Biogas Plant AI: Predict Yield and Digester Instability

Can AI predict biogas yield and warn before a digester turns sour?

Yes: a model on SCADA and FOS/TAC data forecasts gas yield per recipe and flags an acidifying digester days before methane drops. Dutch green-gas plants fed 294 million m3 into the grid in 2024.

Yes, from the data your plant already logs: feedstock intake and recipe, digester temperature, pH, FOS/TAC from the lab, gas flow and methane content from SCADA, and the status of the CHP or the upgrader. A model learns what your digesters look like when they are healthy, forecasts tomorrow's gas and flags the drift that precedes an acidification days before the methane drops. The plant operator keeps every decision. This guide says what changes, what data it needs, where it fails and what it costs.

By Tom Joseph · Last updated: 14 October 2026

What a yield and stability model does that the weekly lab round cannot

Forecasts gas per recipe. From the weighbridge and intake register (tonnes per substrate, dry matter, organic loading rate), digester temperature and hydraulic retention time it predicts the biogas flow and methane content of the next 24 to 72 hours, so the shift knows whether tomorrow's recipe fills the upgrader or starves it.

Warns before a digester acidifies. It learns the plant's own steady state from pH, FOS/TAC, gas flow, CH4 and H2S in the historian and flags relative drift, not a fixed threshold. In a published lab study on corn stalk digestion the intermediate-to-bicarbonate alkalinity ratio gave a warning time of 7 days and volatile fatty acids 5 to 8 days before failure (Wu et al., Biotechnology for Biofuels, 2019); a plant model is calibrated on your digesters, and your lead time is measured in the audit.

Predicts the trip, not just the chemistry. Membrane differential pressure, water-scrubber load, CHP derating and methane slip are patterns too: the model shows which upgrader or CHP fault is building and what it costs in lost feed-in hours at the grid entry.

Closes the production report. Metered Nm3 at the grid entry, the SDE++ production statement and the guarantees of origin at VertiCer (formerly Vertogas) are reconciled against the forecast, so a gap between what the digesters should have made and what was fed in is explained with a cause, not discovered at the annual settlement.

Data source, what it yields and who decides

Source	What the model gets from it	Who decides
SCADA historian (WinCC, Ignition or your integrator's package): temperature, pH, gas flow, CH4, H2S, pressure, stirrer load	Minute-level steady state per digester; drift and yield forecast	Operator confirms or dismisses a warning on the shift log
Lab log: FOS/TAC, VFA, NH4-N, dry and organic matter	Anchor points that calibrate the online signals, two or three per week	Process technologist sets the recipe and the loading rate
Weighbridge and intake register: substrate, tonnes, supplier, recipe	Yield per substrate and overload risk per recipe	Feedstock planner decides what goes in tomorrow
Upgrader and CHP PLC: membrane dP, scrubber load, engine load, methane slip	Fault building up, expected lost hours	Maintenance plans the stop with the CMMS work order
Grid entry meter: Wobbe index, H2S, O2, pressure, delivered Nm3	Off-spec and shut-in risk, feed-in hours against the 8,000-hour target	Plant manager settles SDE++ and guarantees of origin

The data and systems a digester model runs on

Nothing new has to be bolted onto the tank. The model reads the SCADA historian that already exists: Siemens WinCC runs the process visualisation at Gasum's biogas plant in Götene and Ignition runs the digester-gas upgrading plant at La Farfana, and most Dutch digesters run one of those or the integrator's own package on an S7 PLC. It needs twelve months of minute data from that historian so that it has seen a summer, a winter, at least one recipe change and preferably one upset. It needs the lab log, whether that lives in a LIMS or in the biologist's spreadsheet, because FOS/TAC and VFA are the only direct measurements of the chemistry and they arrive two or three times a week. It needs the intake register from the weighbridge with the recipe per day, because a yield forecast without the feed is a guess. And it needs the upgrader and CHP logs and the grid-entry meter, because the money is counted in Nm3 fed in, not in m3 of biogas produced.

Two constraints are fixed. The sensors sit in ATEX zones and the model never asks for new ones: it works with what the historian holds. And it stays outside the safety layer: it writes a warning to the shift log and a forecast to the planner's screen, never a setpoint to the PLC, so it needs no change to the HAZOP and no new notification under the Omgevingswet. The permit regime for the digester itself does not move: a mono-manure digester is notified four weeks in advance under paragraph 4.88 of the Besluit activiteiten leefomgeving, co-digestion needs an environmental permit, and a model that reads data changes none of that.

What a yield and stability model does not fix

It cannot see what nobody measures. Trace-element depletion, antibiotics in a manure batch or a foaming digester do not show in the historian until the gas flow does, and the model can only say that the pattern looks like the week before the last upset. A pH probe that has drifted gives a model that drifts with it; sensor maintenance is the first thing the audit checks. A substrate the plant has never fed before, or a recipe change the planner did not record, produces a forecast that is confidently wrong, which is why the planner sees the inputs the forecast rests on.

It does not replace the process technologist, and it does not shorten the lab round: it makes the two or three FOS/TAC tests a week worth more by filling in the days between them. It does not fix the grid. ACM has described how a producer on a regional distribution network is automatically shut in when the pressure rises in summer; a forecast tells you when that will cost you the most gas, it does not add capacity, and the Netbeheer Nederland commitment of at least 8,000 feed-in hours a year is theirs to deliver, not the model's. And it does not replace the biogas plant's SCADA, its CHP service contract or its ICT partner: Crux Digits builds the model and the integration, and it works alongside the integrator who owns the control system.

What it is worth at a plant with three digesters, and what it costs

Take a green-gas plant with three digesters, 1,500 Nm3/h of biogas and a membrane upgrader delivering about 900 Nm3/h of green gas to the regional grid, aiming at 8,000 feed-in hours a year. Assume two instability episodes a year that each hold one digester at 60% load for ten days, and four unplanned upgrader stops of twelve hours. That is roughly 2 x 240 x 0.4 x 900 plus 4 x 12 x 900, about 216,000 Nm3 of green gas a year not delivered. Assume an all-in value of €0.80 per Nm3 across gas, SDE++ and guarantees of origin; every number here is an illustration and the audit replaces it with your meter data. The lost output is then worth about €173,000 a year. If an early warning turns half of those episodes into a recipe adjustment instead of a recovery, and maintenance plans two of the four stops, the plant keeps roughly €86,000 a year. That is the number to test, on your own digesters, before anything is built.

Three fixed steps, all excl. VAT. An AI audit and strategy of €2,500, one to two weeks on twelve months of historian, lab and intake data: you keep a report that shows how many hours each digester spent outside its steady state, what each episode cost in Nm3 and what warning time your data supports; the fee is credited in full if you continue within 60 days. A Production-ready MVP of €20,000, four to six weeks: the model runs on one digester and the upgrader next to the current lab round, with its warnings on the shift log, and you keep the code and the model whether you continue or not. Production from €50,000, eight to twelve weeks: all digesters, integrated with the historian, the lab log and the production report, with the operator approving every action. See machine learning, data engineering, AI for energy, AI implementation in the Netherlands and, for the shut-in question, grid congestion for industry.

Frequently asked questions

Does the model need new sensors in the digester?

No. It works on the SCADA historian, the lab log and the intake register you already have. The sensors sit in ATEX zones and adding any is a project of its own; the audit first checks whether the existing pH, temperature and gas-analysis signals are maintained well enough to learn from.

How much data does it need?

Twelve months of minute-level historian data per digester, the lab results for the same period and the daily recipe from the weighbridge. The year matters because it must have seen a summer, a winter and at least one recipe change; an upset in the data is useful, not a problem.

Does it work with our SCADA?

Yes. Siemens WinCC, Ignition and the integrator's own packages all keep a historian the model can read, and the warning is written back to the shift log, not to the PLC. The integration path is scoped in the audit with your integrator, who keeps the control system.

Does the model adjust the feeding recipe itself?

No. It forecasts the gas per recipe and flags drift; the process technologist decides what goes in and the operator confirms or dismisses each warning. Nothing writes a setpoint, so the HAZOP, the ATEX zoning and the permit under the Omgevingswet stay as they are.

What does yield and stability prediction cost a green-gas plant?

At Crux Digits: an AI audit and strategy of €2,500, a Production-ready MVP of €20,000 on one digester and the upgrader, production from €50,000 for the whole plant, all excl. VAT, with the code and the model handed over. The audit is credited in full if you continue within 60 days.