Skip to content

Latest commit

 

History

History
121 lines (82 loc) · 10.1 KB

22-network-error.md

File metadata and controls

121 lines (82 loc) · 10.1 KB

22: Error in the network data

Date: 2026-11-12 · Plan phase: the standing screening question

Goal

Every screen so far was given the network's R, X and C exactly. Network data are not exact: resistance moves with temperature, the equivalent of the rest of the system moves with outages, and a capacitor bank may have segments out. How much error can each screen stand, which network quantity matters, and is this a bigger or smaller problem than noise on the scan?

What was done

  • src/sso_sciml/neterror.py: one case with the plant known exactly and the screen handed a wrong network. Five quantities are each wrong by an independent factor: line reactance, line resistance, series capacitor, parallel path, plant-side impedance. Three screens are evaluated on the wrong network: the plant model closed on it exactly (what a clean scan and fit give), the step 14 hand formula, and the slow mode of the phasor-domain model with its operating point also taken from the wrong network.
  • src/sso_sciml/data/t1.py: impedance_root and operating_point now accept a network in place of the one of the case.
  • scripts/network_error.py: the pass over all 98,635 cases of both plant types, the results file and three figures.
  • tests/test_neterror.py: 5 tests.

Commands

uv run scripts/network_error.py            # 98,635 cases, about 25 minutes on 8 processes
uv run scripts/network_error.py --reuse    # numbers and figures from the saved pass
uv run pytest tests/test_neterror.py

Setup of the study

  • Cases: every solved case of steps 18 and 19, in and outside the training range together: 38,511 doubly-fed, 45,056 converters with a phase-locked loop, 15,068 without.
  • Error: each of the five factors is the exponential of a normal number with standard deviation 2 %, 5 %, 10 % or 20 %. Separately, each quantity alone is moved 10 % up and 10 % down.
  • Truth: the modes of the full model on the right network, by band as in step 20.
  • The plant is exact. Its admittance is taken at the true operating point. Only the phasor-domain screen takes its operating point from the wrong network, as a planning model would.

Results

All numbers are in docs/results/network/network_error.json.

Which network quantity matters

Sensitivity to each network quantity

Median shift of the growth rate of the least-damped mode between 2 and 58 Hz when one quantity is wrong by 10 %, in 1/s, for cases within 5 1/s of the boundary. The sign is that of the shift when the quantity is 10 % too large.

Quantity Doubly-fed plant Converter with PLL Converter without PLL
Series capacitor 0.80, less stable 0.23 0.40
Line resistance 0.48, more stable 0.05 0.03
Plant-side impedance 0.43, more stable 0.16, less stable 0.25, less stable
Line reactance 0.39, more stable 0.42, less stable 0.76, less stable
Parallel path 0.25, less stable 0.15, less stable 0.16, less stable
  • For the doubly-fed plant the capacitor matters most and resistance second. A capacitor 10 % larger than assumed moves the mode 0.8 1/s toward instability. More resistance and more reactance both help.
  • For a converter the reactance matters most and resistance not at all. More reactance hurts, the opposite of the doubly-fed plant: there it lowers the resonance frequency, here it weakens the grid.
  • The slow modes do not care. Below 2 Hz the same 10 % errors move the growth rate by less than 0.001 1/s on the doubly-fed plant and by 0.005 1/s or less on the converters.

The exact plant model on a wrong network

Growth-rate error and unstable cases cleared

Sub-synchronous band of the doubly-fed plant, 10,051 unstable of 38,511. Error of the growth rate near the boundary, then share cleared and unstable cases cleared at four margins.

Error in the data Error, median and 95th percentile (1/s) Margin 0 1 1/s 2 1/s 5 1/s
2 % 0.17, 0.66 73.8 %, 1,155 64.3 %, 15 60.8 %, 3 45.4 %, 0
5 % 0.40, 1.65 73.6 %, 1,436 65.4 %, 241 60.9 %, 25 45.4 %, 3
10 % 0.82, 3.27 73.7 %, 1,824 67.3 %, 790 61.7 %, 272 45.0 %, 10
20 % 1.62, 7.29 73.1 %, 2,397 68.2 %, 1,622 63.1 %, 984 44.8 %, 154
  • The error of the growth rate is about 0.08 1/s per percent of error in the data, in the median, and four times that at the 95th percentile.
  • A perfect plant model does not help if the network is wrong. With 5 % error and no margin 1,436 unstable cases are cleared, one in seven.
  • Margins have to be several 1/s. With 2 % error a margin of 2 1/s lets through 3 of 10,051. With 5 % it takes 5 1/s to get down to 3. With 10 % only a margin of 10 1/s clears none, and it clears 9 % of cases.

Converters, all modes, 3,960 and 3,343 unstable. Unstable cases cleared at a margin of 2 1/s: 1 and 3 at 2 % error, 28 and 22 at 5 %, 87 and 98 at 10 %. The share cleared stays at 87 % and 71 %.

  • The converters have a tail that no margin removes. In the band between 2 and 58 Hz alone, a few unstable cases are still cleared at a margin of 10 1/s with 2 % error (4 and 14). The wrong network moves the mode across 2 Hz or changes which mode is least damped. Screening on all modes together catches most of them.

Against noise on the scan

Step 20 gave 0.06 1/s per percent of scan noise for the same band. Network data error costs 0.08 1/s per percent. The two are of the same size, percent for percent. Network data are rarely known to better than a few percent, and a scan from a simulation model can be much better than that. So for a scan taken from a model the network data, not the scan, set the margin.

The screens of the doubly-fed plant

Screens under network error

Unstable cases cleared, of 10,051 unstable in the sub-synchronous band for the first two rows and 1,516 in the slow band for the last two.

Screen No error 2 % 5 % 10 % 20 %
Step 14 rule as it stands (margins 1 and 5 1/s), clears 56 % 5 8 34 180 725
Step 14 rule, margins widened by 5 1/s, clears 30 % 0 0 2 28
Phasor-domain model, slow mode, margin 0.05 1/s, clears 95 % 0 3 14 77
Phasor-domain model, slow mode, margin 0.2 1/s, clears 84 % 0 0 1 13

The "no error" entry is from step 18.

  • The step 14 rule already has margins and they absorb 2 % of error. Beyond that it lets cases through in proportion.
  • The phasor-domain screen is the robust one. The slow mode hardly depends on the network, so even with 10 % error and the operating point from the wrong network it lets through 14 of 1,516 at its small margin and 1 at a margin of 0.2 1/s.
  • The pair of step 21, hand formula and phasor-domain model together, lets through 6, 33, 186 and 763 unstable cases of 11,527 at the four error levels, clearing 52 % to 54 %. Most of them are the hand formula's.

What this says about screening

  • How little information, completed. A screen needs the plant (a scan or a formula) and the network. The first can be made nearly exact from a model. The second cannot, and it now sets the accuracy: 0.08 1/s of growth rate per percent.
  • Which network number to pin down depends on the plant. For a machine on a compensated line: the capacitance in service, then the resistance. For a converter: the reactance, which is to say the short-circuit level. Resistance is irrelevant to the converter and second for the machine.
  • The direction of the error is known for the doubly-fed plant. More capacitor, less resistance, less line reactance and a weaker parallel path all push toward instability. A screen run at the unfavourable end of each range would bound the error from one side instead of padding the margin. Not tested here.
  • The band below 2 Hz needs no care about network data. The phasor-domain screen is robust for the same reason it is cheap: the mode it sees depends on the network through one number, the 60 Hz reactance, and weakly.
  • Topology, again. What the network contributes to the resonance is carried by four numbers, and an error in any of them moves the answer by a fraction of 1/s per 10 %. Topology tells where the resonance is; it does not tell the margin to better than the data allow.

Disagreements

  1. The trial on one case in sixteen suggested margins of 2, 5 and 10 1/s were safe at 2 %, 5 % and 10 % error. On all cases they are not: 3, 3 and 0 unstable cases are cleared at those settings on the doubly-fed plant. The tail is longer than a small sample shows.
  2. The converters have unstable cases that are cleared at any margin in one band. Sorting modes into bands at fixed frequencies is fragile when an error moves a mode across the edge. The all-modes screen is the safe one for a converter.
  3. The errors are independent and log-normal. Real errors are structured: one outage changes the parallel path by a large factor and leaves the rest alone. Not studied.
  4. The plant admittance is taken at the true operating point. With a wrong network the power-flow solution is wrong too, and step 17 showed the operating point matters for a converter. Only the phasor-domain screen includes that here.
  5. With 20 % error 40 of 38,511 doubly-fed cases have no steady state on the wrong network. They are counted as not cleared.

Decisions

  • Margins for the sub-synchronous band are to be sized to the larger of the scan error and the network data error, at 0.08 1/s per percent in the median and about four times that for safety.
  • For a converter the screen is on all modes together, not by band.
  • The phasor-domain screen for the slow band keeps its margin of 0.05 1/s up to 5 % network error and takes 0.2 1/s beyond.

Next

In the order they bear on the result: the screen at the unfavourable end of each network range instead of a wider margin; structured network error (one outage); a source and a load on one bus; control delay and a filter capacitor in the converter. Still open from before: the protection model; a fit whose order grows with the data.