← Back to the map

How the plume forecast works — and how well it works

Methodology and independent validation of the Calvert City industrial plume model.

This page explains, in plain language first and then in technical detail, how the daily plume forecast is produced, what the map is actually showing you, and how it was tested against real air-quality measurements. It covers the plume dispersion model only — the companion Odor Forecaster has its own methodology page.

How this forecast works

Every day, the model estimates how much of each chemical the local industrial facilities release into the air, then uses real, hour-by-hour weather — winds, temperature, and how much the atmosphere is mixing — to carry those releases across the map. The result is a picture of where the plume is likely to drift and roughly how concentrated it is, hour by hour.

The emissions come from the facilities' own annual reports to the U.S. EPA. The weather comes from NOAA's high-resolution national forecast model. The dispersion itself is computed by NOAA's HYSPLIT model — the same tool used by federal agencies to track smoke, ash, and hazardous releases.

How to read the concentration layer

The map includes a concentration layer. It is genuinely a concentration estimate — but like any model, its numbers carry uncertainty, and that uncertainty is not the same everywhere:

• Where the plume goes, and when, is the most trustworthy part. Validation testing (below) shows the model reliably predicts when a plume reaches a location.
• Close to the fencelines, the concentration numbers are likely on the high side. In the winter test the model over-stated near-fenceline concentrations by roughly 5–6×. This over-statement is largest in stable, calm, shallow-air conditions (winter nights) and is expected to be smaller in summer's well-mixed air. Magnitude testing so far is winter-only, so no single fixed multiplier is claimed.
• Farther from the sources, the magnitudes line up better — within about a factor of 1.5 of measured values.
• The map shows a typical day, not a worst case. Emissions are entered as a steady year-round average, so the layer does not capture short, intense upset or batch-release events, which can be much higher than what is shown. This is the single biggest limitation on the absolute numbers.

Bottom line: trust the pattern and timing most; read the absolute concentrations as an order-of-magnitude guide, most reliable away from the immediate fencelines.

How the model runs

1. Emissions. Annual air releases for each facility and chemical are taken from the EPA Toxics Release Inventory (TRI) and converted to a steady hourly emission rate (annual pounds ÷ 8,760 hours), split between tall-stack releases and lower fugitive (leak/area) releases.
2. Weather. Real hourly meteorology comes from NOAA's HRRR model (3 km resolution): winds, temperature, and boundary-layer depth (how much the air is mixing).
3. Dispersion. NOAA HYSPLIT releases model particles at each source and transports them with the HRRR winds, spreading and depositing them realistically.
4. Output. The particles are converted to a ground-level concentration and deposition field for each chemical — the layers you see on the map.

Release height, and why there is no "plume rise" (for now)

Hot gases leaving a tall stack are buoyant and rise before they spread — an effect called plume rise. The deployed model does not simulate plume rise; it releases emissions at a fixed low height (about 15 m for stacks, 2 m for fugitive sources) and lets the plume travel along near the ground. That is a deliberate choice:

• The per-stack engineering parameters (height, exit velocity, gas temperature) that a plume-rise calculation requires are not available for the Calvert facilities — the TRI reports only total mass.
• When plume rise was tested with real stack data in a separate benchmark (Pittsburgh, below), it pushed the plume above the monitors on calm, stable days and lost the model's skill entirely — swinging from over-prediction to large under-prediction. Grounding the plume is the validated, health-protective choice (it errs toward over-warning, not silence).

Plume rise remains a planned future upgrade: source real stack parameters, re-validate, and likely apply it only when the atmosphere is deep enough to avoid pushing the plume over everyone.

How the model was validated

The model was back-tested two ways: directly against Calvert-area EPA canister measurements, and against a data-rich benchmark city (Pittsburgh) with dense hourly monitoring, which stress-tests the same method with real statistical power. Two questions matter: does it get the timing right (when does a plume arrive?), and does it get the amount right (how much is emitted)?

Timing — does the model predict when a plume arrives?

Yes, and significantly so. When the wind opens a corridor from a source to a monitor, measured concentrations spike as predicted; when it doesn't, they don't.

The Pittsburgh benchmark (25 days, ~1,670 site-hours of hourly SO₂) is what gives the timing result its statistical power; the Calvert canister record is sparser but points the same way.

Amount — how accurate are the emissions?

Emission accuracy can only be judged away from the sources, in the "grounding zone" where an elevated plume has fully mixed down to the surface. There, predicted and measured concentrations agree to within about a factor of 1.5 — i.e. the reported inventory is the right order of magnitude, not grossly over- or under-stated.

Right next to a source the story is different, and this is why a single concentration accuracy is not quoted: near-field amounts depend heavily on the release-height assumption, swinging from over- to under-prediction. This sensitivity is exactly why the deployed model keeps the plume grounded and why the concentration layer should be read as a guide near the fencelines.

Limitations

• Steady-state emissions. A single year-round average rate is used, so short upset or batch-release events (which can dominate real exposure) are not represented.
• Near-field magnitude is uncertain and depends on the release-height assumption; the magnitude test so far is winter-only.
• Terrain and resolution. Real river-valley terrain interacts with weather at scales finer than the 3 km meteorology, weakening accuracy very close to the sources.
• Not every chemical is modeled. The inventory occasionally lists high-emitting chemicals the model doesn't yet track (for example, methanol at ~780,000 lb/yr in a recent year); these are flagged automatically so the model can be updated.
• This is a screening and awareness tool, not a substitute for regulatory monitoring or a health assessment.