Menu
All evidence

Fouling Forecast

Can existing measurements give useful warning before an exchanger’s duty becomes unacceptable? We compared seven forecasting methods on complete campaigns the models had never seen.

Simulated data. No real plant results. Open the interactive report

What we found

On campaigns shifted away from the training conditions, the simple trend was the most accurate at timing, with 1.38 days of error. Our grey hybrid came in at 2.25 days.

A white hybrid had the lowest temperature error. No single method won on every measure, so a blanket claim for hybrid modelling isn’t supported.

The grey model covered 100% of shifted cases, but its band averaged 44.65 K. High coverage from a band that wide isn’t useful for planning a cleaning.

Simulated data

Error in predicting when a fouling limit is reached

Mean timing error in days on 24 shifted synthetic test campaigns. Lower is better.

  1. Simple trend 1.38 Simple trend Timing error 1.38 days Temperature error 1.70 K Mean band width 8.86 K
  2. Grey hybrid 2.25 Grey hybrid Timing error 2.25 days Temperature error 1.72 K Mean band width 44.65 K
  3. White hybrid 3.17 White hybrid Timing error 3.17 days Temperature error 1.53 K Mean band width 8.26 K
  4. Black hybrid 3.23 Black hybrid Timing error 3.23 days Temperature error 2.05 K Mean band width 7.99 K
  5. Fixed rate 5.85 Fixed rate Timing error 5.85 days Temperature error 2.59 K Mean band width 10.72 K
  6. Direct ML 15.15 Direct ML Timing error 15.15 days Temperature error 5.87 K Mean band width 9.78 K
View as table
Method Timing error (days) Temperature RMSE (K) Mean band width (K) Band coverage
Simple trend 1.38 1.70 8.86 96%
Grey hybrid 2.25 1.72 44.65 100%
White hybrid 3.17 1.53 8.26 98%
Black hybrid 3.23 2.05 7.99 78%
Fixed rate 5.85 2.59 10.72 80%
Direct ML 15.15 5.87 9.78 19%

5 model seeds on one synthetic fleet. Persistence is left out because it never predicted a crossing, so it has no timing error.

How it was tested

A two-stream countercurrent thermal model generated 88 synthetic campaigns with disturbances, sensor noise and cleaning residuals. Forecasts used only data available on the forecast day.

Whole campaigns were held out: 24 training candidates, 20 for calibration, 20 nominal tests and 24 shifted tests. Five seeds varied the models on one fixed fleet, so they are not independent plants.

What it doesn’t show

This is a lumped single-phase model. It doesn’t represent a full refinery train, and it doesn’t establish savings, reliability or readiness for use on a plant.

It does show the method we’d apply to your historian data: simple baselines first, unseen test campaigns, and failures in plain view.

Run this test on your exchangers.

See the pilot