Choosing a CEMS: Extractive, In-Situ, or PEMS

Most continuous emissions monitoring systems get specified to pass a compliance test and then forgotten until the next audit. That's a waste. The same gas analyzers and flow monitors that prove your stack is inside its limits also carry a live signal about how the combustion plant underneath them is actually running. A NOx trend that climbs every afternoon, a CO spike that tracks a specific waste batch, an O2 reading that drifts after a soot-blow — these are process data first and compliance records second. The question worth asking is which CEMS architecture gives you both, because the choices trade off against each other and the right answer depends on the plant.

This piece compares the main approaches — extractive sampling (hot-wet, cold-dry, and dilution), in-situ analysis, and predictive emissions monitoring — against the criteria that decide which one you should live with: accuracy and calibration burden, latency, reliability and availability, maintenance, and cost. First, though, the regulatory floor, because that's what sets the non-negotiables.

What "continuous" has to mean before it means anything else

A CEMS isn't continuous because the vendor says so. It's continuous because it meets a defined performance specification and keeps meeting it. In the United States, gas monitors on stationary sources are certified against the EPA Performance Specifications in 40 CFR Part 60, Appendix B. For SO2 and NOx, Performance Specification 2 sets the bar: the relative accuracy of the monitor, checked against a reference method, must be no greater than 20 percent of the mean reference value, per EPA Performance Specification 2. Pass that once and you're certified. Keeping the certification is the harder part.

That's the job of Appendix F to Part 60, the quality-assurance procedures. They turn a one-time test into a permanent control loop. According to 40 CFR Part 60, Appendix F, the monitor's calibration drift gets checked at least once daily, a cylinder gas audit runs in three of four calendar quarters, and a relative accuracy test audit happens at least once every four quarters. Miss the accuracy targets and the data turns invalid until you fix and re-certify. So the real spec for any CEMS isn't its headline accuracy — it's how cheaply it holds that accuracy across thousands of operating hours.

Availability is the other hard number. For large municipal waste combustors under 40 CFR Part 60 Subpart Eb, valid hourly averages must be captured for at least 90 percent of operating hours each calendar quarter and 95 percent per calendar year while the unit burns waste, per the EPA Subpart Eb standards. An analyzer that's accurate but offline for maintenance two days a month can blow that budget. Availability, not peak accuracy, is where a lot of installed systems quietly fail.

Europe frames the same problem differently but lands in the same place. The quality assurance of automated measuring systems runs under EN 14181, which splits the lifecycle into QAL1 (suitability of the instrument before installation), QAL2 (calibration against a standard reference method after install), QAL3 (ongoing drift control during normal operation), and an annual surveillance test, per the standard EN 14181:2014. QAL3 is the part operators underestimate: it requires you to run control charts on zero and span checks and prove statistically that the instrument hasn't wandered. That's a data-analysis obligation, not just a maintenance one.

And the European limits leave little headroom. Under the Industrial Emissions Directive, Annex VI, a new waste incineration plant has daily-average air limits of 10 mg/Nm³ total dust, 10 mg/Nm³ total organic carbon, 10 mg/Nm³ HCl, 50 mg/Nm³ SO2, 50 mg/Nm³ CO, and 200 mg/Nm³ NOx for larger units, per Directive 2010/75/EU. The same annex caps how much the measurement itself is allowed to be wrong: the 95 percent confidence interval of a single result may not exceed 10 percent of the limit for CO, 20 percent for SO2 and NO2, 30 percent for total dust and TOC, and 40 percent for HCl. Read that carefully — for dust at a 10 mg/Nm³ limit, your whole measurement uncertainty budget is about 3 mg/Nm³. That constrains which technology you can even use.

The bar has also been rising. The 2019 BAT conclusions for waste incineration extended the list of pollutants that must be measured continuously to include not just NOx, CO, SO2, HCl, HF, dust and total VOC but also mercury, with ammonia added continuously wherever SNCR or SCR is in use, per Commission Implementing Decision (EU) 2019/2010. Continuous mercury monitoring on a waste line is a real engineering step up, and it pushes hard on the architecture choice below.

The architectures, in one pass

Strip away the brand names and there are three families.

Extractive systems pull a sample out of the stack, carry it down a heated line, condition it, and present it to analyzers sitting in a shelter at grade. They come in three flavors. Hot-wet keeps the whole path above the dew point so nothing condenses and water-soluble gases like HCl and SO2 survive to the analyzer intact. Cold-dry chills the sample to drop the water out, then measures a dry gas — simpler, but you can lose soluble species in the condensate and have to correct back to wet basis. Dilution injects a precise ratio of clean dry air at the probe so the sample never reaches its dew point, letting analyzers run near ambient concentrations. Each conditioning step is also a source of bias if it's not controlled; the EPA's own analysis of sources of bias in extractive CEM systems walks through how leaks, adsorption, and incomplete moisture removal each pull the reading off true.

In-situ systems skip the sample line. The measurement happens in the stack itself — typically a light beam across the duct (cross-stack) or into a short probe (in-stack). There's no conditioning train to maintain and no transport lag, but the optics live in the flue gas, alignment matters, and you measure on a wet, hot, particulate-laden basis whether you like it or not.

Predictive emissions monitoring systems (PEMS) don't measure the pollutant at all. They infer it. A model — regression or a learned function — maps process variables you already log (fuel flow, combustion air, temperatures, load) onto an emissions estimate. The EPA recognizes the approach: Performance Specification 16, finalized in 2009, sets out how a PEMS has to pass a relative-accuracy test before it can stand in for a hardware monitor, per EPA Performance Specification 16. A PEMS is only as good as the process instrumentation feeding it and the conditions it was trained across.

Accuracy and the calibration you'll actually do

On paper, a well-run hot-wet extractive system is the reference-grade choice. It preserves soluble gases, it lets you span analyzers with traceable cylinder gases at the analyzer inlet, and it's what most reference-method comparisons are built around. That accuracy isn't free. You're maintaining a heated line, a probe filter, a chiller or Nafion dryer, pumps, and flow controllers, and every one of them can introduce the bias the EPA bias analysis describes.

Dilution-extractive trades some of that for simplicity downstream, but the dilution ratio becomes the thing you must hold dead steady — a 1 percent error in the ratio is a 1 percent error in every reading, and ratios drift with probe temperature and supply-air quality. In-situ optical systems can be very accurate for the species they suit (NOx, SO2, sometimes HCl by laser), and they sidestep sample-handling bias entirely, but they're harder to span check because you can't easily flood the stack path with a known gas; you lean on internal calibration cells and audit gases instead.

PEMS is the outlier. Its accuracy is conditional. Inside the operating envelope it was trained on, a good model can meet PS-16's relative-accuracy requirement. Push the plant to a load point or fuel mix it never saw, and the estimate degrades with no physical sensor to catch the error. For a waste-to-energy line burning a heterogeneous feed, that's a real limitation — the input varies more than a gas turbine's ever will. PEMS earns its place as a backup, a validation cross-check, or a monitor on a stable auxiliary burner, not usually as the sole certified monitor on a variable waste combustor.

Latency: what the number is fast enough to do

Here's where the compliance-versus-control split gets concrete. For a daily-average limit, a sixty-second response time is irrelevant. For closing a combustion loop, it's everything.

In-situ wins on latency outright. The measurement is in the stack, so there's no transport delay — the lag is essentially the analyzer's own response time. Extractive systems add the time for gas to travel the sample line plus the conditioning train; on a long heated line that can be tens of seconds before a change at the stack shows up at the analyzer. Tens of seconds is fine for reporting. It's marginal if you want emissions feedback inside an air-control or SNCR-dosing scheme, where a fresher signal lets you trim reagent and catch a CO excursion before it becomes a violation.

PEMS is, paradoxically, the lowest-latency option of all, because it's computed from process tags that update every scan of the control system. There's no gas to move. The catch is the same as before — it's a prediction, so it leads the true value only as far as the model is trustworthy. But for early warning, a PEMS running alongside a hardware CEMS can flag a developing excursion seconds before the stack analyzer confirms it.

Reliability and availability: the 90 percent you have to hit

Remember the availability budget — 90 percent per quarter, 95 percent per year under Subpart Eb. Architecture drives how easily you make it.

Extractive systems have more parts that fail: probe filters blind, lines clog or develop leaks, chillers and pumps wear. The upside is that almost all of it sits at grade in a shelter, so you can service it without a stack entry, and a maintenance-friendly layout keeps mean time to repair short. In-situ systems have fewer consumables, but when the optics need attention the work is up the stack, often at height, and an alignment knocked out by thermal cycling or vibration can take a channel offline until someone gets to it. Redundancy economics differ too: doubling up an extractive analyzer in a shelter is cheap relative to mounting and aligning a second in-situ head on the duct.

PEMS has no field hardware of its own to fail — but it inherits every fault of the sensors it reads. Lose the fuel-flow transmitter and the model goes blind. That's why PEMS shines as a gap-filler: when a hardware monitor drops out for calibration or repair, a certified PEMS can supply substitute data and protect the availability budget, which is exactly the role many operators give it.

Maintenance and lifecycle: who fixes it at 3 a.m.

So who fixes it at 3 a.m., and with what skills? The honest comparison here isn't about the instrument. It's about your people. Extractive systems demand a technician comfortable with gas handling, leak checks, and a conditioning train — a known, teachable skill set, with spares you can stock. In-situ systems demand optical alignment and stack access, a rarer skill and a permit-to-work every time. PEMS demands something different again: someone who can tell when a model has gone stale and needs retraining, which is a data skill, not a wrench skill.

EN 14181's QAL3 makes this explicit on the European side. You're obligated to keep control charts on the instrument's zero and span and to act when they signal drift, per EN 14181:2014. That's a recurring analytical task regardless of architecture, and it's the seam where monitoring and process engineering meet. A plant that treats QAL3 as a checkbox tends to discover its drift problems during the annual surveillance test, which is the worst time to find them.

Cost: where the money actually goes

Capital cost roughly tracks complexity: a full hot-wet extractive system with a shelter and redundant analyzers is the heaviest spend, in-situ sits lower on hardware but can need structural and access work on the stack, and PEMS is mostly software and engineering with little field hardware. But capital is the smaller story over a fifteen-year asset life. The operating cost — consumables, calibration gases, technician hours, and the cost of invalid data when availability slips — is what separates the architectures, and it runs opposite to the capital ranking. The system that's cheap to buy can be expensive to keep certified, and vice versa. Specify on total cost of ownership and the QA workload, not the purchase order.

The comparison, side by side

Note the table is organized by what you're deciding on, not by vendor. No column wins every row. That's the point — the architecture follows the criteria that matter most at your stack, and on a waste-to-energy line the criteria that matter most are usually soluble-gas fidelity and availability, which is why hot-wet extractive remains the workhorse there even though it's the most maintenance-hungry.

Turning CEMS data into control, not just records

The compliance system is already the densest, best-calibrated sensor array on the plant. Treating it only as a regulatory recorder leaves most of its value unused. Three moves change that.

The first is drift intelligence. The QAL3 control charts EN 14181 already requires — Shewhart or CUSUM on zero and span — don't just keep you compliant; read the other way, a slow span drift is an early sign an analyzer cell or a dilution probe is degrading, well before it fails an audit. Wiring those charts into the same historian that holds process data lets you catch the maintenance need on a planned outage instead of an emergency one.

The second is contextualizing the emissions trend against operations. A CO peak means more when you can line it up against the grate speed, the feed crane cycle, and the secondary-air damper position at the same instant. That's a data-integration job: emissions tags, control-system tags, and a common timestamp in one place, queryable. When emissions, combustion, and feed data sit together, the afternoon NOx climb stops being a mystery and starts being a setpoint you can adjust. Building that layer — ruggedized acquisition at the stack, time-aligned with the rest of the plant, with models watching for the patterns operators can't eyeball — is the work behind the Zoniax edge telemetry and analytics platform.

The third is the soft sensor. The same modeling that makes a PEMS work can run alongside a hardware CEMS purely as a validator. If the model and the analyzer agree, confidence is high. If they diverge, one of them is wrong and you've caught a fault either way — a clogging probe, a stale model, a drifting cell. That cross-check costs almost nothing once the process data is already flowing, and it directly protects the availability budget by flagging questionable hours before they're reported. For operators weighing how to stand this up against in-house resources, it's the kind of integration the Zoniax industrial AI deployment services are built around.

Which fits your plant

Start from the gas, not the brochure. If your priority pollutants are soluble — HCl, SO2, mercury on a waste line — hot-wet extractive is hard to beat on fidelity, and you accept the maintenance load that comes with it. If you have several stacks feeding one analyzer shelter over long runs, dilution-extractive earns its keep by simplifying the downstream side. If you want a fast signal to close a combustion or reagent loop and your duty is low on consumables, in-situ optical pays back in latency and uptime. And if you need a backup that protects availability, a validation cross-check, or coverage on a stable auxiliary source, a PS-16-compliant PEMS is the cheapest insurance you can buy — as a complement, rarely as the sole monitor on a variable feed.

Whatever you pick, specify it for the QA workload, not the certification day. The system you can keep accurate and available for 95 percent of the year, with the people you actually have, is the one that turns a compliance cost into an operational asset. The rest is just paperwork on a stack you're not really watching.