Data Center Corridor Air Monitoring

Every summary figure near this chart is given twice, once over the whole record and once with 2026-07-16 to 2026-07-19 removed. A statistic pooled across the smoke episode and ordinary conditions sits between two regimes and describes neither. At the regulatory monitor the mean over the full record is 8.44 µg/m³ across 3,040 valid hours, and 6.46 µg/m³ across 2,944 hours with the smoke window removed. The highest single hour in the whole record, 282.5 µg/m³, falls inside that window; outside it the highest hour is 28.5 µg/m³.

Four separate periods are voided in the source data, and the chart shows each as a gap. DEQ's August 7 report records regulatory PM2.5 voided 2026-05-07 to 05-10 and 2026-06-16 to 07-14 for failed data quality analysis and instrumentation issues, regulatory NO2 voided 2026-05-07 to 05-10, sensor PM2.5 voided 2026-03-03 to 03-08 at every site for high humidity and sensor data quality, and PM2.5 collection at one retired site ending 2026-05-30 after an air flow failure in the particle sensor. In the released record the fully void days are narrower than the stated windows, 05-08 to 05-09 and 06-17 to 07-13, because the days at each boundary are partly valid.

The DEQ column is the August 7 edition apart from the daggered row, and the GENARCH column stops at 2026-08-07 07:00 EST. DEQ produces each edition on Friday morning from data through 07:00 that day. The cutoff is the agency's specification, not an assumption made here (DEQ staff, correspondence, 2026-08-16). DEQ reissues this analysis every week, so a comparison that names no edition is measured against a target that moves.

Three later editions have since shipped, and the column stays on August 7 anyway. That is the window the GENARCH column is computed over, and it is where the regulatory record released under FOIA 26-4646 ends, on 2026-08-10. Setting a coefficient refitted on a longer record beside a fit that stops at August 7 would compare two spans rather than two calculations. Extending the comparison to the August 28, 2026 edition needs regulatory data that has been requested and not yet released.

The APEX 14 collocation period closed on 2026-04-08, which fixes the two APEX 14 regressions: no later week of data can enter them, and they print identically in all four editions. The APEX 5 window is still open, so each edition refits those two on a week more data than the last, and their coefficients move accordingly.

Three of the four comparisons are checked by the build, against a tolerance of 0.05 on slope and intercept and 0.03 on R², and all three agree. DEQ publishes these coefficients to two significant figures, so the claim is that these fits round to DEQ's values at DEQ's own precision. Agreement to three decimals is not something this data can establish, and part of every difference in the table is that rounding rather than a difference in the underlying fit.

The APEX 5 NO2 row is recorded rather than checked. Its DEQ side is the 2026-08-19 correspondence, refit over DEQ's record to that date, while the GENARCH fit beside it stops at 2026-08-07 07:00 EST. The fit cannot be extended to meet it, because a collocation fit needs both sides of each hourly pair and the regulatory record released under FOIA 26-4646 ends 2026-08-10. Neither side can be moved, so no change here makes the comparison valid, and a check that passed on it would not mean what the other three mean. Both figures stay in the table and the check returns when the extended regulatory record arrives.

The rows do not carry equal weight. The two APEX 14 rows compare against a closed window whose published values cannot move, so they are the same calculation on both sides. The APEX 5 rows compare against a window DEQ is still refitting, and the checked one lines up only because the column above is pinned to the same August 7 span. Read as a count of matches the four would say more than they can support.

The fourth took a round of correspondence to get there. The APEX 5 NO2 figure printed in the August 7 edition did not reproduce here, so the question went to DEQ. DEQ traced it to the plotting call behind the chart: the axis limits there bounded the model fit as well as the plotted view, so points outside them never entered the regression. Refitting without the limits over the August 7 window gives y = 1.8 + 0.32x, R² = 0.21, and DEQ sent the paired hourly measurements the regression runs on (DEQ staff, correspondence, 2026-08-19).

The correction is now in print. The August 21, 2026 edition carries the unbounded fit, at y = 1.8 + 0.31x, R² = 0.2 over a record two weeks longer than the one this table compares on. Those coefficients describe that longer window, so this page cites them here rather than setting them in the column above.

The same fit computed from the source tables gives y = 1.847 + 0.317x, R² = 0.204 on 2,746 paired hours, inside tolerance of the corrected coefficients. That fit keeps the sensor's exact-zero readings, 17.6 percent of the APEX 5 window's NO2 hours in runs as long as 45 hours, because DEQ keeps them (DEQ staff, correspondence, 2026-08-19). Everywhere else on this page they stay excluded as a per-pollutant floor clamp. Dropping them here returns slope 0.340 and R² 0.199 on 2,261 pairs, the figure this page carried before, which left zero handling as the only remaining difference between the two calculations.

Agreement on the coefficients does not make the relationship a strong one. At an R² near 0.21 this is the weakest of the four, and its coefficients are loosely determined. At 10 ppb on the sensor the two lines sit 0.02 ppb apart, on a series whose regulatory mean is 3.5 ppb.

Applying the same null-code and flag exclusions and an 18-of-24 completeness rule, recomputed daily averages match DEQ's published daily file on 118 of 119 comparable days after truncation to one decimal, with a mean absolute difference of 0.0017 µg/m³. DEQ truncates published values rather than rounding them.

The single day that differs is 2026-03-10, where the recomputation gives 5.0 against a published 5.2. DEQ staff reviewed that date, said that recomputing from hourly values also gives 5.0, and noted that they were unsure why the invalid hour at 10:00 was not excluded from the published daily export (DEQ staff, correspondence, 2026-08-16).

DEQ's analysis: Data Center Air Quality Analysis, August 7, 2026. Project page: Virginia DEQ Data Center Air Monitoring.

The sentence naming this sensor as the only one above 35 µg/m³ stood word for word through the first three editions while the figure it described fell in every one: 40.4 on Aug 7, 38.2 on Aug 14, 36.0 on Aug 21, 33.8 on Aug 28. The August 28, 2026 edition drops that sentence and states instead that no site's 98th percentile of daily averages sits above the 35 µg/m³ level. DEQ staff said they expected the figure to keep falling as the record lengthened (DEQ staff, correspondence, 2026-08-16). The first three editions are archived under docs/deq-reports/ and the Aug 28 edition under pipeline/sources/deq/, because DEQ's published page carries only the current one.

DEQ staff note that the wildfire smoke data heavily skews the 98th percentiles at every site, and that before the smoke arrived they sat at or below about 20 µg/m³ everywhere (DEQ staff, correspondence, 2026-08-16). The percentiles in the table below include that window.

Restricting every sensor to the window in which all twelve were collecting, 2026-06-18 to 2026-08-07, changes the ordering. Five of nine sites sit above 35 µg/m³ on matched windows, and sterling-ms ranks fourth rather than first. The common window is not a clean control on its own: it falls entirely in high summer and contains the smoke window, so it removes the unequal-length difference and leaves the unequal-seasonality one in place.

All 46 pairwise correlations fall between -1.000 and 1.000. This is a description of how the series move together and nothing more. High spatial correlation does not establish that local sources contribute nothing: the deployed sensors may sit largely upwind, and a local increment can be masked by a dominant regional signal.

The ratio between the highest and lowest site reading in a given hour is not a usable discriminator here. It fails below roughly 30 µg/m³, where small absolute differences produce large ratios. Across 5,076 hours the median ratio is 1.86, and low-concentration hours reach ratios above 235 from differences under a microgram per cubic metre.