Mine air quality studies tend to fail in one of two ways: they model the right pollutant at the wrong scale, or they report a correlation and call it prediction. This sets out how to design a study for an open pit that survives technical review, and how to read the error metrics honestly.
Define the question before choosing the model
A study asking what the air quality of the mine is has not asked a question. Split it into the two questions regulators and community groups actually pose:
- Attribution — how much of the measured concentration comes from this operation, versus regional transport and background?
- Prediction — given tomorrow production, haul routing and weather, what will the concentration be at the receptor?
Attribution is a dispersion problem. Prediction is a time-series problem. Running one method for both is the most common design error, and it is why so many site studies produce defensible-looking numbers with no operational use.
Build the network around sources, not around convenience
Receptors belong where people are, at the lease boundary, and upwind of nothing. Sources must be instrumented at emission height: a crusher monitored at ground level describes the dust on the crusher, not the plume leaving it. Meteorological measurement is not optional — without wind speed, wind direction and boundary layer height a dispersion model has no forcing function and its output cannot be interpreted.
Anchor to the regulatory network
Any study intended for compliance or community engagement should anchor to CPCB NAAQS 2009 and, for health framing, to the WHO 2021 guidelines. Where a mine sits inside the Indian standard but outside the WHO value, say so plainly: that gap is usually why a community dispute escalates beyond the technical file.
Dispersion model versus machine learning
These are not competitors. They answer different parts of the problem.
- Regulatory-grade dispersion models (the AERMOD family) need meteorological and source characterisation, generalize to conditions never measured, and are what an environmental clearance condition will reference.
- Machine learning on historical monitoring captures site-specific behaviour a generic model misses and forecasts short horizons well. Its weakness is extrapolation: a model trained on calm dry-season data has no basis for a still, humid monsoon week.
- Hybrid approaches use the dispersion model to generate features and a learned model to correct its bias. This is the most defensible structure where reviewers expect both methods.
Reporting error honestly
Report the distribution of the error, not a single headline number. For a continuous concentration series the standard set is:
- MAE in µg/m³ — interpretable, in the units of the thing being predicted
- RMSE — penalises large misses, which matters because a regulatory exceedance is by definition a large miss
- R² — necessary but not sufficient. A high R² with a biased model is common, because R² rewards reproducing the variance of the series rather than its level
- Mean signed error — the metric most often omitted, and the one that reveals whether a model systematically over- or under-predicts
- Exceedance detection rate — the share of true breach events correctly flagged. This is the operational metric and it is almost never reported
For exceedance prediction, plain accuracy is the wrong headline. A model that never predicts an exceedance is highly accurate at a site where breaches are rare and completely useless. Report precision, recall and the confusion matrix for breach events.
Validation needs genuinely independent data
Training and testing on the same months produces optimistic results that will not survive at a second site or in a second season. Hold out at least one full monsoon period, or an entire production shift cycle, and report performance on that holdout separately. Report across seasons: a model validated only in the dry season has not been validated.
Limitations to state in the write-up
- Single-site results do not transfer to another orebody or climate zone
- Monitored parameters are limited by instrument cost, so unmonitored fractions of the PM mass are inferred rather than measured
- Attribution to individual sources depends on how well the receptor geometry separates them
- Regulatory limits are themselves under revision, so a compliance conclusion has a shelf life
Where the hardware fits
None of this works without a defensible input stream. Continuous PM2.5 and PM10 monitoring with a real-time display at the gate, the crusher and the village-facing boundary gives the network the temporal resolution that both attribution and prediction require — and gives the site a live number to put in front of a community instead of a quarterly figure. Our ambient air quality monitoring system provides exactly that, with a DMD display readable from the control room and on site.