ERCOT model quality tests: a blind test across six plants

A top renewable asset owner holds a large ERCOT fleet of inverter-based plants. Each plant needs the ERCOT model quality tests under PGRR 109. The owner had a set of consultant studies, run by hand in the first half of 2026: one report and one model package per plant, in PSCAD and PSS/E, with TSAT on one plant.

The owner wanted to know whether those studies hold up before relying on them. GridStrong ran a blind test: all 17 tests on six of those plants, on the delivered models, with no reference to the consultant results until our runs were complete. Then we overlaid the two case by case, matched where the model and the report agreed, and documented every place they did not.

PGRR 109 defines 17 standard test cases for every plant: a flat run, voltage steps, low- and high-voltage ride-through, voltage dips, frequency response with and without headroom, phase-angle jumps and a short circuit stepped down to a weak grid. Each case starts from a defined operating point, including a 0.95 power factor, and the plant's response is judged on the traces it produces.

A study holds up when the delivered model reproduces the report, the text agrees with the plots, and the PSCAD and PSS/E models give one answer.

We took each plant's delivered model package unchanged: the PSS/E case and dynamic data, and the PSCAD plant, with only our standard test harness in place of the study's grid source and test controls. We ran all 17 tests per plant in PSCAD and PSS/E through GridStrong's automated test harness, which sets every case up the same way from the test definition: setpoints, power factor, event timing and short-circuit stages.

Only then did we recover the plotted curves from each consultant report and overlay our results on them, case by case, with settled values, recovery times and trip times measured the same way on both sides. Where a result differed, we reran with the one setting the consultant run appeared to use, to name the cause rather than guess. Each plant got one report: matched charts first, then every difference with its cause.

What the blind test found in the consultant studies, with the number of plants affected out of six:

• Report text contradicts its own plots (6 of 6). A phase-angle ride-through limit in the text that the traces do not show; a voltage recovery to 0.98 pu described where the plot holds 0.9 pu; an event time of 21 s in the text where the plots show 1 s; a TSAT run called stable at the weakest grid where it stops one stage earlier. The automated run follows the test definition, so each claim is checked against a trace, not against prose.
• "Passed all tests" beside pages that show inverter trips (3 of 6). In the automated run, pass or fail comes from the results, case by case.
• The delivered model does not reproduce the report without undocumented changes (3 of 6). Template feeder sections left in the PSCAD case, controller start values fixed at one operating point for every case, a power setpoint on the wrong base, capacitor banks and tap timers set per case by hand. The harness starts every case from that case's own setpoints; each needed change is named and shown with and without it.
• Starting reactive power off the required 0.95 power-factor point (2 of 6), by up to 10 Mvar at one plant. Every automated case starts at the required setpoint.
• A plotted point-of-interconnection reactive power curve the plant cannot produce (1 of 6). The curve carries a fixed 6 to 9 Mvar offset in all eight lead and lag cases, which no element of the plant network can produce; the study's own inverter readings, carried through its own network, land within about 1 Mvar of our result. The harness measures at the point of interconnection and checks against the inverter readings.
• PSCAD and PSS/E presented as one study while differing by up to 7.5 Mvar (3 of 6), and in places one platform trips where the other rides through. Each platform is compared with itself, and the gap between the delivered models is reported, not hidden.
• Reactive power drifting where it should hold steady (4 of 6), by up to 12 Mvar. The same model holds steady when set up from the test definition.
• The plant stopping, or not returning to output, at the phase-angle jump back to zero (4 of 6). The same model rides through in every automated run of that jump, including runs that copy the consultant start.
• Paperwork errors (3 of 6). A verification checklist carrying another plant's controller values, a table with resistance and reactance swapped, a rating base that differs from the model, two cases with the same profile name run two different ways. In the automated run, settings and test parameters come from one source and print into the report.

On all six plants we matched the consultant results where the delivered model and the report agreed: voltage steps within 1.3 Mvar, frequency response within 1.4 MW, and the same short-circuit behavior at the same grid strength. Every remaining difference is documented with its cause, or with the one question only the original analyst can answer. The owner now has a second, independent run of each plant's tests, built from the same files, with every run linked so an engineer can open the traces.

Model quality tests are judged on traces, not on summaries. A hand-built study can pass review on its summary page while its own plots say otherwise. An automated harness sets every case up from the test definition, every time, so the same model gives the same answer and a difference has a cause you can name. A blind test of a fleet's existing studies is a low-risk way to find out what an owner is relying on.