arXiv CSopen access

PROTECT-90: A Fault Dataset for Power System Protection

PROTECT-90: A Fault Dataset for Power System Protection Julian Oelhaf1 † * , Georg Kordowich2 † , Christian Bergler3 , Andreas Maier1 , Johann Jäger2 , Siming Bayer1 1

2

Pattern Recognition Lab, Friedrich-Alexander-Universität Erlangen-Nürnberg, Erlangen, Germany Institute of Electrical Energy Systems, Friedrich-Alexander-Universität Erlangen-Nürnberg, Erlangen, Germany 3 Department of Electrical Engineering, Media and Computer Science, Ostbayerische Technische Hochschule Amberg-Weiden, Amberg, Germany † These authors contributed equally.

arXiv:2606.24298v1 [eess.SP] 23 Jun 2026

* Corresponding author: [email protected]

Abstract—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 digitalfault-recorder-like measurements, publicly released with this work. The dataset comprises 9,022 physically consistent shortcircuit 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, machinereadable 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, PROTECT90 establishes a standardized foundation for reproducible benchmarking of protection-oriented signal processing and learningbased methods. Index Terms—Power system protection, fault dataset, electromagnetic transient (EMT) simulation, high-voltage systems, reproducible benchmarking

I. I NTRODUCTION Reliable high-voltage-level protection relies on accurate interpretation of voltage and current waveforms during shortcircuit events. Electromagnetic transient (EMT) simulation is widely used to analyze fault inception and fast transient phenomena not captured by steady-state phasor-domain models. As high-voltage systems operate under increasingly diverse conditions, including varying load levels, short-circuit strengths, voltage setpoints, and line characteristics, protection validation requires waveform data that reflect realistic operating variability. Despite the central role of EMT simulation in protection engineering, openly accessible waveform © 2026 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works.

datasets with clearly documented topology definitions, parameter ranges, and labeling conventions remain scarce. Many studies rely on proprietary grid models or task-specific data generation procedures that are not publicly released, restricting reproducibility and cross-study comparability. Standardized test systems such as the IEEE 9-bus and IEEE 39-bus networks are widely used for steady-state and phasordomain validation [1], [2]. For waveform-based analysis, publicly available resources include EMT-simulated datasets such as the Inverter-Rich Transmission System Disturbance (IRTSD) dataset [3], as well as measured waveform repositories including the SoCal 28-Bus distribution grid dataset (SoCal) [4], the RTE digital fault recording database [5], and the Grid Event Signature Library (GESL) [6]. As summarized in Table I, these resources provide either synthetic events tailored to specific grid models or real-world recordings without explicitly documented and controllable system parameter bounds, limiting their suitability for transparent and reproducible EMTbased benchmarking. Taken together, existing public resources either provide controlled EMT simulations without releasing complete waveform corpora and parameter metadata, or offer measured recordings without structured and fully documented simulation assumptions. As reported in a recent scoping review of machine learning applications in power system protection [7], only 3.4 % of surveyed studies released their datasets or relied on publicly accessible data, with most based on proprietary simulations or unpublished grid models. Moreover, topology definitions, parameter ranges, and data generation procedures are often only partially documented, limiting independent reproduction of reported results. To the best of our knowledge, no openly available high-voltage level EMT waveform dataset simultaneously provides raw instantaneous measurements, complete machine-readable metadata, and explicitly documented parameter bounds for systematic and reproducible evaluation. This lack of transparent and standardized waveform corpora remains a structural barrier to cross-study comparability. This paper addresses this gap by introducing PROTECT90, a publicly released 90 kV EMT-simulated high-voltage level waveform dataset comprising 9,022 physically consistent simulation episodes. The dataset provides raw instantaneous

Bus 3

Bus 2 Line 2–3 A

Bus 1 Line 1–2 A

Ext. Grid 3

Ext. Grid 1

Line 2–3 B

Load 3

Line 1–2 B

Load 2

Protection relay location

Fig. 1. Standardized 90 kV high-voltage double-line topology used for EMT dataset generation. Filled squares denote protection relay measurement locations.

three-phase voltage and current measurements together with structured, machine-readable metadata describing fault characteristics, grid parameters, and topology states. A central design principle is the use of physically constrained domain randomization, whereby operating conditions are systematically varied within explicitly defined and physically plausible bounds to enable controlled yet diverse scenario generation. The main contributions of this work are: • The release of PROTECT-90, a publicly accessible high-voltage EMT waveform dataset with 9,022 labeled fault scenarios and structured metadata available at 10.5281/zenodo.18418330. • A physically constrained domain-randomization framework with explicitly documented parameter bounds. • Full documentation of modeling assumptions and datageneration procedures under an open license, enabling transparent and reproducible benchmarking. II. D ESIGN R ATIONALE AND S COPE The PROTECT-90 dataset is designed as a standardized high-voltage EMT reference case rather than a large-scale meshed network model. A nominal voltage level of 90 kV is selected to represent a widely deployed sub-transmission class while maintaining alignment with practical high-voltage digital fault recording conditions. This choice enables realistic transient behavior without restricting the dataset to extra-highvoltage or inverter-dominated system configurations. The adopted double-line topology reflects common highvoltage corridor structures featuring parallel current paths and intermediate infeeds [8]. These characteristics introduce nontrivial current distribution during short-circuit events while remaining sufficiently compact for controlled parameter variation and full documentation. The topology is intentionally fixed to provide a reproducible and interpretable reference scenario, enabling systematic comparison across studies. The scope of the dataset is limited to EMT simulations under physically constrained domain randomization of grid and fault parameters. It does not aim to represent large meshed networks or specific protection schemes but instead provides a

transparent and standardized waveform corpus suitable for reproducible transient analysis and protection-oriented research. III. S IMULATION F RAMEWORK This section describes the electrical modeling assumptions, EMT simulation configuration, and measurement structure used to generate the dataset. A. Benchmark Topology The benchmark topology, shown in Fig. 1, is a standardized 90 kV high-voltage double-line network comprising three buses (Bus 1–3) connected by two parallel corridors between adjacent bus pairs. Line 2–3 A/B connect Bus 2 and Bus 3, and Line 1–2 A/B connect Bus 1 and Bus 2, forming parallel current paths commonly encountered in practice [8]. Such configurations introduce non-trivial current distribution and impedance measurement errors relevant for distance protection studies. Individual line lengths and electrical parameters are sampled within predefined bounds (Table II). The selected parameter ranges are chosen to represent realistic high-voltage conditions while covering a broad variability spectrum. The bounds are derived from established literature on short-circuit current calculation and high-voltage system modeling [9], [10]. Bus 1 and Bus 3 are connected to external grid equivalents, and aggregated loads are connected at Bus 2 and Bus 3, resulting in multiple power-flow paths during fault conditions. To introduce structural variability, the secondary corridor segments (Line 1–2 B and Line 2–3 B) are switched out in a subset of scenarios, allowing different configurations and their respective impact on power system protection. B. EMT Simulation Configuration and Measurement Model All simulations are performed in EMT mode using DIgSILENT PowerFactory1 . For each scenario, the system is first simulated for an internal warm-up period of 0.1 s to allow the differential equation system to settle to steady-state conditions. 1 https://www.digsilent.de/en/powerfactory.html

TABLE I P OSITIONING OF PUBLICLY AVAILABLE WAVEFORM DATASETS (IRTSD [3], S O C AL [4], RTE [5], GESL [6]). Feature EMT-simulated Measured waveforms High-voltage focus Explicit parameter bounds documented

TABLE II PARAMETER RANGES USED FOR EMT DOMAIN RANDOMIZATION .

PROTECT-90

IRTSD

SoCal

RTE

GESL

✓ –

✓ –

– ✓

– ✓

– ✓

✓ ✓

✓ –

– –

✓ –

– –

Notes: SoCal, RTE, and GESL provide real-world waveform measurements, while IRTSD and the proposed dataset are EMT-simulated. The proposed dataset is designed as a high-voltage level benchmark with explicitly documented parameter bounds.

The subsequent 1 s analysis window is then recorded and exported, sampled at 6,400 Hz, resulting in 6,400 time steps per episode. The 6.4 kHz sampling rate corresponds to 128 samples per 50 Hz fundamental cycle, representing a typical digital fault recording resolution in high-voltage protection while enabling high-resolution transient analysis with manageable data size. A fixed-step EMT solver with a step size of 10 µs is used to ensure consistent and accurate results across all simulations. Three-phase voltages and currents are recorded at both terminals of each line segment, yielding six measurement channels per relay location and a total of 48 synchronized waveform channels. To ensure consistent signal availability across all scenarios, voltage and current signals are extracted directly at the corresponding bus terminals within the simulation model. No secondary instrument transformer models (CTs/VTs) are considered; all exported quantities represent ideal primary voltage and current signals. IV. S CENARIO G ENERATION AND AUTOMATION The automation concept underlying the scenario generation process extends a previously published generic data generation framework for short-circuit studies in PowerFactory [11]. While the present implementation was redesigned and developed independently for the structured PROTECT-90 benchmark, the earlier work established the principle of automated fault injection, parameter randomization, and parallelized simulation execution. Each simulation scenario (episode) contains a single short-circuit event. Faults are defined by the affected line segment, fault type (SLG, LL, LLG, LLL), and involved phase(s). The fault resistance Rf and inception time tf are selected within the ranges specified in Table II. Fault locations are uniformly distributed along the normalized line length, ensuring spatial coverage across the full extent ℓ. To represent operating variability, grid parameters are varied according to a domain randomization strategy. The objective is to generate a broad spectrum of physically plausible operating conditions, thereby increasing waveform diversity and improving the robustness of data-driven methods with respect to domain shift from simulation to real-world measurements. The electrical parameters listed in Table II are independently sampled within their specified bounds prior to each simulation. These include the line length ℓ, series resistance R′ , series

Parameter

Unit

Min

Max

Fault Parameters Fault resistance Rf Fault inception time tf

Ω s

0.1 0.2

10 0.5

Line Parameters Line length ℓ Series resistance per km R′ Series reactance per km X ′ Shunt capacitance per km C ′

km Ω/km Ω/km nF/km

10 0.01 0.35 8.5

60 0.20 0.45 10

Load Parameters Active power P Reactive power Q

MW Mvar

20 −20

50 20

External Grid Parameters Short-circuit power Sk′′ Voltage magnitude V Voltage angle ϕ

MVA pu deg

90 0.95 −180

1000 1.05 180

reactance X ′ , and shunt capacitance C ′ per kilometer, as well as the load active and reactive power (P, Q). External grid parameters – short-circuit power Sk′′ , voltage magnitude V , and voltage angle ϕ – are likewise varied within their respective bounds. The selected short-circuit power range is intentionally biased toward comparatively low values to reflect future grid conditions with increasing penetration of inverter-based resources and consequently reduced fault levels. The voltage angle ϕ is randomized to prevent systematic phase alignment across scenarios and avoid artificial waveform synchronization. In addition, the external grid at Bus 3 is deactivated in a subset of cases to emulate reduced infeed or topology reconfiguration. To ensure physical realism, all parameter combinations must satisfy load-flow convergence and EMT numerical stability before, during, and after the fault event. Furthermore, the R/X ratio of all lines is constrained to the interval [0.05, 0.5], reflecting typical high-voltage line characteristics. This physically constrained randomization preserves electrical plausibility while maintaining a wide variability spectrum representative of real-world high-voltage system operation. V. PROTECT-90: DATASET S TRUCTURE AND C OVERAGE Each simulation episode is stored as an individual .pkl file containing a time-series dataframe with 49 columns. The first column corresponds to the time vector, while the remaining 48 columns represent synchronized three-phase voltage and current measurements from eight relay locations (six channels per relay). Signal naming follows a consistent convention combining bus, line section, quantity, phase, and unit, e.g., Bus_3_Line_02_03A_vol_L3_V, denoting the phase 3 voltage at the Bus 3 terminal of line section 2–3A. Currents are labeled analogously using the suffix _cur_Lx_A. All quantities correspond to secondary voltage (V) and current (A) signals sampled at 6,400 Hz.

Voltage (kV)

A. Coverage and Dataset Characteristics 50 0

Current (kA)

−50

5

A B C

0

−5 0.4

0.5

0.6

0.7

0.8

Time (s) Fig. 2. Representative three-phase voltage (top) and current (bottom) waveforms for a two-phase short-circuit between phases B and C on Line 1–2 B. 1 The fault is applied at 31.8 % of the line length with a resistance of 6.56 Ω. Dashed vertical lines indicate fault inception (t = 0.458 s) and the end of the injected fault (fault element removed at t = 0.688 s). TABLE III OVERVIEW AND KEY CHARACTERISTICS OF THE PROTECT-90 DATASET. Property

Specification

Topology Simulation type Sampling rate Episode duration Episodes Location Domain Randomization Metadata Train/test split License DOI

90 kV double-line system Electromagnetic transient (EMT) 6.4 kHz (128 samples per 50 Hz cycle) 1 s (6,400 samples) 9,022 fault scenarios Normalized line length Line, load, grid, and fault parameters (Table II) CSV (fault and grid parameters) None (episode-wise partitioning recommended) CC BY 4.0 10.5281/zenodo.18418330

At the episode level, the dataset can be formally represented as X ∈ RNPR ×6×Ts , where NPR = 8 denotes the number of relay locations, six channels correspond to three-phase voltage and current measurements per relay, and Ts = 6,400 is the number of time steps per episode. A representative example of the recorded three-phase voltage and current waveforms during a fault event is shown in Figure 2. Structured metadata for all episodes are provided in a CSV file indexed by a unique sample_id. Metadata fields include fault type, affected phases, faulted line section, normalized fault location, fault resistance, inception time, randomized grid parameters (with physical units), and topology switching states (boolean flags). No derived or preprocessed features are included, allowing users to define task formulations explicitly. A concise summary of the dataset characteristics and core technical specifications is provided in Table III, facilitating transparent reporting and cross-study comparability.

The dataset comprises 9,022 fault scenarios without a nonfault class. Figure 3 summarizes the distribution of fault types, faulted line sections, and spatial fault locations. Fault categories (SLG, LL, LLG, LLL) are approximately uniformly distributed within each line section. The primary corridor sections (1–2A and 2–3A) each account for roughly 35 % of the dataset, while the secondary corridor sections (1–2B and 2–3B) contribute approximately 15 % each, reflecting the applied topology switching strategy. The spatial distribution of fault locations is approximately uniform along the full line length. Topology switching of Line 1–2B, Line 2–3B, and the external grid at Bus 3 is balanced (approximately 50/50), ensuring representation of both single- and double-circuit configurations as well as reduced infeed conditions. Basic integrity checks are performed prior to export. Simulations that fail load-flow convergence are discarded. Exported episodes are verified to contain no missing values. No predefined train/test split is provided; dataset partitioning is intentionally left to the user to enable application-specific evaluation protocols. Overall, the waveform structure, metadata schema, and balanced scenario coverage establish a reproducible and systematically documented benchmark for high-voltage fault studies. VI. R EPRODUCIBILITY AND DATA ACCESS PROTECT-90 is publicly available via Zenodo under DOI: 10.5281/zenodo.18418330 [12] and released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license. The archive (12.5 GB compressed, approximately 31 GB uncompressed) contains the complete waveform corpus and all domain-randomized grid and fault parameters described in Table II. Waveforms are provided as Python pickle files (Pandas DataFrames, Python 3.11, float32), accompanied by a structured CSV metadata file (7.3 MB) containing all labels and scenario parameters. No predefined train/test split is imposed, allowing users to define task-specific partitioning protocols. VII. I NTENDED U SAGE AND P RACTICAL R ECOMMENDATIONS The PROTECT-90 dataset and associated metadata enable multiple protection-oriented task formulations. For fault classification [13], the fields sc_type and phase_select provide categorical targets. For fault line identification [14], fault_target serves as a discrete label. For fault localization, the sc_location variable enables regression-based estimation of the fault position. Randomized grid parameters (e.g., Sk′′ , V , ℓ, R′ , X ′ ) allow controlled analysis of robustness and domain shift effects under varying operating conditions [15], [16]. Boolean topology flags support evaluation under single- and double-circuit configurations or reduced infeed conditions. For benchmarking, episode-wise data partitioning is recommended. Sliding windows derived from the same simulation

2400

3200

2100 1800

0.008

2400

1500

2000

1200

1600

900

1200

600

800

300

400

0

Density

Count

0.010

2800

LL

LLG

LLL

0.004 0.002

0

SLG

0.006

0.000

Li

–2 ne 1

A

2–3 Line

(a)

A

2–3 Line

B Li

–2 ne 1

B

0

20

40

60

80

100

Location along line (%)

(b)

(c)

Fig. 3. Coverage statistics of the EMT dataset. (a) Distribution of short-circuit fault types, showing balanced sampling across SLG, LL, LLG, and LLL 1 categories. (b) Distribution of faulted line segments, reflecting corridor switching states and resulting episode counts per segment. (c) Spatial fault-location distribution along the line segments, illustrating approximately uniform sampling across the full line length.

episode should not be distributed across training and test sets to prevent temporal leakage. Because the dataset contains only fault scenarios, detection tasks require explicit definition of windowing and labeling assumptions. Preprocessing steps, normalization procedures, and evaluation metrics should be reported to ensure comparable results across studies. The dataset is intentionally task-agnostic and does not prescribe specific window lengths, model architectures, or validation protocols, enabling flexible and application-dependent evaluation designs. VIII. C ONCLUSION This paper presents PROTECT-90, a publicly released electromagnetic transient dataset for high-voltage fault studies based on a standardized 90 kV double-line topology. The dataset comprises 9,022 physically consistent simulation episodes with synchronized three-phase voltage and current measurements and fully documented domain-randomized grid and fault parameters. All modeling assumptions, parameter ranges, and topology configurations are explicitly disclosed, enabling transparent and reproducible research. Unlike proprietary or partially documented waveform repositories, the dataset provides machine-readable metadata and complete traceability between measurements and underlying system conditions. By combining high-resolution EMT waveforms with structured parameter documentation, it establishes a physically grounded and fully documented benchmark for principled evaluation of protection-oriented signal processing and learning-based methods. The dataset is openly accessible under a permissive license and intended to support reproducible benchmarking, crossstudy comparison, and future extensions toward more complex grid topologies and measurement assumptions. ACKNOWLEDGMENT This project was funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) - 535389056.

R EFERENCES [1] T. Athay, R. Podmore, and S. Virmani, “A Practical Method for the Direct Analysis of Transient Stability,” IEEE Transactions on Power Apparatus and Systems, vol. PAS-98, no. 2, pp. 573–584, Mar. 1979. [Online]. Available: http://ieeexplore.ieee.org/document/4113518/ [2] V. Vittal and J. D. McCalley, Power system control and stability. Hoboken, New Jersey: Wiley, 2020. [3] Brett Ross and Kaveri Mahapatra, “IRTSD: Open-Source Data and Toolset for Electromagnetic Transient Analysis of Disturbances and IBR Control Malfunctions in Transmission Systems,” 2024. [Online]. Available: https://dx.doi.org/10.21227/mp6d-j677 [4] Y. Xie, L. Werner, K. Chen, T.-L. Le, C. Ortega, and S. Low, “A Digital Twin of an Electrical Distribution Grid: SoCal 28-Bus Dataset,” Apr. 2025, arXiv:2504.06588 [eess]. [Online]. Available: http://arxiv.org/abs/2504.06588 [5] C. Presvôts and T. Prevost, “Database of Voltage and Current Samples Values from the French Electricity Transmission Grid, Réseau de Transport d’Electricité (RTE), France,” 2024. [Online]. Available: https://github.com/rte-france/digital-fault-recording-database/ [6] A. J. Wilson, A. Riza Ekti, J. Follum, S. Biswas, C. Annalicia, J.-Y. Joo, O. Aziz, and J. Lian, “The Grid Event Signature Library: An Open-Access Repository of Power System Measurement Signatures,” IEEE Access, vol. 12, pp. 76 207–76 218, 2024. [Online]. Available: https://ieeexplore.ieee.org/document/10538319/ [7] J. Oelhaf, G. Kordowich, M. Pashaei, C. Bergler, A. Maier, J. Jäger, and S. Bayer, “A Scoping Review of Machine Learning Applications in Power System Protection and Disturbance Management,” International Journal of Electrical Power & Energy Systems, vol. 172, p. 111257, Nov. 2025. [Online]. Available: 10.1016/j.ijepes.2025.111257 [8] G. Ziegler, Numerical distance protection: principles and applications, 4th ed. Erlangen: Publicis Publishing, 2011. [9] R. Roeper and Mitlehner, Friedrich, Kurzschlußströme in Drehstromnetzen, 6th ed. Berlin, München: Siemens: Publicis Corporate Publishing, 1984. [10] D. Oeding and B. R. Oswald, Elektrische Kraftwerke und Netze, 8th ed. Berlin, Heidelberg: Springer Berlin Heidelberg, 2016. [Online]. Available: http://link.springer.com/10.1007/978-3-662-52703-0 [11] M. Wang, G. Kordowich, and J. Jäger, “A Generic Data Generation Framework for Short Circuit Detection Training of Neural Networks,” in PESS + PELSS 2022; Power and Energy Student Summit, 2022. [12] G. Kordowich, J. Oelhaf, C. Bergler, A. Maier, S. Bayer, and J. Jäger, “PROTECT-90: A Fault Dataset for Power System Protection: Open, Standardized Voltage and Current Waveforms for Reproducible Protection and Transient Analysis,” Mar. 2026. [Online]. Available: https://zenodo.org/doi/10.5281/zenodo.18418330 [13] J. Oelhaf, G. Kordowich, C. Kim, P. A. Pérez-Toro, C. Bergler, A. Maier, J. Jäger, and S. Bayer, “Controlled Comparison of Machine Learning Models for Fault Classification and Localization in Power System Protection,” 2025, version Number: 2. [Online]. Available: https://arxiv.org/abs/2510.00831

[14] J. Oelhaf, G. Kordowich, P. A. Pérez-Toro, T. Arias-Vergara, A. Maier, J. Jäger, and S. Bayer, “A Systematic Evaluation of Machine Learning Methods for Fault Detection and Line Identification in Electrical Power Grids,” in ICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). Hyderabad, India: IEEE, Apr. 2025, pp. 1–5. [Online]. Available: https://ieeexplore.ieee.org/document/10890544/ [15] J. Oelhaf, G. Kordowich, C. Kim, P. A. Pérez-Toro, A. Maier, J. Jäger, and S. Bayer, “Impact of Data Sparsity on Machine Learning for Fault Detection in Power System Protection,” in 2025 33rd European Signal Processing Conference (EUSIPCO). Palermo, Italy: IEEE, Sep. 2025, pp. 1997–2001. [Online]. Available: https://ieeexplore.ieee.org/document/11226584/ [16] J. Oelhaf, M. Pashaei, G. Kordowich, C. Bergler, A. Maier, J. Jäger, and S. Bayer, “Robustness Evaluation of Machine Learning Models for Fault Classification and Localization in Power System Protection,” in 20th IET International Conference on Developments in Power System Protection (DPSP Global 2026). IET, 2026.