Desalination plant performance monitoring that predicts membrane fouling and staggers every clean against the city's demand
Clean membranes when they need it, not by the calendar.
See one case, screen by screen ↓A seawater RO plant's historian logs feed, concentrate and permeate pressure, flow, conductivity and temperature for every train, every minute. To see real fouling, a process engineer pastes those readings into a normalisation spreadsheet — the ASTM D4516 method, with its temperature correction factor and net driving pressure — once a day or once a week. Most plants still clean on a fixed calendar and look at the cleaning trigger only when something is already wrong.
So the calendar is wrong both ways. At Corran Point, Train 7 was cleaned at +2.4 % pressure drop, when it did not need it yet; Train 3 will cross its trigger 14 days before its calendar date. Cleans also collide: on the calendar, Train 3 and Train 5 both land on 22 October, the hottest day of a heatwave, and two trains out leaves 162,500 m³ against a demand of 172,400 m³. Membranes are replaced by age, not by how well they recover after each clean.
Estimated split for one engineer looking after eight RO trains, in hours a week, by hand and with the solution.
Six specialist agents normalise, forecast, diagnose and stagger every clean; the process engineer books and the manager signs.
Corran Point Desalination Plant runs 8 seawater RO trains of 25,000 m³/day each for Lumina Utilities. It is Tuesday 13 October 2026; the sea is at 33.4 °C, an algal bloom reached the intake on 2 October, and demand rises to 172,400 m³ on 22 October in a heatwave. Here is what happens next, screen by screen, in the working solution.
Before the day shift, the data normaliser corrects 11,520 readings across the 8 trains to each train's post-clean baseline. Two conductivity readings on RO-06 look wrong; they are marked suspect and left out, not deleted. Yesterday's view still has Train 3 on Watch at +9.6 % differential pressure.
“8 trains · 11,520 readings normalised to their post-clean baselines; 2 suspect conductivity readings on RO-06 marked, not used”
Farah opens the plant schematic: intake, flotation, filter banks A and B, cartridges and the 8 trains, each vessel grid coloured by drift. Train 3 is at +11.8 % pressure drop since its 29 August clean, rising 0.36 points a day, twice its August rate. The fouling forecaster puts its trigger on 22 October, in 9 days (band 20–24 October) — while its calendar clean is 5 November. The output band under the trains shows one day below demand.
“Thu 22 Oct: Trains 5 and 3 out → 162,500 m³ against a demand of 172,400 m³”
The clean scheduler reads the demand forecast and tries the options in order. Moving Train 5 earlier is rejected: it is at +8.6 %, and cleaning it 7 days before its forecast wastes a clean. Moving Train 3 to Tuesday 20 October works — forecast pressure drop on the day +14.1 %, still under the trigger, with 181,250 m³ available against 168,900 m³ (+7.3 %). Trains 8, 2 and 1 are matched to their own forecasts, and the CIP skid and crew are free.
“No day below demand · lowest margin +5.1 % on 22 Oct · 4 bookings changed → Daniel Okoye approves”
In Train 3, the rise sits in rows E–H, fed first from filter bank B, and in the lead elements — where a biofilm starts. The diagnostician ranks biofouling at 78 %, particulate at 15 %, scaling at 4 % and organic at 3 %, each with the evidence for and against. Chlorophyll-a is 6.8 µg/L and TOC is up since 2 October; against it, salt passage is rising slower than a pure biofilm usually gives.
Against scaling: “Tail elements normal; antiscalant dosing in range; LSI −0.4 in the concentrate”
August's clean, CIP-2026-037, was acid only and recovered 91 % of baseline flow; earlier cleans recovered 96–98 %. The recommended clean is alkaline Coral Chem CL-12 at pH 12 and 35 °C for 6 hours, a non-oxidising biocide, then citric acid at pH 2, then flush and a new baseline at 24 hours. Cleaning on 20 October is expected to recover 97 %; waiting for the calendar date, 90 %.
“An acid clean dissolves scale; it does not lift a biofilm.” The diagnostician recommends — Farah chooses.
Farah books the clean-in-place for Tue 20 Oct 2026 at 06:00: 18 hours offline, back in service at midnight, with day crew B and Coral Chem skid 2. Her signature records its meaning, and Tomas Varga takes the train out of service on the day. Priya Raman is notified to have the crew and chemicals ready.
“Signed by Farah Nasser, process engineer. Meaning: “Book CIP for Train 3”.”
Cleaning plan v2 runs from 13 October to 11 November: 4 bookings changed and 5 cleans down to 4 in the 30 days. Train 1 moves from 10 to 17 November and Train 2 from 29 October to 3 November, because the calendar was early. Train 8 comes forward from 2 November to 31 October, because its trigger falls before the calendar date. Train 5 stays on 22 October.
Train 3: “Needs 22 Oct, which already has Train 5 in a heatwave; 20 Oct is under the trigger (+14.1 %)”
Farah sends the plan to Daniel with each change and its reason. The lowest margin in the 30 days is +5.1 % on Thursday 22 October, 181,250 m³ against 172,400 m³, and there is no day below demand. Daniel approves at 09:40, signing “approve cleaning plan, 13 Oct–11 Nov”, and the crew request goes to Priya Raman.
A plan with any day under +5 % margin needs the water production manager.
Replacing by age would buy new membranes for Trains 1, 2 and 6 in 2027, $2.23M. Train 6 has to go in Q1 2027 — salt passage +31 % and recovery down to 89 % — at $742,560 for 952 elements. Trains 1 and 2 still recover 97–98 % after each clean with salt passage at +2–3 %, so the planner moves them to 2028: $1.49M deferred.
“This supports the case to defer; it is not an order.”
The same fouling train, seen by the five people who run the plant — what their month looked like, and what it looks like now.
Six specialist agents normalise, forecast, diagnose and stagger every clean; the process engineer books and the manager signs.
Every 15 minutes, normalises each train's permeate flow, differential pressure and salt passage to its post-clean baseline with the ASTM D4516 method — temperature correction factor and net driving pressure — and flags suspect sensors.
Fits each train's fouling rate since its last clean from 90 days of normalised data, weighting the last 10 days, and forecasts the day it crosses each cleaning trigger, with a confidence band.
Compares lead and tail elements, salt passage and pressure-drop patterns, feed-water lab results and the last cleans; ranks biofouling, particulate, scaling, organic fouling and ageing; and recommends the clean chemistry.
Places each clean before its forecast trigger and staggers the cleans against the demand forecast, the trains-in-CIP limit and crew availability, moving the train closest to its trigger by the fewest days and explaining each move.
Computes kWh/m³ for each train and the plant every day from the energy meters and permeate flow, and prices the energy above each train's clean baseline at the tariff.
Builds the five-year membrane replacement plan from recovery after each clean, salt passage and age, compares it with replace-by-age and drafts the capital request.
Books each clean; the manager approves the plan. The agents propose; a named person decides.
Ask about any train, clean, energy or membrane — or say what to change. Every answer points to its readings.
Why is Train 3 fouling faster than the others?
Biofouling, started by the bloom and helped by August's clean. Its pressure drop has risen 0.36 points a day since 2 Oct. Chlorophyll-a at the intake went from 1.1 to 6.8 µg/L and filters B are backwashing every 14 h; Train 3's rise sits in rows E–H, fed first from filters B. Its last clean was acid only and recovered 91 % — acid does not lift a biofilm.
Can we clean Train 3 and Train 5 in the same week?
Yes — not on the same day. Both on Thu 22 Oct would leave 162,500 m³ against a demand of 172,400 m³. Train 3 on Tue 20 Oct is still under its trigger (+14.1 % forecast) and leaves 181,250 m³ against 168,900 m³; Train 5 stays on 22 Oct.
What is fouling costing us in energy this month?
About $27,100 — 361 MWh above what the trains would use if they were all clean (12,040 kWh a day at $0.075). Train 3 is 39 % of it at 3.61 kWh/m³ against its 3.39 baseline. Cleaning it on 20 Oct instead of 5 Nov avoids about 72 MWh.
Which trains can wait for new membranes?
Trains 1 and 2 can wait to 2028. Replacing by age would buy them in 2027, but both still recover 97–98 % after each clean and their salt passage is only +2–3 %. Train 6 cannot: recovery has fallen 95 → 89 % over four cleans and salt passage is +31 %. That moves $1.49M out of 2027.
13 screens from the working solution, on its sample data. Pick one to see it large.
The cleaning triggers, the output the scheduler protects, the costs and who decides are settings, chosen by the plant.
| Setting | Default | Choose from |
|---|---|---|
| Permeate flow down | 10 % | Normalised flow; membrane maker guidance 10–15 % |
| Differential pressure up | 15 % | Normalised ΔP across the train |
| Salt passage up | 15 % | Normalised salt passage; Train 6 watched at +25 % |
| Minimum margin over demand | 2 % | Every day of the plan — what the scheduler protects |
| Manager approval below | 5 % | A plan with a day under this margin needs Daniel Okoye |
| Trains in CIP at once | 1 | Two allowed only below 150,000 m³/day demand |
| Train offline per clean | 18 h | Used to compute the output on a clean day |
| Energy tariff and element price | $0.075/kWh · $780 | Internal transfer price · Halden SW-440, 2026 contract |
Every figure is labelled: a target the solution is built to, an estimate, a typical published result, or a proven one.
“demo” = seen in the working solution, on its sample plant data · “estimated” = our arithmetic, inputs shown: $0.9M = 62.8M m³ a year × 3.42 kWh/m³ × $0.075 × 3 % less energy ($0.48M) + one third fewer of 52 cleans a year × $24,000 ($0.41M), for the demo plant; the hours-a-week split is also our estimate · “typical” = published figures: Water Online (2026) — energy is 45–60 % of RO operating cost, fouling cuts productivity 15–40 %; Mahmoud et al., Frontiers in Membrane Science and Technology (2026) — up to 12–15 unplanned shutdowns a year at some Gulf SWRO plants, and a literature-based Gulf model, not field-validated, of 3 % less energy and 50 % fewer cleans (we use a third). Normalisation method: ASTM D4516. Lumina Utilities, Corran Point Desalination Plant, Halden Membranes, Coral Chem, the people and the records named on this page are fictional — characters and sample data in the working solution.
Software that tracks how every reverse-osmosis train in a desalination plant is performing — normalised flow, pressure drop, salt passage and energy per cubic metre — and tells you when each needs cleaning. Desalination Plant Performance Monitoring adds six agents that normalise the data, forecast fouling, name its type, stagger the cleans against demand and plan membrane replacement. The process engineer books each clean and the water production manager signs the plan.
The data normaliser corrects every train's readings to its post-clean baseline every 15 minutes, using the ASTM D4516 method. The fouling forecaster fits each train's fouling rate since its last clean and forecasts the day it crosses each cleaning trigger, with a confidence band. In the sample plant it put Train 3's trigger on 22 October, band 20–24 October, 14 days before its calendar clean.
The fouling diagnostician reads where the pressure rise sits in the train, the feed-water lab results and the clean history, and ranks biofouling, particulate, scaling, organic fouling and ageing with the evidence for and against each. It recommends a sequence within the membrane maker's limits; the process engineer chooses.
That is what the clean scheduler is for. It places each clean before its trigger, never puts more trains in CIP than the plant allows, and keeps output above demand plus the minimum margin on every day. In the sample plan it moved Train 3 off Train 5's day, and the lowest margin in 30 days was +5.1 %.
It shows what fouling costs in energy, train by train, so each clean is timed by its forecast rather than the calendar. The energy tracker computes kWh/m³ for each train against its clean baseline every day; at the sample plant fouling cost $27,100 of energy this month, and cleaning Train 3 on 20 October instead of 5 November avoids about 72 MWh. Our estimate for a 200,000 m³/day plant is $0.9M a year in energy and cleaning.
Yes. The membrane planner builds a five-year plan from recovery after each clean, salt passage and age, and compares it with replace-by-age. In the sample plant it keeps Train 6 in Q1 2027 and supports the case to defer Trains 1 and 2 to 2028. It drafts the capital request and never orders membranes.
Operators stay in control. The agents read historian, lab and meter data and never write to the control system, set-points or dosing. Only the process engineer books a clean, with a signature; the water production manager signs any plan that changes bookings; the shift supervisor takes the train out of service.
The Agentic Solution Engine builds and deploys it from your requirements and documents — your historian tags, cleaning triggers, CIP records and membrane data — and it goes live once every quality gate has passed. We will walk you through it on your own plant data first.
We'll run Desalination Plant Performance Monitoring on a sample of your own historian data and clean records.