Methods and data

Sample data: LBNL's public simulated chiller plant. The Plant Check samples and every sample-data figure on this site use two files from this public research dataset:

Granderson, J., Lin, G., Chen, Y., Casillas, A., et al. LBNL Fault Detection and Diagnostics Datasets. Lawrence Berkeley National Laboratory, 2022. Open Energy Data Initiative. https://doi.org/10.25984/1881324 (CC BY 4.0)

• Licence: CC BY 4.0.
• Changes: July slices, resampled from 1-minute values to 15-minute averages; point names unchanged.

The two files:

• Sample plant, no faults (July): Three chillers and three cooling towers serving a 12-storey Chicago office, fault-free case.
• Sample plant, cooling-tower fouling (July): The same plant with cooling tower 1's heat-transfer coefficient reduced to 65% (LBNL's 'fouling' fault).

Both files have the same 77 points and a timestamp column, Datetime, with LBNL's point names as published.

The 77 points include no make-up, blowdown, conductivity or chemistry points, so this site shows no water-side numbers for the sample plant. Water-side numbers appear only in the calculators, driven by your own inputs.

LBNL's point inventory labels OA_TEMP as outdoor dry-bulb temperature and OA_TEMP_WB as outdoor wet-bulb temperature. Mapped by those labels in the fouled-tower file, Plant Check's physics check reports:

The reference assignment comes from the plant's documented control sequence, LBNL inventory §1.2, Eq. 3 (tower leaving-water setpoint = wet-bulb + 8 °F). Over 2,976 rows, the setpoint column CT_SW_TEMPSPT minus each outdoor column:

The setpoint follows OA_TEMP at a mean offset of 7.93 °F with a standard deviation of 0.76 °F, as the sequence requires of the wet-bulb; against OA_TEMP_WB the standard deviation is 4.86 °F.

Plant Check's check flags the pair when mapped by name; our reference mapping assigns them by the control-sequence evidence. The check never changes a mapping. The files are used as LBNL published them.

Plant Check runs 6 fixed engineering rules on every proposed mapping and again on the mapping you confirm. A rule that fires shows its warning; it never changes a mapping.

Below is every check on the fouled-tower file with the points mapped by their inventory labels: 24 checks, 1 flagged. With the reference mapping the same file gives 24 checks, 0 flagged.

Plant Check calculates every number it reports directly from the points you confirm, with the formulas below. It reports a metric only when the points it needs are mapped: the base requirement, plus any listed against it. We verify that each requirement is both sufficient and necessary.

After the metrics are calculated, AI writes the explanation: a headline, findings, questions for the site and caveats. It is given the calculated metrics, notes on missing points, and which physics checks flagged. For the samples it is told where the data came from, but not which fault was introduced, so an explanation that points at the towers got there from the numbers. Apart from the metric values themselves, nothing it is given states a number: labels that name a percentile or threshold are reworded, and a flagged check is described without its measured share.

The rule it works under: it may not write a number, in digits or in words. Where it needs a figure, it cites one of the calculated metrics, and we insert the value and unit.

Every explanation is checked before you see it. A digit, a number word (such as eight, forty-four, half, twice, doubled or percent), or a citation of a figure we did not calculate rejects the whole explanation, and you see an error instead. In an accepted explanation, every number traces to a calculated metric.

These drafts are examples we wrote, not AI output. AI output appears on this site only from a recorded run, tagged Written by AI with its date, and shown unedited.

In Plant Check, AI proposes a role and unit for every point, from its name and a profile of its values. The AI proposes and you confirm: you check or change each row before anything is calculated.

The sample-data figures on this site don't use an AI proposal. They use a reference mapping we wrote from LBNL's point inventory and checked against the data's physics: 23 of the 77 points carry a role.

One recorded AI run per file, shown unedited: how its proposed mapping compares with the reference mapping.

The dataset is public. Download it from the DOI above, take the July slices at 15-minute averages, and apply the reference mapping and the metric definitions above to reproduce any sample-data number on this site. Questions about a number? Write to us at [email protected].