A Waste-to-Energy Digital Twin Can Only Optimize Inside Your Last Stack Test

"If we put a digital twin on this plant, can we run the boiler harder than we run it now?" That question opens almost every scoping call I take about plant simulation, and it usually comes from an operations director with a capacity problem and a board asking about throughput. The straight answer is no, not on the strength of a model. A digital twin of a waste plant will find you an optimum. Whether you're permitted to operate at that optimum was settled by a stack test somebody ran at one load, for four consecutive hours, quite possibly before the current control system was installed.

I'm not arguing against simulation. I build these systems, and the modeling earns its keep. But it earns it inside a box drawn by a document the data team usually hasn't read, and a WTE digital twin scoped without that document in the room will spend its first year generating recommendations nobody can legally act on. There are 57 US power plants burning municipal waste for electricity, per the EIA as of May 2026, most of them permitted long before anyone used the word twin in a vendor pitch.

The permit drew the box before the model existed

For a large municipal waste combustor in the United States, the monitoring section at 40 CFR 60.58b does something that surprises engineers arriving from other process industries. It defines the plant's allowable operating envelope by reference to a past test rather than to a design document. Maximum demonstrated unit load is the highest four-hour arithmetic average load the unit reached during four consecutive hours of its most recent compliant dioxin and furan test. Maximum demonstrated control device temperature works the same way: the highest four-hour average flue gas temperature at the inlet to the particulate control, measured across those same four hours.

So the envelope is historical, and it belongs to a particular afternoon with a contractor's sampling trains on the stack rather than to the equipment itself.

Compressed to one line, 60.58b says this: your legal maximum was set by whatever the equipment happened to be doing the last time somebody measured dioxins and the lab number came back passing.

That has a blunt consequence for anyone buying waste facility simulation to chase thermal efficiency. If the twin's optimum sits above the demonstrated load, acting on it means a new performance test, a new set of reported maximums, and a conversation with the permitting authority (in practice a retest window negotiated around the outage schedule, so budget a season for it). Teams tend to discover that late rather than price it early, and a model that comes out accurate and unusable at the same time is hard to explain to the people who funded it.

The rule does something similar to the data. Carbon monoxide compliance runs on four-hour block arithmetic averages at most units, and on 24-hour daily averages for mass burn rotary waterwall units and refuse-derived fuel stokers, per 60.58b. Each hourly average needs at least two valid data points. Valid data has to cover 90 percent of operating hours in a quarter and 95 percent across the year, and when the analyzer is down you fall back to an approved reference method, not to your own estimate. None of that is negotiable in software.

Feedstock is the input nobody measures

Everything genuinely hard about predictive plant modeling on a waste line lives in the bunker. Steam flow, oxygen, grate speeds, flue gas temperatures: you know those to within a few percent. You don't know the moisture, ash fraction or calorific value of what the crane just dropped into the chute, and those are the inputs driving every variable downstream of them.

Davide Manca and colleagues at Politecnico di Milano published the clean statement of this problem in 1998, in Computers & Chemical Engineering. Steam flow in a municipal waste incinerator responds inversely to a change in waste feed. Add waste and steam output can fall, because the feed's moisture content and initial temperature impose a heat demand that pulls combustion temperature down before it pushes anything up. A simple feedback controller chasing that signal produces, in their words, unacceptable oscillations of the controlled variable. They built a compensation algorithm on a dynamic simulation model and ran the strategy on-line at an Italian plant through a PC wired into the DCS. That was nearly thirty years ago, and it's still the sharpest explanation I've read of why waste combustion loops misbehave in ways gas loops never do. Autonomous process control only earns trust after the loop survives a day when the bunker lies to it, and the inverse steam response is that day.

Which points at an inversion most twin projects get backwards. The useful version of a twin on a waste line works as a soft sensor for the feedstock rather than as a predictor of the plant. Wien Energie's hazardous waste facility in Vienna is the cleanest production example I know of: a dynamic physical model of two rotary kilns runs alongside the real kilns, and the team matches simulated sensor readings against measured ones. The mismatch between the two is the product, because it's what lets them classify what each delivery was and how it burned, with TU Wien supplying the thermal classification of the input streams. According to the project write-up, permanent monitoring of that kind could cut downtime by up to 25 percent and avoid roughly 500 tons of CO2 from the fuel oil otherwise burned on restarts, both labeled as projections rather than measured results.

Let me own the twin-flavored version of a mistake I've made for years. When the residual between model and plant goes wide, the instinct is to add physics: finer grid, better kinetics, one more reaction. The residual is usually an instrument. A drifted thermocouple, a flow meter reporting set point instead of actual, a tag remapped during a control upgrade that nobody reflected in the historian. On our vision lines we've measured classification accuracy above 98% on live material [from our deployment data], and the step changes came from illumination and belt loading rather than from deeper networks (recall is the first thing to sag when the stream turns wet, and recall is the number that costs money). It's the same lesson one layer down from the combustion model.

The identical failure mode turns up in predictive maintenance across waste processing lines, where a model confidently flags a healthy bearing and misses the one that's about to go, because the vibration channel moved and the training data didn't.

And none of this protects you from the pumps. Amager Bakke in Copenhagen is about as heavily modeled a waste plant as exists anywhere, and in June 2017, roughly a year into commercial operation, one of its two lines broke down on its feedwater pumps. According to reporting in The Local at the time, the operator had to apply for an extraordinary permit to store incoming waste in bags, and the first line came back after about two weeks. The supplier, B&W Volund, had already taken a heavy write-down on technical problems the previous autumn, and the operator's then-director said in October 2016 that earlier delays alone would cost around 100 million kroner. Simulation is not a maintenance program.

Reconcile the twin against the compliance record before anyone trusts either

A plant with both systems running ends up keeping two accounts of itself. One is the compliance record: four-hour blocks, daily calibration drift checks and quarterly accuracy audits under the rule's appendix procedures, with gaps filled by an approved reference method. The other is the twin's state estimate, assembled from one-second tags pulled off the historian over OPC UA. They won't agree, and when they diverge the compliance record wins, because it's the account with a defined measurement procedure behind it, and a closed loop whose numbers can't be lined up against that record is a liability at audit time rather than an asset.

So the first deliverable on a twin project shouldn't be the solver. It should be a reconciliation layer: every model input mapped to a specific instrument tag, with that tag's calibration history and last drift check attached, and a logged residual for every tag that also feeds a reported number. Dull work, and it's the work that decides whether the twin survives its first public disagreement with the stack monitor.

And once you trust it, point it at the right objective. We wired the real-time PLATTS index into our own platform so output mix follows the price signal instead of a fixed recipe, and a twin needs the same treatment: optimizing steam when the money sits in the spark spread is a well-engineered wrong answer. Gren's plant in Klaipeda published the clearest version of this. Per the operator's June 2025 account, its twin was used to analyze and then implement a deliberate cut in boiler load through periods of negative electricity prices, with waste throughput unaffected.

ISO 23247-1, published in October 2021, is the framework document people cite in these projects, and the useful thing about it is what it declines to do. It defines digital twins of observable manufacturing elements, equipment and materials and processes and people among them, then explicitly declines to prescribe data formats or communication protocols, on the grounds that different manufacturing domains use different standards. Read that as a warning about where the money goes. The solver is maybe a tenth of waste plant simulation; the tag map, the clock synchronization and the calibration provenance are the rest, and they're the parts no vendor demo shows you. That gap is where the integration debt accumulates.

Where does this stop holding? Below a certain scale the whole construct changes shape. A small line without continuous emission monitoring has no second account of itself to reconcile against, so the twin becomes the only record, which is a worse problem than disagreement. Outside the United States the envelope is drawn differently, since the demonstrated-maximum construct is an artifact of the US rule and European plants work from BAT conclusions instead. Still, the harder constraint is physical: a lower-temperature thermal conversion line has a narrower stable window than a mass burn grate, which makes feedstock inference harder rather than easier. And neither the Vienna nor the Milan work transfers onto a plant whose historian retains 30 days.

There's a scale caveat buried in the academic record too. Martin Hammerschmid's 2024 dissertation at TU Wien sets out a four-stage ladder, digital model to digital shadow to digital twin to digital predictive twin, and tests it on four processes. The one where a model predictive controller closed the loop on real hardware was a 100 kWth biomass-to-gas pilot. The commercial hazardous waste incineration application stayed at the concept stage. That's a fair read of the field: closed loops get demonstrated at pilot scale with controlled feedstock, while the commercial waste plant twins in production today are overwhelmingly advisory.

Which brings up the greenfield case, since simulating before you build is the usual pitch. For a plant that doesn't exist yet, what you can build is a digital model rather than a twin: no live data to synchronize against, no instrument to be wrong. Still worth doing, and the cheapest way to catch a bunker sized for the waste study instead of the waste. Expect it to be good at sizing and poor at control.

If you're scoping one of these, start by getting your most recent compliant dioxin and furan test report out of the file and writing down two numbers from it: the four-hour average load the unit held through the test, and the inlet temperature at the particulate control over those same four hours. Hand both to whoever is specifying the twin, as hard constraints, before anybody picks a solver or signs a vendor. If nobody can locate the report, you've found your first result, and the scoping conversation should wait until someone does.

Disclosure: I'm Chief Automations Officer at Renewable Waste Energy and the architect of its Optimal Waste Intelligence platform, so I have a commercial interest in people building these well.

Sources & Notes

The operating-envelope and monitoring language comes straight out of 40 CFR 60.58b. It governs large municipal waste combustors; existing units pick up equivalent requirements through state plans, so confirm which path covers yours before quoting a maximum to a regulator.

On the inverse steam response, read Manca, Rovaglio, Pazzaglia and Serafini, Inverse response compensation and control optimization of incineration plants with energy production, Computers & Chemical Engineering, 1998. Still the paper I'd hand to anyone who owns a combustion loop.

Cloudflight documents the Vienna rotary kiln work as a data-driven optimization project for Wien Energie. Its downtime and CO2 numbers are projections, which is how I've cited them.

The four-stage taxonomy and the pilot-versus-commercial gap come from Martin Hammerschmid's open-access TU Wien dissertation, From digital model to digital predictive twin (2024).

US fleet context is from the EIA waste-to-energy explainer, figures dated May 2026. ISO 23247-1:2021 sits behind a paywall; the scope language above comes from its published preview pages.