Home / Science

Science & validation

Forecasts you can check.

EcoFlow Predict rests on a simple physical idea: the signals sensors already record carry information about the pollutants they don’t. Here is how we build on it, and how we prove it works at each site.

The principle

Why proxy signals carry pollutant information.

Pollutants change the physics and chemistry of water in measurable ways, and they travel with the flows that sensors already track.

  • Conductivity rises with dissolved ions such as nitrate, and shifts when groundwater, runoff or effluent change the mix of water in the river.
  • Turbidity tracks suspended particles, which carry particle-bound phosphorus and metals after rainfall and runoff.
  • Dissolved oxygen and temperature reflect biological activity and nutrient enrichment through the day and the seasons.
  • Ammonium, flow and rainfall signal effluent, agricultural runoff and dilution.

The evidence

What published research already shows.

At three river sites in the US national ecological monitoring network, models using only in-situ sensor variables (conductance, dissolved oxygen, temperature, turbidity and water level) explained 75–85% of the variation in nitrate, rising to 99% when time-series structure was included (Kermorvant et al., PLOS ONE, 2023).

Inferring bacterial indicators and organic load from low-cost sensors is already in commercial use in the UK. What has not been shown is inference of several pollutant classes from the same sensors, transfer to catchments with little laboratory data, and forecasts that follow pollution downstream. That is the focus of our R&D.

Validation

Every site is calibrated and scored before it goes live.

Each deployment is tested against laboratory reference samples. Clients receive a written calibration report with the metrics below before the site moves from supervised to live operation.

R² ≥ 0.6

Target agreement between nitrate estimates and laboratory results at pilot sites

≥ 12 h

Target warning lead time for confirmed pollution events

≤ 1 / month

Target false high-risk alerts per catchment

≥ 95%

Target share of hourly forecasts produced despite sensor gaps

These are the acceptance targets for our pilot programme. Actual results for each site are reported to the client, including where a target is not met.

Research programme

The open questions we are working on.

Now – month 12

Several pollutants from one set of sensors

Estimating nitrate, phosphate and dissolved metals together, using physically informed features rather than a separate model for each.

Months 6–24

Transfer to data-poor catchments

Using catchment characteristics and targeted sampling so new sites reach useful accuracy within weeks, not years.

Months 12–30

Catchment-to-coast forecasting

Combining hydrological routing with machine learning to forecast pollution arriving at towns, bathing waters and estuaries.

Ongoing

Robustness to real-world sensors

Detecting drift and fouling automatically and reporting uncertainty with every estimate.

Validation study under way. We are testing the approach on the Environment Agency’s open Water Quality Archive for English rivers, on sites the models have never seen. We will publish a summary of the results here.

References

  1. Kermorvant, C. et al. (2023). Understanding links between water-quality variables and nitrate concentration in freshwater streams using high frequency sensor data. PLOS ONE 18(6): e0287640. Link
  2. Environment Agency (2026). New data to drive action to improve England’s rivers and lakes (2025 classifications). Link
  3. Environment Act 2021, Part 5, sections 80–82. Link

Have a river, catchment or discharge point you need to understand better?

Join the pilot programmeHow we validate