neurips26-PSML/SIIB-Time

1# 1. Motivation2 3## Q1. For what purpose was the dataset created? 4 5This dataset was created to support research on machine learning (ML)-based surrogate modeling for power systems time-domain simulation, with a specific focus on simulation time step-invariance. The increasing penetration of inverter-based resources (IBRs) in modern power grids introduces ultrafast dynamic phenomena that require electromagnetic transient (EMT) simulation at microsecond time steps to capture accurately. This makes system-wide time-domain simulation computationally intractable, creating a critical bottleneck for stability analysis, contingency planning, and control design. In this context, researchers have been taking initial steps towards ML-based surrogate models for power system time-domain simulation.6The dataset fills a specific gap: no publicly available dataset provides paired EMT and RMS phasor-domain simulation trajectories of an inverter-based system under both grid-forming (GFM) and grid-following (GFL) control modes across a large and diverse set of operational scenarios. Existing datasets either provide data at a single fixed temporal resolution, cover only RMS simulation, or are not designed with machine learning applications in mind. This dataset directly addresses that gap by providing 3,000 scenarios simulated in both PSCAD (EMT, 50 µs resolution) and MATLAB Simulink (RMS, 1 ms resolution), enabling research on resolution-invariant operator learning methods.7 8## Q2. Who created the dataset (for example, which team, research group) and on behalf of which entity (for example, company, institution, organization)?9 10Will be added upon publication.11 12## Q3. Who funded the creation of the dataset? 13 14Will be added upon publication.15 16## Q4. Any other comments?17 18The dataset was generated between November 2025 and March 2026. It is publicly released under the Creative Commons Attribution 4.0 International License (CC BY 4.0) to maximize accessibility to the research community.19 20# 2. Composition21 22## Q5. What do the instances that comprise the dataset represent?23 24Each instance in the dataset represents a single operational scenario of the single inverter infinite bus (SIIB) system, a canonical power systems test case consisting of one inverter-based resource connected to a stiff upstream grid through a LCL filter and a feeder. Each instance captures the complete time-domain dynamic response of a random initialization to a random disturbance, either a short circuit fault or a load step, simulated under one of two inverter control modes: grid-forming (GFM) or grid-following (GFL). Each instance consists of two types of sub-records corresponding to two simulation domains as follows:25 26- EMT sub-record: Four CSV files recording the system's electromagnetic transient response at 50 µs resolution in PSCAD27- RMS sub-record: Four CSV files recording the system's phasor-domain response at 1 ms resolution in MATLAB Simulink28 29Both sub-records cover the same physical scenario under matched conditions, making each instance a paired multi-resolution observation of the same dynamic event. A metadata file accompanies each instance, recording the complete set of initial conditions and disturbance parameters that define the scenario.30 31## Q6. How many instances are there in total?32 33No labels were created prior to simulation or by external human annotators. Annotation is fully automated and embedded within the simulation models. The dataset contains 3,000 scenarios, each identified by a four-digit, zero-padded index (0001-1600). There are 1600 scenarios simulated for the GFM control mode, named xxxx_M where xxxx is in [0001, 1600], and 1400 scenarios simulated for the GFL control mode, named xxxx_L where xxxx is in [0001, 1400]. Each control mode scenario contributes one metadata file and 2 signal CSV files (one from the EMT simulation and one from the RMS phasor-domain simulation). The signal file follow the naming convention xxxx_[M|L]_[EMT|RMS].csv, where the EMT or RMS denotes the simulation domain. 34 35The metadata file follows the naming convention xxxx_[M|L]_meta_data.csv. Given there are 1,600 GFM scenarios and 1,400 GFL scenarios, the total number of CSV files are 1,600 × 3 + 1,400 × 3 = 9,000.36 37## Q7. Does the dataset contain all possible instances or is it a sample (not necessarily random) of instances from a larger set?38 39The dataset is a sample from the theoretically infinite space of all possible time-domain trajectories producible by the SIIB system under GFM and GFL control across all possible combinations of initial conditions and disturbances. The 3,000 scenarios were generated by changing the control mode as well as stochastic sampling of the scenario parameter space, active and reactive power references, grid impedance scaling, grid voltage sag scaling, disturbance type, disturbance timing, and disturbance configuration, using uniform distributions over predefined ranges encoded in the scenario generation scripts. The sample is representative of the parameter space as defined by those ranges, but is not representative of the real-world distribution of operating conditions observed in actual systems. The sampling procedure does not weight scenarios by their likelihood of occurrence in practice; all parameter combinations within the defined ranges are equally probable. Additionally, the 3,000 scenarios are not stratified to guarantee equal representation of stable and unstable outcomes; the class proportions are emergent properties of the physics. Users performing classification tasks should assess class balance before training. Further discussion of representativeness is provided in the accompanying dataset card.40 41## Q8. What data does each instance consist of?42 43The metadata and singal files contain the following information:44 45- xxxx_[M|L]_meta_data.csv: located in ./metadata, these files provides information needed to reconstruct or verify the simulation conditions for any scenario, and serves as the primary scenario-level annotation for downstream tasks. It is the authoritative annotation of each scenario's initial and disturbance conditions and records 16 rows, as follows:46  - Pref: float, active power reference (per unit)47  - Qref: float, reactive power reference (per unit; GFL mode only)48  - grid_impedance_scale: float, initial condition of grid impedance scaling factor (null if not applicable).49  - voltage_sag_factor: float, initial condition of grid voltage magnitude sag factor (null if not applicable)50  - disturbance type: string, either "Short circuit" or "Load Step up"51  - disturbance duration: float, duration of the short circuit in seconds, otherwise 052  - sc type: integer, 0 for load step up, 1-10 for short circuit type identifier53    - 1: a-g, 2: b-g, 3:c2g, 4:a-b-g, 5:a-c-g, 6: b-c-g, 7: a-b-c-g, 8: a-b, 9: a-c, 10: b-c54  - R1, R2, R3: float, per-phase resistance values of the stochastically sampled random load (Ω), used for load disturbance.55  - L1, L2, L3: float, per-phase inductance values of the random load (H), used for load disturbance.56  - C1, C2, C3: float, per-phase capacitance values of the random load (F), used for load disturbance.57 58- xxxx_[M|L]_[EMT|RMS].csv: located in ./data, these files are structured into 8 columns as follows:59  - Time column, t, in seconds, and the rows correspond to simulation time steps; simulations run from t=0 to t=6.5 s. 60  - Column Theta, measuring phase angle of the voltage in radians, derived from the inverter's internal synchronization signal (PLL output for GFL; droop-based frequency integration for GFM).61  - Columns, P and Q, measuring active power in MW and reactive power in Mvar, respectively.62  - Columns, Vd and Vq, measuring the d- and q-axis voltages in kV, respectively.63  - Columns, id, iq, measuring the d- and q-axis currents in kA, respectively.64 65## Q9. Is there a label or target associated with each instance?66 67Yes. A metadata file is associated with each scenario, as described before. 68 69## Q10. Is any information missing from individual instances?70 71No information is missing from any instance. 72 73## Q11. Are relationships between individual instances made explicit?74 75The primary relationship between instances is the pairing of EMT and RMS sub-records within each scenario; both sub-records share the same four-digit scenario index and the same xxxx_meta.csv file, making the pairing explicit through the naming convention. No other explicit relationships between different scenarios are encoded in the dataset. 76 77## Q12. Are there recommended data splits?78 79No fixed train/validation/test split is prescribed in the distributed dataset. 80 81## Q13. Are there any errors, sources of noise, or redundancies in the dataset?82 83No errors or redundancies have been identified. High-frequency numerical artifacts at the PWM switching frequency (8,000 Hz) are present in the raw PSCAD outputs but are attenuated by PSCAD's internal measurement filtering before export. This inherent noise must not be confused with processing or measuring noise.84 85## Q14. Is the dataset self-contained, or does it link to or rely on external resources?86 87The dataset is fully self-contained. 88 89## Q15. Does the dataset contain data that might be considered confidential?90 91No. 92 93## Q16. Does the dataset contain data that, if viewed directly, might be offensive, insulting, threatening, or might otherwise cause anxiety?94 95No.96 97## Q17. Does the dataset identify any subpopulations?98 99No. 100 101## Q18. Is it possible to identify individuals (that is, one or more natural persons), either directly or indirectly (that is, in combination with other data) from the dataset?102 103No.104 105## Q19. Does the dataset contain data that might be considered sensitive in any way?106 107No.108 109## Q20. Any other comments?110 111No.112 113## Q21. How was the data associated with each instance acquired?114 115All data was generated through physics-based time-domain simulation. It was not collected from human subjects, sensors, or third-party sources. Each instance was produced by running a fully parameterized simulation scenario in two platforms: PSCAD (EMT domain) and MATLAB Simulink (RMS phasor domain). The simulation models implement established inverter control architectures, cascaded voltage-current control with active power droop for GFM, and PLL-based current control for GFL, parameterized with representative but non-proprietary control gains, filter components, and system parameters, as follows:116- A three-phase upstream grid source:117    - Nominal voltage: 3.3 kV LL RM118    - Nominal frequency: 60 Hz119    - Grounded Neutrak120- Three-phase transformer:121    - Nominal rating: 5 MVA122    - Leakage reactance (EMT): 0.05 pu 123    - Leakage reactance (RMS): L_1 = 0.03 pu, L_2 = 0.02 pu (Matlab)124    - Grid-side winding: Delta125    - Load-side winding: Yg126- Inverter parameters (EMT):127    - V_DC = 1.5 kV, C_DC = 3900 µF128    - Switching frequency: 8000 Hz.129    - The active power reference P_ref (EMT and RMS) and the reactive power reference Q_ref (RMS) are varied across scenarios to sample a wide range of operating points. it is drawn from a uniform distribution over the interval [0.5, 1.7] pu.130- Inverter LCL filter:131    - L_1f = 60 µH, C_f = 1 mF, L_2f = 35 µH, 132    - Series damping resistor, R_f = 0.01 Ω133- A constant baseline load (R_L = 2 Ω, L_l = 0.01 H).134Within each simulation run, signals are recorded directly from measurement blocks (multimeters) embedded at the point of common coupling in the simulation model.135 136## Q22. What mechanisms or procedures were used to collect the data? How were these mechanisms validated?137 138Data generation was orchestrated through Python scripts that interfaced with the simulation platforms via their respective APIs. For each scenario, the Python scripts: (1) sampled scenario parameters stochastically from predefined distributions; (2) injected those parameters into the simulation model via the API; (3) executed the simulation; and (4) retrieved and organized the output files. PSCAD outputs were written natively in COMTRADE format and subsequently converted to CSV. MATLAB Simulink outputs were written directly to CSV by the simulation model at the end of each run. The scenario generation and data collection pipeline was validated by manually inspecting the CSV files of a subset of scenarios and verifying that the signal files correctly reflect the events. 139 140# 3. Collection Process141 142## Q23. If the dataset is a sample from a larger set, what was the sampling strategy?143 144Scenario diversity is achieved through stochastic parameterization of initial conditions and two disturbance categories; these are generated in Python and injected into the simulation models via their respective APIs.145 146Initial conditions include the power references set-points, grid impedance scale, and grid voltage sag scales. These initial conditions are sampled from uniform distributions. Active power in both the GFM and GFL modes is uniformly sampled from [0.5-1.7]. Reactive power in the GFL mode is uniformly sampled from [0.2-0.8]. The voltage sag scale is randomly selected from a uniform distribution between 0.8 and 0.99. The grid impedance scale is sampled from a uniform distribution between 1 and 7. The two scales are mutually exclusive, i.e., one is scaled and the other scaling factor remains 1 based on a 50-50% selection chance.147 148- Load disturbances: A random load of stochastically sampled magnitude is connected to the network at a randomly sampled time and disconnected at a later randomly sampled time. The three-phase random load can be unbalanced across phases in the EMT models. The load parameters are independently drawn from uniform distributions over predefined ranges: R_L ∈ [0.2,2] Ω, L_L ∈ [0.001,0.05] H, and C_L ∈ [1×10^(-6),50×10^(-6)] F. To account for phase imbalance, per-phase parameters are independently resampled from uniform distributions within ±15% of their respective average values, yielding a maximum inter-phase imbalance of 30%. The load connection time is uniformly sampled over the interval [0.5, 5] s.149- Fault disturbances: Short-circuit events are introduced at randomly sampled occurrence times with randomly sampled durations. The fault type is randomly selected among all the possible 10 three-phase fault types. A A fault ride-through (FRT) behavior is implemented; upon fault detection, the active power reference is set to zero and the inverter prioritizes reactive current injection for voltage support. The fault occurrence time is uniformly sampled over the interval [0.5, 5] s, and the fault duration is uniformly sampled within [0.02, 0.2] s.150 151## Q24. Who was involved in the data collection process (for example, students, crowdworkers, contractors) and how were they compensated?152 153Will be discussed upon publication.154 155## Q25. Over what timeframe was the data collected? Does this timeframe match the creation timeframe of the data associated with the instances?156 157The dataset was generated between November 2025 and March 2026. 158 159## Q26. Were any ethical review processes conducted?160 161No formal ethical review process was conducted, as the dataset contains no data about human subjects and poses no identifiable risks to individuals or communities. 162 163## Q27. Did you collect the data from the individuals in question directly, or obtain it via third parties or other sources?164 165N/A166 167## Q28. Were the individuals in question notified about the data collection?168 169N/A170 171## Q29. Did the individuals in question consent to the collection and use of their data?172 173N/A174 175## Q30. If consent was obtained, were the consenting individuals provided with a mechanism to revoke their consent in the future or for certain uses?176 177N/A178 179## Q31. Has an analysis of the potential impact of the dataset and its use on data subjects (for example, a data protection impact analysis) been conducted?180 181N/A182 183## Q32. Any other comments?184 185No.186 187# 4. Preprocessing/Cleaning/Labeling188 189## Q33. Was any preprocessing, cleaning, or labeling of the data done? If so, please provide a description.190 191No cleaning or normalization was applied at any stage to the data exported from the simulation environments. Two preprocessing operations were performed, one during format conversion and one during wrangling:192 193- COMTRADE-to-CSV conversion: PSCAD natively writes simulation outputs in COMTRADE format, consisting of a binary data file (out.dat) and a channel configuration file (out.cfg). A dedicated Python script converts these files into the five-CSV structure described before. This is a lossless recovery procedure, not a normalization or transformation. 194- Structural wrangling: Python scripts organize all simulation outputs into a consistent hierarchical structure and apply the standardized file naming convention (xxxx_[M|L]_[EMT|RMS_meta_data].csv). These operations are purely structural and do not modify signal values.195 196## Q34. Was the raw data saved in addition to the preprocessed/cleaned/labeled data?197 198Partially. The raw COMTRADE files produced by PSCAD (out.dat, out.cfg, out.hdr) are not distributed as part of the public dataset; only the converted CSV outputs are included. PSCAD COMTRADE files were not included and processed into CSVs to avoid proprietary formats. Users who require access to the native PSCAD COMTRADE files for any purpose should contact the dataset maintainers. 199 200## Q35. Is the software used to preprocess/clean/label the data available?201 202The simulation models and the scenario generation scripts and simulation API orchestration scripts are not publicly archived with the dataset at this time. Users wishing to extend or reproduce the data generation pipeline should contact the dataset maintainers. The simulation platforms used, PSCAD and MATLAB Simulink, are proprietary software products available under license from their respective vendors. They are not open-source and are not bundled with the dataset.203 204## Q36. Any other comments?205 206No.207 208# 5. Uses209 210## Q37. Has the dataset been used for any tasks already? If so, please provide a description.211 212Can be discussed upon publication.213 214## Q38. Is there a repository that links to any or all papers or systems that use the dataset?215 216Can be discussed upon publication.217 218## Q39. What other tasks could the dataset be used for?219 220The dataset is suitable for a range of adjacent research tasks:221 222- Machine learning for power systems dynamics: The paired multi-signal time-series structure makes the dataset suitable for training and evaluating a broad class of sequence-to-sequence models, recurrent neural networks, transformers, and physics-informed neural networks applied to power system trajectory prediction and stability assessment223- Educational use: The dataset is suitable for graduate-level courses on machine learning for power systems, providing a well-documented, physically grounded benchmark that students can use to develop and evaluate surrogate modeling methods without requiring access to commercial simulation software.224 225## Q40. Is there anything about the composition or collection of the dataset that might impact future uses? Is there anything a dataset consumer might need to know to avoid uses that could result in unfair treatment of individuals or groups?226 227The SIIB system is a canonical reduced-order abstraction of a single IBR connected to a strong grid; it is not a model of any specific real grid or installation. The population sampled is therefore the population of trajectories producible by this specific model family under the parameter ranges encoded in the scenario generation scripts, not the population of trajectories observable in real inverter installations. The produced sample does not represent multi-inverter interactions, network topology effects, and large-scale grid dynamics, among others. Moreover, the data representativeness depends on parameter selection and the underlying assumptions in control and system design. In this context, the distribution of scenarios is determined by the dataset design process rather than real-world statistical distributions. Some operating conditions may be over- or under-represented, and rare or extreme events may not be fully captured. Additionally, the 3,000 scenarios are not stratified to guarantee equal representation of stable and unstable outcomes. The proportions of each are emergent properties of the physics rather than design targets, and users performing classification tasks should assess class balance before training.228 229In synthetic engineering datasets, extrinsic bias operates differently than in datasets derived from human-generated text or behavioral data. That said, several structural biases warrant explicit acknowledgment. The dataset is built around power system conventions, standards, and test cases that predominantly originate from North American and European grid infrastructure traditions, specifically 60 Hz nominal frequency and grid parameters typical of North American distribution systems. Grids in the Global South, particularly in Sub-Saharan Africa, South Asia, and rural and remote communities in the Arctic and northern regions, often operate under fundamentally different conditions: weaker grids with lower short-circuit ratios, 50 Hz nominal frequency, different fault standards, and less standardized inverter hardware. The SIIB system, as parameterized here, may not fully reflect those conditions. 230 231## Q41. Are there tasks for which the dataset should not be used?232 233- Direct deployment in real grid operations: The dataset is a synthetic academic benchmark based on a simplified test system. It should not be used as the sole basis for training models that will be deployed in real grid control, protection, or planning systems without extensive additional validation on realistic, high-fidelity, and system-specific data.234- Claims of universal generalizability: Results obtained on this dataset should not be presented as evidence of general performance across all inverter types, control architectures, grid topologies, or geographic contexts. The dataset's scope is explicitly bounded, and conclusions should be scoped accordingly.235- Representation of real inverter hardware: The dataset should not be used to make claims about the behavior of specific commercial inverter products or manufacturer implementations, as the control parameters used are representative but not derived from any specific hardware.236 237## Q42. Any other comments?238 239The dataset is intentionally designed as a proof-of-concept benchmark for a nascent research direction. Its primary value lies in enabling controlled, reproducible experimentation with resolution-invariant operator learning methods in a physically grounded setting. As the research field matures, we anticipate that this dataset will be extended to more complex test systems, additional control architectures, and higher scenario counts. Users are encouraged to treat this dataset as a starting point rather than a definitive benchmark, and to contribute to the development of more comprehensive public datasets for ML-based power systems simulation research.240 241# 6. Distribution242 243## Q43. Will the dataset be distributed to third parties outside of the entity on behalf of which the dataset was created?244 245Yes. The dataset is publicly released and freely available to any researcher, practitioner, or institution worldwide. It is distributed externally via the Hugging Face dataset repository without restriction on who may access or use it, subject only to the terms of the CC BY 4.0 license described in Q46.246 247## Q44. How will the dataset be distributed? Does the dataset have a digital object identifier (DOI)?248 249The dataset is assigned a persistent Digital Object Identifier (DOI) through Hugging Face. The DOI resolves to both the dataset metadata and the dataset files.250 251DoI: 10.57967/hf/8655252 253## Q45. When will the dataset be distributed?254 255The dataset is publicly available as of April 30, 2026. 256 257## Q46. Will the dataset be distributed under a copyright or other intellectual property license and/or under applicable terms of use?258 259Yes. The dataset is released under the Creative Commons Attribution 4.0 International License (CC BY 4.0). 260 261## Q47. Have any third parties imposed intellectual property-based or other restrictions on the data associated with the instances?262 263No.264 265## Q48. Do any export controls or other regulatory restrictions apply to the dataset or to individual instances?266 267No. 268 269## Q49. Any other comments?270 271No.272 273# 7. Maintenance274 275## Q50. Who will be supporting, hosting, and maintaining the dataset?276 277The dataset is hosted on the Hugging Face platform, which provides infrastructure for long-term storage, versioning, and access. Dataset maintenance people will be discussed upon publication.278 279## Q51. How can the owner, curator, or manager of the dataset be contacted?280 281Will be discussed upon publication.282 283## Q52. Is there an erratum?284 285No erratum exists at the time of initial release. If errors are identified after release, an erratum will be posted in the Hugging Face repository's dataset card.286 287## Q53. Will the dataset be updated? If so, how often, by whom, and how will updates be communicated?288 289No major updates to the dataset are planned. If there is an update, the documentation and datacard on Huggingface will be updated by the maintainers and will reflect the changes made in the new version. 290 291## Q54. If the dataset relates to people, are there applicable limits on the retention of the data associated with the instances?292 293N/A294 295## Q55. Will older versions of the dataset continue to be supported, hosted, and maintained?296 297N/A298 299## Q56. If others want to extend, augment, build on, or contribute to the dataset, is there a mechanism for them to do so?300 301Given the proprietary nature of the simulation platform, there is no planned mechanism for community contributions to the dataset.302 303## Q57. Any other comments?304- This datasheet has been anonymized due to the double-blind peer-review process. Once accepted, these details will be updated.305- This datasheet follows the framework proposed by Gebru et al. (2021): "Datasheets for Datasets," Communications of the ACM, 64(12), 86–92. https://doi.org/10.1145/3458723. The datasheet should be read in conjunction with the accompanying Hugging Face dataset card, which provides more detailed technical documentation of the dataset's scope, data pipeline, quality, and management. Together, these two documents are intended to provide dataset consumers with all the information needed to make informed decisions about using this dataset for their chosen tasks.