fail means the issue is significant enough that any downstream risk view should reduce its confidence or abstain.

Importantly, a value should never be marked Pass simply because it exists. A reading of 2,400 m³/h on a line with a plausible maximum of 2,000 m³/h is present, but it is not valid. A dissolved-oxygen probe that has reported exactly the same number for nearly an hour is present, but it is probably not measuring.

Connecting data quality to the risk view

The value of a data-quality layer is realised only if it changes what the operator sees next. In the Waste Water Care workflow, a Fail on a signal that feeds the nutrient-risk view lowers the stated confidence of that view or causes it to abstain entirely, with an explanation.

This is a design choice: it is better for a decision-support tool to say "unable to assess" than to show a confident-looking trajectory based on a broken probe.

Takeaway

Data quality is not a preliminary chore to be completed before the interesting work begins. It is part of the product. Any wastewater forecasting effort that cannot show, signal by signal, whether its inputs are trustworthy is asking operators to accept outputs on faith.

The current Waste Water Care MVP demonstrates these checks on synthetic sample data. Applying them to a real plant's history is part of the next validation stage.