A Standardized Framework for Machine Learning in Power System Protection

Julian Oelhaf; Georg Kordowich; Paula Andrea Pérez-Toro; Christian Bergler; Johann Jäger; Andreas Maier; Siming Bayer

cs.LGAug 20, 2026

A Standardized Framework for Machine Learning in Power System Protection

Authors: Julian Oelhaf, Georg Kordowich, Paula Andrea Pérez-Toro, Christian Bergler, Johann Jäger, Andreas Maier, Siming Bayer

Organizations: Pattern Recognition Lab, Friedrich-Alexander-Universität Erlangen-Nürnberg, Martensstr. 3, 91058, Erlangen, Germany · Institute of Electrical Energy Systems, Friedrich-Alexander-Universität Erlangen-Nürnberg, Cauerstr. 4, 91058, Erlangen, Germany · Department of Electrical Engineering, Media and Computer Science, Ostbayerische Technische Hochschule Amberg-Weiden, Kaiser-Wilhelm-Ring 23, 92224, Amberg, Germany

Abstract

Studies of machine-learning-based power-system protection increasingly report near-perfect scores, yet the meaning of those scores depends strongly on the evaluation setting. Protection task, physical scope, measurements, timing, targets, preprocessing, and validation often vary jointly and remain incompletely specified. This paper proposes a standardization-oriented framework that treats evaluation design as part of the scientific contribution. It defines seven required study dimensions: protection objective, physical scope, observability, timing and decision windows, targets and sample validity, validation protocol, and evaluation outputs. The framework is instantiated in a bounded case study on the public PROTECT-90 electromagnetic-transient benchmark, comprising 9022 simulated episodes from a 90 kV double-line topology, for onset-conditioned fault classification and localization. Under centralized sensing, simulation-metadata-aligned 20 ms windows, and episode-grouped validation, a multi-layer perceptron (MLP) achieved a five-fold mean macro-averaged F1 score of 0.991 +/- 0.001 for classification and a localization mean absolute error of 10.20 +/- 0.25% of line length (mean +/- std across episode-grouped folds). Extending the decision horizon to 50 ms preserved this task-dependent performance asymmetry, while reduced observability approximately doubled the MLP localization error but had little effect on classification. A synchronized two-ended conventional locator outperformed the learning locators under its richer clean information set, and measurement degradation showed that clean predictive performance did not determine robustness. The framework turns evaluation assumptions into explicit, reproducible evidence and provides a basis for more comparable, auditable evaluation and future certification-oriented assessment of machine-learning protection functions.

Figures & tables

Source family	Obj.	Scope	Obs.	Time	Targets	Validation	Outputs
Dataset documentation [ 40 ]	$\circ$	$\circ$	$\circ$	–	$\circ$	$\circ$	–
ML reporting and validation guidance [ 39 , 38 , 41 ]	✓	$\circ$	$\circ$	$\circ$	✓	✓	✓
Protection reviews and scoping studies [ 27 , 28 , 30 ]	✓	✓	$\circ$	$\circ$	$\circ$	$\circ$	✓
Protection-testing standards [ 8 , 42 ]	✓	✓	✓	✓	$\circ$	$\circ$	✓
General ai risk and assurance guidance [ 10 ]	✓	$\circ$	$\circ$	–	$\circ$	✓	✓
Proposed framework	✓	✓	✓	✓	✓	✓	✓

Table 1: Comparison of representative guidance against the seven framework dimensions.

Figure 1: Seven-step framework for evaluating machine-learning-based power system protection studies. Step 1 defines the protection objective, Steps 2–5 specify the effective inference setting, Step 6 places the classical training and test pipeline within an explicit validation protocol, and Step 7 structures the resulting evaluation evidence. The final reporting package links conclusions to the complete study definition. The pattern-recognition pipeline is adapted from Niemann [ 47 ] .

Framework step	Minimum reported information
1. Protection objective	Protection task, operational interpretation, learning formulation, and predicted output
2. Physical system scope	Network type, topology class, voltage level, operating range, fault and disturbance space, scenario-generation method, critical-condition coverage, data source, and known representativeness limitations
3. Observability and measurements	Measurement locations, signal domain, channel set, synchronization assumptions, access regime, and auxiliary inputs
4. Timing and decision windows	Event reference, pre-/post-event context, decision horizon, window length, stride, and overlap rule
5. Targets and sample validity	Label or regression-target construction, inclusion/exclusion rules, boundary handling, and retained sample set
6. Training and validation protocol	Split design, unit of independence, leakage-prevention strategy, preprocessing boundary, and model-comparison conditions

Table 2: Minimum reporting package for ml -based power system protection studies under the proposed framework.

Framework step	Reference instantiation
1. Protection objective	fc and fl under shared sensing, timing, and validation assumptions
2. Physical system scope	PROTECT-90; 90 kV double-line topology; 9022 emt episodes; randomized fault and operating conditions
3. Observability	Synchronized three-phase voltage/current measurements from 8 relays; centralized full-observability reference setting
4. Timing	$\pm$ 80 ms around fault inception; 20 ms sliding windows; fixed 5 ms stride
5. Targets	Onset-conditioned 11-class fc with one non-onset class and ten fault-type classes; fl as normalized fault position; fault-onset-containing windows for fl
6. Validation	5-fold grouped cross-validation by simulation episode; shared preprocessing and split design across tasks and models

Table 3: Case-study instantiation of the proposed standardized evaluation framework.

Panel of Figure 2 — Figure 2: Case-study data generation flow for the PROTECT-90 dataset [ 54 ] . A fixed grid topology is combined with randomized fault and operating scenarios, simulated in the emt domain, and processed into synchronized waveform episodes with labels and metadata.

Task	Metric	mlp	gb	knn	Ridge
fc	Macro- f1 $\uparrow$	0.991 $\pm$ 0.001	0.745 $\pm$ 0.017	0.799 $\pm$ 0.006	0.095 $\pm$ 0.003
fl	mae [%] $\downarrow$	10.20 $\pm$ 0.25	14.78 $\pm$ 0.25	20.10 $\pm$ 0.16	26.12 $\pm$ 0.36

Table 4: Reference performance across representative case-study models for the fixed 20 ms timing configuration. fc is reported by macro- f1 ; fl by mae in percent of normalized line length. Values are mean

\pm

std over 5 folds.

Window	mlp	gb	knn	ridge
10 ms	0.985 $\pm$ 0.011	0.418 $\pm$ 0.009	0.738 $\pm$ 0.012	0.097 $\pm$ 0.002
20 ms (ref.)	0.991 $\pm$ 0.001	0.745 $\pm$ 0.017	0.799 $\pm$ 0.006	0.095 $\pm$ 0.003
30 ms	0.988 $\pm$ 0.005	0.977 $\pm$ 0.001	0.831 $\pm$ 0.004	0.093 $\pm$ 0.002
40 ms	0.988 $\pm$ 0.007	0.982 $\pm$ 0.002	0.851 $\pm$ 0.003	0.091 $\pm$ 0.002
50 ms	0.990 $\pm$ 0.005	0.982 $\pm$ 0.001	0.863 $\pm$ 0.003	0.089 $\pm$ 0.002

Table 5: Controlled timing sensitivity for fault classification. Macro- f1 across decision horizons for the representative case-study models. Mean

\pm

std over 5 folds; higher is better. The 20 ms row corresponds to the fixed reference timing configuration.

Window	mlp	gb	knn	ridge
10 ms	10.64 $\pm$ 0.36	14.65 $\pm$ 0.19	20.38 $\pm$ 0.32	26.14 $\pm$ 0.32
20 ms (ref.)	10.20 $\pm$ 0.25	14.78 $\pm$ 0.25	20.10 $\pm$ 0.16	26.12 $\pm$ 0.36
30 ms	10.18 $\pm$ 0.35	14.64 $\pm$ 0.18	19.65 $\pm$ 0.19	26.11 $\pm$ 0.35
40 ms	10.46 $\pm$ 0.52	14.59 $\pm$ 0.19	19.33 $\pm$ 0.16	26.12 $\pm$ 0.32
50 ms	9.92 $\pm$ 0.27	14.66 $\pm$ 0.20	19.12 $\pm$ 0.14	26.13 $\pm$ 0.31

Table 6: Timing sensitivity for fl . mae is reported as percent of normalized line length across decision horizons. Values are mean

\pm

std over 5 folds; lower is better. The 20 ms setting is the reference configuration.

Observability	mlp (20 ms)	gb (20 ms)	mlp (50 ms)	gb (50 ms)
Full	0.991 $\pm$ 0.001	0.745 $\pm$ 0.017	0.990 $\pm$ 0.005	0.982 $\pm$ 0.001
Relay pair	0.989 (0.988–0.990)	0.750 (0.717–0.777)	0.987 (0.987–0.989)	0.971 (0.968–0.975)
Single relay	0.986 (0.975–0.993)	0.734 (0.629–0.811)	0.986 (0.966–0.991)	0.952 (0.900–0.973)

Table 7: Controlled observability sensitivity for fault classification. Macro- f1 under full, relay-pair, and single-relay sensing for mlp and histogram-based gb models at 20 ms and 50 ms. Full-observability values are the mean

\pm

standard deviation over the five episode-grouped folds; the relay-pair and single-relay values are the mean over the four same-line relay pairs and the eight individual relays, respectively, with the min–max range across those configurations in brackets. Higher is better.

Observability	mlp (20 ms)	gb (20 ms)	mlp (50 ms)	gb (50 ms)
Full	10.20 $\pm$ 0.25	14.78 $\pm$ 0.25	9.92 $\pm$ 0.27	14.66 $\pm$ 0.20
Relay pair	19.65 (17.40–21.67)	21.43 (20.30–22.42)	19.62 (17.23–21.63)	21.37 (20.12–22.38)
Single relay	22.22 (20.17–23.69)	23.20 (22.46–23.94)	22.17 (20.05–23.62)	23.07 (22.19–23.77)

Table 8: Controlled observability sensitivity for fault localization. mae in percent of normalized line length under full, relay-pair, and single-relay sensing for mlp and histogram-based gb models at 20 ms and 50 ms. Full-observability values are the mean

\pm

Method	Observability	20 ms	50 ms
mlp	full	10.20 $\pm$ 0.25	9.92 $\pm$ 0.27
gb	full	14.78 $\pm$ 0.25	14.66 $\pm$ 0.20
knn	full	20.10 $\pm$ 0.16	19.12 $\pm$ 0.14
Ridge	full	26.12 $\pm$ 0.36	26.13 $\pm$ 0.31
Two-ended (per window)	relay pair	5.42 $\pm$ 0.09	3.62 $\pm$ 0.23
Two-ended (settled)	relay pair	2.74 $\pm$ 0.05	1.57 $\pm$ 0.03

Table 9: fl performance of learning and conventional impedance-based methods under shared episode partitions and task-specific metrics. mae [% line length], mean

\pm

standard deviation over five folds. Per-window conventional results use the onset-conditioned evaluation windows, whereas settled results use a separately reported per-episode post-onset validity rule. Conventional locators additionally receive per-episode line parameters and ground-truth loop selection.

Task (metric)	Model	Clean	Worst case
fl ( mae [%], $\downarrow$ )	mlp	10.20	24.36
fl ( mae [%], $\downarrow$ )	gb	14.78	20.56
fc (macro- f1 , $\uparrow$ )	mlp	0.991	0.959
fc (macro- f1 , $\uparrow$ )	gb	0.745	0.274

Table 10: Compact measurement-fidelity summary at 20 ms for the focused sensitivity panel: clean value and worst-case level across the three degradation axes (additive noise, current-transformer saturation, synchronization jitter). Full per-level results in Appendix B (Tables B.1 and B.2 ).

Figure 3: Row-normalized confusion matrix for fc on fault-only cases under the 20 ms mlp configuration. Rows denote the true fault class and columns the predicted fault class. Diagonal entries indicate per-class recall, while off-diagonal cells represent misclassifications.

Figure 4: fl error as a function of the true fault position for the reference 20 ms mlp configuration (out-of-fold evaluation). Results are aggregated in 10% line-length bins; the solid curve shows the median absolute error, and the shaded band spans the median to the 95th percentile.

Model	Window	f1	Train [s]	Infer. [ $\mu$ s]	Thr. [k/s]
mlp	20 ms	0.991	603.7	7.32	136.7
gb	20 ms	0.745	132.6	24.47	41.1
mlp	50 ms	0.990	1289.5	16.37	61.1
gb	50 ms	0.982	1968.8	82.91	12.1

Table 11: Compact runtime summary for fault classification under the reference assumptions. Macro- f1 is shown for context.

Model	Window	mae	Train [s]	Infer. [ $\mu$ s]	Thr. [k/s]
mlp	20 ms	10.20	346.3	6.45	155.0
gb	20 ms	14.78	27.2	22.49	44.5
mlp	50 ms	9.92	2258.0	15.91	62.9
gb	50 ms	14.66	140.4	49.36	20.3

Table 12: Compact runtime summary for fault localization under the reference assumptions. mae [%] is shown for context.

Model	fc 20 ms	fc 50 ms	fl 20 ms	fl 50 ms
mlp (reference)	0.991 $\pm$ 0.001	0.990 $\pm$ 0.005	10.20 $\pm$ 0.25	0 9.92 $\pm$ 0.27
cnn1d (from scratch)	0.999 $\pm$ 0.001	0.999 $\pm$ 0.001	0 9.65 $\pm$ 0.67	0 9.09 $\pm$ 0.28

Table 13: Main-text learned-baseline comparison under the shared episode-grouped five-fold protocol. fc is evaluated using macro- f1 (

\uparrow

), and fl using mae in percent of line length (

\downarrow

). Values are mean

\pm

standard deviation across the five held-out folds; the best mean in each column is shown in bold. The pre-trained MOMENT-1-large results are reported separately in Appendix E (Table E.1 ).

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Line	Single-terminal mae [%]		Two-terminal	Improvement
	Terminal 1	Terminal 2	mae [%]	[% points]
01–02A	21.948	20.055	17.229	2.826
01–02B	22.316	22.712	20.814	1.502
02–03A	20.510	23.202	18.797	1.712
02–03B	22.997	23.620	21.630	1.367

Appendix

Table A.1: Relay-level fl error for the 50 ms mlp regressor under reduced observability. The improvement is computed relative to the better of the two single-terminal settings for each line.

Observability	Conventional	Conv.	mlp	gb
Full	(n/a)	–	10.20	14.78
Relay pair	two-ended synchr.	2.74	19.65	21.43
Single relay	one-ended reactance	41.9	22.22	23.20

Appendix

Table A.2: fl performance by observability at 20 ms for sensing-matched conventional and learning methods. Conventional values use the separately reported per-episode settled estimates and additionally receive episode-specific line parameters and ground-truth loop selection; the comparison is therefore sensing-matched but not input- or sample-validity-matched. mae [% line length]; lower is better.

Axis	Level	mlp	gb
Clean	–	10.20 $\pm$ 0.25	14.78 $\pm$ 0.25
Noise	20 dB	10.48 $\pm$ 0.26	14.91 $\pm$ 0.24
Noise	10 dB	12.55 $\pm$ 0.35	15.75 $\pm$ 0.22
ct saturation	$c=0.5$	15.91 $\pm$ 0.24	17.44 $\pm$ 0.35
ct saturation	$c=0.3$	24.36 $\pm$ 0.67	20.56 $\pm$ 0.37
Jitter	2 samples	12.30 $\pm$ 1.35	15.05 $\pm$ 0.26

Appendix

Table B.1: fl robustness to measurement degradation at 20 ms. The clean row reproduces the five-fold reference statistic. For each perturbed level, the five realizations are first averaged within each fold; values are then reported as mean

\pm

standard deviation across the five fold-level means. Lower is better.

Axis	Level	mlp	gb
Clean	–	0.991 $\pm$ 0.001	0.745 $\pm$ 0.017
Noise	20 dB	0.989 $\pm$ 0.003	0.592 $\pm$ 0.024
Noise	10 dB	0.982 $\pm$ 0.011	0.274 $\pm$ 0.028
ct saturation	$c=0.5$	0.982 $\pm$ 0.002	0.682 $\pm$ 0.021
ct saturation	$c=0.3$	0.959 $\pm$ 0.003	0.534 $\pm$ 0.021
Jitter	4 samples	0.983 $\pm$ 0.006	0.682 $\pm$ 0.034

Appendix

Table B.2: fc robustness to measurement degradation at 20 ms. The clean row reproduces the five-fold reference statistic. For each perturbed level, the five realizations are first averaged within each fold; values are then reported as mean

\pm

standard deviation across the five fold-level means. Higher is better.

Task	Model	Five-fold reference	High- $R_{f}$ test	Low- $R_{f}$ test
fc (macro- f1 )	mlp	0.991	$0.975\pm 0.004$	$0.979\pm 0.007$
fc (macro- f1 )	gb	0.745	0.731	0.725
fl ( mae [%])	mlp	10.20	$10.33\pm 0.22$	$12.34\pm 0.63$
fl ( mae [%])	gb	14.78	16.79	14.30

Appendix

Table C.1: Directional fault-resistance holdout at 20 ms. Models are trained on the lower 80% and tested on the upper quintile of

R_{f}

, or vice versa. fc : macro- f1 ; fl : mae [% line length]. The episode-grouped five-fold result is shown for context and is not a training-size-matched control. Shifted mlp columns are mean

\pm

standard deviation across five explicitly varied model seeds (0–4); the histogram-based gb columns report one run with seed 42.

Class	Type	Windows	Share
Non-onset	–	234 572	89.66%
AG / BG / CG	1ph–g	2295 / 2259 / 2283	0.88 / 0.86 / 0.87%
AB / BC / CA	2ph	2175 / 2262 / 2316	0.83 / 0.86 / 0.89%
ABG / BCG / CAG	2ph–g	2175 / 2205 / 2277	0.83 / 0.84 / 0.87%
ABC	3ph	6819	2.61%
Total		261 638	100%

Appendix

Table D.1: Class distribution of the 20 ms fault-classification windows ( event_type ;

N=261{,}638

windows over

9{,}022

episodes).

Model	fc 20 ms	fc 50 ms	fl 20 ms	fl 50 ms
mlp (reference)	0.991 $\pm$ 0.001	0.990 $\pm$ 0.005	10.20 $\pm$ 0.25	0 9.92 $\pm$ 0.27
cnn1d (from scratch)	0.999 $\pm$ 0.001	0.999 $\pm$ 0.001	0 9.65 $\pm$ 0.67	0 9.09 $\pm$ 0.28
MOMENT – linear probe	0.953 $\pm$ 0.005	0.975 $\pm$ 0.002	14.82 $\pm$ 0.10	14.99 $\pm$ 0.09
MOMENT – mlp head	0.985 $\pm$ 0.002	0.989 $\pm$ 0.001	10.59 $\pm$ 0.22	0 8.90 $\pm$ 0.12

Appendix

Table E.1: Pre-trained foundation-model (MOMENT-1-large) baseline under the shared episode-grouped five-fold protocol, with the main-text mlp and cnn1d repeated for reference. fc : macro- f1 (

\uparrow

); fl : mae [% line length] (

\downarrow

). Mean

\pm

standard deviation over 5 folds; best mean per column in bold.

Model	Window	Baseline f1	$\Delta$ @ 10 ms	$\Delta$ @ 20 ms	$\Delta$ @ 50 ms
mlp	10 ms	0.985	+0.003	–	–
mlp	20 ms	0.991	-0.008	-0.010	–
mlp	50 ms	0.990	-0.004	-0.002	-0.003
gb	10 ms	0.418	+0.100	–	–
gb	20 ms	0.745	+0.221	+0.225	–
gb	50 ms	0.982	-0.005	-0.003	+0.004

Appendix

Table F.1: Compact stride sensitivity for fault classification. Baseline performance is reported at the default 5 ms stride; additional columns show

\Delta

macro- f1 relative to that baseline for larger strides. Positive values indicate improvement.

Model	Window	Baseline mae	$\Delta$ @ 10 ms	$\Delta$ @ 20 ms	$\Delta$ @ 50 ms
mlp	10 ms	10.64	+0.00	–	–
mlp	20 ms	10.20	+0.93	+0.13	–
mlp	50 ms	9.92	+1.10	+1.24	+1.44
gb	10 ms	14.65	+0.00	–	–
gb	20 ms	14.78	+0.28	-0.36	–
gb	50 ms	14.66	+0.49	+0.65	-0.20

Appendix

Table F.2: Compact stride sensitivity for fault localization. Baseline performance is reported at the default 5 ms stride; additional columns show

\Delta

mae relative to that baseline for larger strides. Negative values indicate improvement.

Model	Fixed configuration
Ridge	Regularization strength $\alpha=1.0$ ; automatic solver selection; pseudorandom seed 42. RidgeClassifier is used for fc and Ridge for fl .
knn	Five neighbors; uniform weighting; Minkowski distance with $p=2$ ; automatic neighbor-search algorithm; parallel prediction enabled.
Histogram-based gb	100 boosting iterations; learning rate 0.1; unrestricted tree depth; minimum 20 samples per leaf; no L2 regularization; pseudorandom seed 42.
mlp	One hidden layer with 100 rectified-linear units; Adam optimizer; L2 penalty $10^{-4}$ ; initial learning rate $10^{-3}$ ; automatic batch size; maximum 200 iterations; no early stopping; pseudorandom seed 42.

Appendix

Table G.1: Fixed configurations of the classical learning models used in the reference evaluation. Parameters not listed retain the scikit-learn 1.8.0 defaults. The same settings are used for the corresponding classifier and regressor unless stated otherwise.

Model	Hyperparameter	Default	Tested values
gb	Learning rate	0.1	0.03, 0.2
gb	Maximum tree depth	unlimited	3, 5, 10
gb	Boosting iterations	100	50, 300
gb	Minimum samples per leaf	20	5, 50
gb	L2 regularization	0	$10^{-4}$ , $10^{-2}$
mlp	Hidden layers	(100)	(50), (100, 50), (256, 128)

Appendix

Table G.2: Single-parameter hyperparameter ablation design.

Task	Model	Window	Default	Best metric	$\Delta_{\text{best}}$	Max. spread
fc	gb	20 ms	0.745	0.967	+0.222 (min leaf)	0.327 (learn. rate)
fc	gb	50 ms	0.982	0.988	+0.006 (iters)	0.116 (learn. rate)
fc	mlp	20 ms	0.990	0.996	+0.006 (init. lr)	0.015 (init. lr)
fc	mlp	50 ms	0.991	0.994	+0.003 (init. lr)	0.011 (init. lr)
fl	gb	20 ms	14.782	13.232	+1.550 (iters)	5.530 (depth)
fl	gb	50 ms	14.662	12.942	+1.720 (iters)	5.671 (depth)

Appendix

Table G.3: Condensed hyperparameter robustness summary across single-parameter ablations. For fc ,

\Delta_{\text{best}}

denotes the increase in macro- f1 ; for fl , it denotes the reduction in mae . Positive values therefore indicate improvement in both tasks. The maximum spread reports the largest best–worst difference within any single-parameter ablation group. The default cells are independent results from the ablation harness and can differ slightly from the Table 4 reference; for example, fl mlp at 20 ms is

9.998

here versus

10.20

in the reference evaluation. Comparisons within each ablation group therefore use the corresponding default generated by the same harness. This difference is not a repeated-seed uncertainty estimate.

Model	Task	Ablation default	Best	$\Delta_{\text{best}}$	Max. spread
mlp	fc	0.991	0.996	+0.005	0.015
gb	fc	0.982	0.988	+0.006	0.327
mlp	fl	9.955	8.296	+1.659	6.209
gb	fl	14.662	12.942	+1.720	5.671

Appendix

Table G.4: Compact hyperparameter robustness summary for the representative mlp and histogram-based gb models based on the single-parameter ablations reported in this appendix. Ablation default refers to the best tagged default result within the independent ablation campaign across the reported decision horizons. Best denotes the strongest result observed in the corresponding ablation setting. For fc ,

\Delta_{\text{best}}

denotes the increase in macro- f1 ; for fl , it denotes the reduction in mae . Positive values therefore indicate improvement in both tasks.

Explore similar work

Sep 23, 2026

EvEMTBench: An Open Benchmark for Machine Learning in Power System Protection

Studies of machine-learning-based power system protection are difficult to compare because task definitions, measurement access, data partitions, metrics, and generalization conditions often differ. EvEMTBench addresses this gap with an open, executable, and versioned benchmark that fixes these evaluation choices while leaving model design open. Across four grids spanning 20-345 kV, it defines 12 protection and event-analysis functions instantiated as 24 scored tasks and supports structured evaluation across observability conditions, predefined distribution shifts, and zero-shot and fine-tuned cross-grid transfer. Committed partitions, leakage controls, and reproducible reporting provide a common basis for comparing future methods. A reference evaluation spanning trivial, conventional, feature-based, and deep-learning baselines shows that wider observability is not uniformly beneficial, shifted conditions can reveal failures not apparent in-distribution, and cross-grid transfer is substantially stronger for fault detection than for fault localization. Protection-relevant diagnostics identify failure modes not apparent from primary metrics alone. EvEMTBench therefore makes generalization in machine-learning-based protection an explicit and reproducible evaluation problem.

Julian Oelhaf, Georg Kordowich, Christian Bergler +3

Pattern Recognition Lab, Friedrich-Alexander-Universität Erlangen-Nürnberg, Erlangen 91058, Germany · Institute of Electrical Energy Systems, Friedrich-Alexander-Universität Erlangen-Nürnberg, Erlangen 91058, Germany · Department of Electrical Engineering, Media and Computer Science, Ostbayerische Technische Hochschule Amberg-Weiden, Amberg 92224, Germany

Domain Generalization Fault Detection Fault Localization Power Systems

Ask PDF

Oct 2, 2026

One-Cycle Fault Classification and Faulted-Line Identification on the PROTECT-90 Dataset: An Initial Application Benchmark

Open electromagnetic-transient datasets are beginning to make reproducible learning-based protection studies possible, but the practical use of these datasets still requires application-level benchmarks that define timing, sensing, and validation assumptions. This paper presents an initial application benchmark on the recently released PROTECT-90 dataset for two protection-oriented tasks: fault-type classification and discrete faulted-line identification. A compact one-dimensional convolutional neural network (CNN) is evaluated using post-inception windows of 0.25, 0.5, 1, and 2 cycles under strict episode-wise splitting. A non-convolutional multilayer perceptron (MLP) is also trained as an architecture-control baseline. The results show that both tasks are nearly saturated under full observability, with one-cycle test accuracies of 99.84% for fault type and 100.00% for line identification. The main performance variation appears under reduced observability: current-only inputs preserve line identification accuracy at 100.00%, whereas voltage-only inputs reduce line identification accuracy to 53.09% with the CNN and 50.57% with the MLP. This indicates that the limiting factor is measurement information rather than neural architecture. Additional stratified checks show stable performance across topology states and fault-resistance bins, while CPU inference contributes only 0.528 ms to the one-cycle total decision time of 20.53 ms.

Emad Abukhousa, Abdulaziz Qwbaiban, Saman Zonouz +1

School of Electrical and Computer Engineering, Georgia Institute of Technology, Atlanta, GA, USA · Electrical Engineering Department, King Fahd University of Petroleum and Minerals, Dhahran, Saudi Arabia · Interdisciplinary Research Center for Sustainable Energy Systems (IRC-SES), KFUPM, Dhahran 31261, Saudi Arabia

Fault Diagnosis Power Systems Convolutional Neural Networks

Ask PDF

Jun 23, 2026

PROTECT-90: A Fault Dataset for Power System Protection

The increasing interest in data-driven methods for power system protection is accompanied by a lack of standardized, publicly available high-voltage waveform datasets that enable transparent and reproducible evaluation. To address this gap, this paper introduces the PROTECT-90 dataset, an open electromagnetic transient (EMT)-simulated reference benchmark for high-voltage fault studies with consistent digital-fault-recorder-like measurements, publicly released with this work. The dataset comprises 9,022 physically consistent short-circuit simulation episodes generated on a standardized 90 kV double-line topology with systematically documented domain randomization of grid operating points, line parameters, and fault conditions. For each episode, synchronized three-phase voltage and current waveforms are recorded at eight measurement locations and released together with structured, machine-readable metadata describing fault type, fault location, inception time, and operating conditions. All modeling assumptions, parameter ranges, and data-generation procedures are explicitly documented to ensure transparency and cross-study comparability. By combining physically grounded EMT simulation, balanced scenario coverage, and open accessibility, PROTECT-90 establishes a standardized foundation for reproducible benchmarking of protection-oriented signal processing and learning-based methods.

Julian Oelhaf, Georg Kordowich, Christian Bergler +3

1Pattern Recognition Lab, Friedrich-Alexander-Universität Erlangen-Nürnberg, Erlangen, Germany · Institute of Electrical Energy Systems, Friedrich-Alexander-Universität Erlangen-Nürnberg, Erlangen, Germany · Department of Electrical Engineering, Media and Computer Science, Ostbayerische Technische Hochschule Amberg-Weiden, Amberg, Germany

Benchmark Design Synthetic Benchmark Generation Power Systems

Ask PDF