Pith. sign in

REVIEW 3 major objections 6 minor 25 references

A Survey of Open-Source Power System Dynamic Simulators with Grid-Forming Inverter for Machine Learning Applications

T0 review · 3 major / 6 minor · reviewed 2026-08-11 · deepseek-v4-flash

Pith's one-line read According to the paper's IEEE 14-bus benchmark, PSID.jl generates 1,000 ML training samples in 31.1 seconds, 7–12× faster than the other timed simulators.

desk verdict A useful but uneven survey of open-source GFM-capable dynamic simulators; the qualitative comparison is worth publishing after the timing claims are either substantiated or removed. read the letter →

arxiv 2412.08065 v1 pith:R6B64HXG submitted 2024-12-11 eess.SY cs.SY

classification eess.SYcs.SY
keywords open-sourcepowersystemdynamicsimulatorgrid-forminginvertertime-domainsimulationmachine-learningdatagenerationquasi-staticphasorANDESSimulationsDynamics.jlIEEE14-bus
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper sets out to tell machine-learning researchers which open-source power system dynamic simulator to use when generating training data for tasks like transient stability prediction and reinforcement learning, now that grid-forming (GFM) inverters—units that establish voltage and frequency rather than tracking the grid—need to be modeled. It surveys five tools—ANDES, PowerSimulationsDynamics.jl, Dynaωo, OpenDSS, and GridLAB-D—and compares them on modeling capability, GFM control support, programmability, and measured simulation speed. Its headline empirical result is a benchmark on the IEEE 14-bus system: PowerSimulationsDynamics.jl completes one 10-second dynamic simulation in 0.05 seconds and 1,000 Monte Carlo samples in 31.1 seconds, while ANDES takes 1.27 seconds and 224 seconds and Dynaωo takes 0.34 seconds and 374 seconds. The paper frames these numbers as practical guidance for ML data generation, and cross-checks ANDES and PSID.jl on a modified 14-bus case with GFM droop and virtual synchronous machine controls. If the comparison is fair, the practical takeaway is that simulator choice can change the cost of an ML dataset by an order of magnitude.

What carries the argument

The load-bearing device is the two-table comparison. Table I catalogs base functions—programming language, unbalanced modeling, power flow, small-signal stability analysis, TDS type, renewable and inverter model library, and parallel-computing support—so a reader can see what each tool can simulate. Table II catalogs ML-facing features—manual quality, ease of code-based parameter modification and result retrieval, external data formats, and simulation time per one run and per 1,000 samples—and supplies the quantitative ranking. The numerical anchor is a quasi-static phasor (QSP) simulation of a modified IEEE 14-bus system, a fast phasor-domain approximation of transmission network dynamics, run with a 0.01 s step size and a 10 s horizon, with generator 4 tripped at 1 s while generator 5 is set successively as a synchronous generator, a GFM droop controller, and a GFM virtual synchronous machine. The GFM model families (droop, VSM, dVOC, and matching control) are the shared objects whose support determines each simulator's relevance to this workload.

What would settle it

Re-run the 1,000-sample benchmark on the same IEEE 14-bus case on identical hardware with each simulator's documented solver and parallelization settings; if PowerSimulationsDynamics.jl no longer produces the shortest wall-clock time by the reported margin, the paper's ranking fails. A weaker check: repeat the comparison on a larger bus system or with a randomized fault scan and see whether the ordering changes.

Watch

Extended reading notes

Core claim

The paper's central claim is that, viewed through the needs of machine-learning data generation, the five surveyed open-source simulators differ substantially and measurably, and that the differences are captured in two comparison tables plus a small numerical case study. The decisive numbers are the simulation-time rows of Table II: for a modified IEEE 14-bus system with step size 0.01 s and a 10 s horizon, PowerSimulationsDynamics.jl takes 0.05 s per run and 31.1 s for 1,000 samples, compared with ANDES at 1.27 s per run and 224 s per 1,000 samples and Dynaωo at 0.34 s per run and 374 s per 1,000 samples; OpenDSS and GridLAB-D are marked not applicable because they target distribution systems. The paper frames this as practical guidance for ML professionals, and supports it with the observation that ANDES and PSID.jl produce similar GFM-Droop and GFM-VSM trajectories on the same 14-bus test case, suggesting that the speedup does not come from modeling a different system.

Load-bearing premise

The load-bearing premise is that the timing comparison in Table II is fair and representative: the simulators ran on comparable hardware with comparable solvers, the IEEE 14-bus case mirrors typical machine-learning data-generation workloads, and wall-clock time is the right criterion for suitability.

Editorial extensions

If this is right

  • For transmission-level machine-learning data generation at the scale of a 14-bus-like system, PowerSimulationsDynamics.jl cuts single-run cost by roughly 6–25× and 1,000-sample cost by roughly 7–12× compared with the other timed simulators, assuming the benchmark conditions hold.
  • ANDES, despite its slower single run, remains in contention for 1,000-sample batches when multi-CPU parallelization is used, as the paper explicitly notes.
  • OpenDSS and GridLAB-D are positioned for distribution-system studies rather than the transmission QSP workloads typical of bulk-system ML datasets, so the timing comparison intentionally excludes them.
  • The agreement between ANDES and PSID.jl trajectories under GFM droop and GFM VSM control supports using either of the two as a dynamical reference while picking the faster one for bulk data generation.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because the paper does not report hardware, DAE solver choice, or parallelization settings, the cleanest reading is that the ranking is a demonstration that large speedups exist, not a guarantee on every machine; users should re-time on their own cluster before scaling up.
  • A natural extension the paper leaves implicit is to benchmark the same tools on larger systems and on fault scans that sample many locations, since ML datasets are rarely built from a single trip scenario.
  • The comparison highlights an unresolved trade-off: QSP-based tools are fastest, but ML models that must learn converter switching or fast electromagnetic transient dynamics would need Dynaωo's quasi-EMT mode, which is much slower at 374 s per 1,000 samples; the paper does not quantify that trade-off.
  • If the ranking holds across system sizes, the choice of simulator is itself a practical hyperparameter of any ML pipeline, and reporting it alongside model performance would make power-system ML studies more reproducible.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The manuscript surveys five open-source power system dynamic simulators (ANDES, PowerSimulationsDynamics.jl, Dynaωo, OpenDSS, and GridLAB-D) that support grid-forming (GFM) inverter models, with an emphasis on their suitability for machine-learning data generation. It categorizes time-domain simulation approaches, reviews GFM control structures, and presents feature comparisons in Tables I and II. The authors report time-domain simulation case studies on a modified IEEE 14-bus system for ANDES and PSID.jl with generator 5 configured as SG, GFM-droop, and GFM-VSM. The central quantitative contribution is Table II's simulation-time comparison, which ranks PSID.jl fastest (0.05 s/31.1 s per run/1,000 samples), Dynaωo intermediate (0.34 s/374 s), and ANDES slowest (1.27 s/224 s), and the paper concludes that these numbers provide practical guidance for ML practitioners.

Significance. The survey addresses a timely need: open-source simulators with GFM capability are evolving quickly, and ML researchers require guidance on which tools support flexible scenario generation and fast data production. The paper's strengths are organizational: it consolidates GFM model availability (e.g., REGCA1/REGCV1/REGF1-3 in ANDES, GFM-droop/VSM/dVOC/matching in PSID.jl and Dynaωo), interface features, and data formats from project documentation, and it provides a concrete two-simulator case study. If the timing comparison can be made reproducible and properly scoped, the survey would be a useful reference. At present, however, the headline numerical ranking is not supported by the evidence reported in the manuscript.

major comments (3)
  1. [Section IV, Table II] The Dynaωo timing entries (0.34 s/374 s) in Table II are not supported by any described measurement. Section IV states that 'TDSs were performed on these two simulators' (ANDES and PSID.jl) and no Dynaωo simulation is described anywhere in the paper. Please state explicitly whether the Dynaωo values were measured by the authors, taken from project documentation, or estimated, and under what conditions. Without this provenance, the ranking implied by Table II—PSID.jl fastest, Dynaωo intermediate, ANDES slowest—is not established.
  2. [Section IV, Table II] The wall-clock-time comparison is under-specified to the point of not being reproducible. The caption reports only step size (0.01 s) and duration (10 s); it omits hardware, solver algorithm and tolerances, parallelization method, warm-up, the procedure used to generate 1,000 samples, and, for Dynaωo, whether the timing used DynaSwing (QSP) or DynaWave (quasi-EMT). If Dynaωo's entry reflects a quasi-EMT simulation while ANDES and PSID.jl use positive-sequence QSP, the entries are not directly comparable. Please provide a full measurement protocol or restrict the quantitative claim to the two simulators actually benchmarked.
  3. [Table II] The relationship between the single-run and 1,000-sample timings is internally inconsistent. The entries imply average per-sample times of 0.224 s for ANDES, 0.031 s for PSID.jl, and 0.374 s for Dynaωo, which are not consistent with the stated single-run times (1.27 s, 0.05 s, and 0.34 s, respectively). If the 1,000-sample runs are simply 1,000 independent simulations, the averages should be close to the single-run times; the large deviations indicate undocumented parallelization, warm-starting, or different sampling procedures. Please define how the 1,000-sample values were generated and confirm that the same procedure was used for all simulators.
minor comments (6)
  1. [Section IV] The text says the modified 14-bus system structure and results are shown in Figs. 3, 4, and 5, but Fig. 3 is the DL/RL framework diagram and no 14-bus system structure figure appears; please renumber the figures or insert the missing diagram.
  2. [Table II] The header 'Simulation time (1/1000 samples)' is ambiguous; please state explicitly that the entries are single-run time / total wall-clock time for 1,000 samples.
  3. [Tables I and II] The symbol '/' is used with two different meanings: 'does not mention this function' in Table I and 'not applicable' in Table II; please define the symbols separately for each table.
  4. [Table II, Section III-A] The statement that 'due to Python's slower speed compared to Julia and C++, ANDES has the longest simulation time' is a causal attribution that is not established by the reported measurements; solver implementation, model complexity, and numerical settings are also relevant.
  5. [Section IV] The claim that the ANDES and PSID.jl results are 'very similar' is qualitative; please add a quantitative comparison (e.g., maximum or RMS trajectory error) and specify the GFM control parameters used, so that the case study is reproducible.
  6. [Tables I and II] No simulator version numbers or documentation access dates are given; since capabilities and model lists change between releases, please state the versions reviewed.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the survey's comparisons are self-contained descriptions from simulator documentation; the only self-citation is peripheral, and the under-specified Dynaomegao timing is a transparency issue, not a circular derivation.

full rationale

This is a survey paper with no mathematical derivation chain that could reduce to its own inputs. The capability and model-support claims in Tables I and II are presented as descriptions of each open-source simulator, sourced from the simulators' own documentation and cited papers, which is the appropriate evidentiary basis for a tool survey. The only self-citation is reference [21], a safe-reinforcement-learning review by the authors, used to support the background statement that 'safe RL has been proposed to maximize rewards while ensuring constraint satisfaction through a series of specially designed techniques'; this is peripheral and not load-bearing for the survey's conclusion. The one evidentiary weakness flagged by the review is the provenance of the Dynaomegao timing: Section IV states 'Since ANDES and PSID.jl provide detailed manuals and programming guidance, TDSs were performed on these two simulators,' while Table II reports Dynaomegao at 0.34s/374s without a described measurement, hardware, solver, or fidelity details. This is a transparency and comparability problem for the empirical ranking, but it is not circularity: no fitted parameter is renamed as a prediction, and no conclusion is defined in terms of its own inputs. Per the hard rules, unsupported measurements and non-standard comparison conditions belong to correctness risk, not to the circularity score.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

No fitting or derivation occurs; the only hand-chosen numbers are the benchmark settings (step size, duration). The survey's claims rest on the accuracy of project documentation and on the representativeness of the single benchmark case.

free parameters (2)
  • Simulation step size = 0.01 s
    Chosen for the benchmark in Table II; measured simulation times scale with step size, so the comparison depends on this setting.
  • Simulation duration = 10 s
    Chosen for the benchmark in Table II; the observable period affects total runtime and the plotted dynamics.
assumptions (3)
  • domain assumption The capability claims in Table I accurately describe the cited versions of each simulator.
    The survey copies feature claims from each project's documentation ([4], [11], [12], [22], [23]) without independent testing.
  • domain assumption The IEEE 14-bus case is representative of machine-learning data-generation workloads.
    Section IV uses one modified 14-bus test system; generality to other system scales and GFM models is asserted without evidence.
  • ad hoc to paper Wall-clock times for QSP and quasi-EMT simulations are directly comparable.
    Table II lists timings for Dynaomegao's quasi-EMT alongside ANDES and PSID.jl QSP runs, an implicit but unsupported comparability assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Survey of Open-Source Power System Dynamic Simulators with Grid-Forming Inverter for Machine Learning Applications." pith.science (2026). https://pith.science/paper/R6B64HXG

@misc{pith2026241208065,
  author       = {Pith},
  title        = {Pith review of: A Survey of Open-Source Power System Dynamic Simulators with Grid-Forming Inverter for Machine Learning Applications},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/R6B64HXG}},
  note         = {Machine review of arXiv:2412.08065}
}
read the original abstract

The emergence of grid-forming (GFM) inverter technology and the increasing role of machine learning in power systems highlight the need for evaluating the latest dynamic simulators. Open-source simulators offer distinct advantages in this field, being both free and highly customizable, which makes them well-suited for scientific research and validation of the latest models and methods. This paper provides a comprehensive survey and comparison of the latest open-source simulators that support GFM, with a focus on their capabilities and performance in machine-learning applications.

Figures

Figures reproduced from arXiv: 2412.08065 by the authors.

Figure 2
Figure 2. Structure of the GFM control: (a) Droop control. (b) VSM control. [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 1
Figure 1. Classification of power system TDS. B. Inverter-Based Resource Traditional inverters use grid-following (GFL) control, re￾lying on the phase-locked loop (PLL) to estimate and track the angle of the terminal voltage and regulate the injected current into the grid. However, GFL inverters, while unable to actively control voltage and frequency, behave similarly to a current source. When facing a contingency, GFL will a… view at source ↗
Figure 3
Figure 3. Framework of interaction between DL, RL, and dynamic simulators. [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: ANDES simulation results: from top to bottom, the three rows correspond to G5 configured as SG, GFM-Droop, and GFM-VSM, respectively; From [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: PSID.jl simulation results: from top to bottom, the three rows correspond to G5 configured as SG, GFM-Droop, and GFM-VSM, respectively; From [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 23 canonical work pages

  1. [21]

    A review of safe reinforcement learning methods for modern power systems,

    T. Su, T. Wu, J. Zhao, A. Scaglione, and L. Xie, “A review of safe reinforcement learning methods for modern power systems,” arXiv preprint arXiv:2407.00304, 2024

  2. [1]

    Revisiting power systems time-domain simulation methods and models,

    J. D. Lara, R. Henriquez-Auba, D. Ramasubramanian, S. Dhople, D. S. Callaway, and S. Sanders, “Revisiting power systems time-domain simulation methods and models,” IEEE Trans. Power Syst., vol. 39, no. 2, pp. 2421–2437, Mar. 2024

  3. [2]

    Machine learning driven smart electric power systems: Current trends and new perspectives,

    M. S. Ibrahim, W. Dong, and Q. Yang, “Machine learning driven smart electric power systems: Current trends and new perspectives,” Appl. Energy, vol. 272, 2020, Art. no. 115237

  4. [3]

    Reinforcement learning for selective key applications in power systems: Recent advances and future challenges,

    X. Chen, G. Qu, Y . Tang, S. Low, and N. Li, “Reinforcement learning for selective key applications in power systems: Recent advances and future challenges,” IEEE Trans. Smart Grid , vol. 13, no. 4, pp. 2935–2958, Jul. 2022. 5 /uni0000002a/uni00000048/uni00000051/uni00000048/uni00000055/uni00000044/uni00000057/uni00000052/uni00000055/uni00000003/uni000000...

  5. [4]

    Hybrid symbolic-numeric framework for power system modeling and analysis,

    H. Cui, F. Li, and K. Tomsovic, “Hybrid symbolic-numeric framework for power system modeling and analysis,” IEEE Trans. Power Syst. , vol. 36, no. 2, pp. 1373–1384, Mar. 2021

  6. [5]

    MAT- POWER: Steady-state operations, planning, and analysis tools for power systems research and education,

    R. D. Zimmerman, C. E. Murillo-S ´anchez, and R. J. Thomas, “MAT- POWER: Steady-state operations, planning, and analysis tools for power systems research and education,” IEEE Trans. Power Syst., vol. 26, no. 1, pp. 12–19, Feb. 2011

  7. [6]

    Pandapower—an open-source Python tool for convenient modeling, analysis, and optimization of electric power systems,

    L. Thurner, A. Scheidler, F. Sch ¨afer, J.-H. Menke, J. Dollichon, F. Meier, S. Meinecke, and M. Braun, “Pandapower—an open-source Python tool for convenient modeling, analysis, and optimization of electric power systems,” IEEE Trans. Power Syst. , vol. 33, no. 6, pp. 6510–6521, Nov. 2018

  8. [7]

    PowerSystems.jl—a power system data management package for large scale modeling,

    J. D. Lara, C. Barrows, D. Thom, D. Krishnamurthy, and D. Callaway, “PowerSystems.jl—a power system data management package for large scale modeling,” SoftwareX, vol. 15, 2021, Art. no. 100747

Show all 25 references
  1. [8]

    PyPSA: Python for power system analysis,

    T. Brown, J. H ¨orsch, and D. Schlachtberger, “PyPSA: Python for power system analysis,” J. Open Res. Softw. , vol. 6, no. 4, Jan. 2018

  2. [9]

    PowerModels.jl: An open-source framework for exploring power flow formulations,

    C. Coffrin, R. Bent, K. Sundar, Y . Ng, and M. Lubin, “PowerModels.jl: An open-source framework for exploring power flow formulations,” in Proc. Power Syst. Comput. Conf. , Jun. 2018, pp. 1–8

  3. [10]

    PowerModelsDistri- bution. jl: An open-source framework for exploring distribution power flow formulations,

    D. M. Fobes, S. Claeys, F. Geth, and C. Coffrin, “PowerModelsDistri- bution. jl: An open-source framework for exploring distribution power flow formulations,” Electric Power Syst. Res. , vol. 189, 2020, Art. no. 106664

  4. [11]

    An open source platform for collaborating on smart grid research,

    R. C. Dugan and T. E. McDermott, “An open source platform for collaborating on smart grid research,” in Proc. IEEE Power Energy Soc. General Meeting , 2011, pp. 1–7

  5. [12]

    GridLAB-D: an agent-based simulation framework for smart grids,

    D. P. Chassin, J. C. Fuller, and N. Djilali, “GridLAB-D: an agent-based simulation framework for smart grids,” J. Appl. Math. , vol. 2014, no. 1, 2014, Art. no. 492320

  6. [13]

    Review: Power system analysis software tools,

    L. Bam and W. Jewell, “Review: Power system analysis software tools,” in Proc. IEEE Power Energy Soc. General Meeting , 2005, pp. 139–144

  7. [14]

    A survey and statistical analysis of smart grid co-simulations,

    M. V ogt, F. Marten, and M. Braun, “A survey and statistical analysis of smart grid co-simulations,” Appl. Energy, vol. 222, pp. 67–78, 2018

  8. [15]

    An open source power system analysis toolbox,

    F. Milano, “An open source power system analysis toolbox,” IEEE Trans. Power Syst., vol. 20, no. 3, pp. 1199–1206, Aug. 2005

  9. [16]

    A toolbox for power system dynamics and control engineering education and research,

    J. H. Chow and K. W. Cheung, “A toolbox for power system dynamics and control engineering education and research,” IEEE Trans. Power Syst., vol. 7, no. 4, pp. 1559–1564, Nov. 1992

  10. [17]

    Frequency stability of synchronous machines and grid-forming power converters,

    A. Tayyebi, D. Groß, A. Anta, F. Kupzog, and F. D ¨orfler, “Frequency stability of synchronous machines and grid-forming power converters,” IEEE J. Emerg. Sel. Topics Power Electron. , vol. 8, no. 2, pp. 1004– 1018, 2020

  11. [18]

    Comparison of grid fol- lowing and grid forming control for a high inverter penetration power system,

    D. Pattabiraman, R. Lasseter, and T. Jahns, “Comparison of grid fol- lowing and grid forming control for a high inverter penetration power system,” in IEEE Power Energy Soc. Gen. Meeting , 2018, pp. 1–5

  12. [19]

    Grid-forming inverters: Are they the key for high renewable penetration?

    J. Matevosyan, B. Badrzadeh, T. Prevost, E. Quitmann, D. Ramasub- ramanian, H. Urdal, S. Achilles, J. MacDowell, S. H. Huang, V . Vital et al. , “Grid-forming inverters: Are they the key for high renewable penetration?” IEEE Power Energy Mag., vol. 17, no. 6, pp. 89–98, 2019

  13. [20]

    Transient stability of voltage- source converters with grid-forming control: A design-oriented study,

    D. Pan, X. Wang, F. Liu, and R. Shi, “Transient stability of voltage- source converters with grid-forming control: A design-oriented study,” IEEE J. Emerg. Sel. Topics Power Electron. , vol. 8, no. 2, pp. 1019– 1033, 2019

  14. [22]

    PowerSimulationsDynamics.jl–an open source modeling package for modern power systems with inverter-based resources,

    J. D. Lara, R. Henriquez-Auba, M. Bossart, D. S. Callaway, and C. Barrows, “PowerSimulationsDynamics.jl–an open source modeling package for modern power systems with inverter-based resources,” arXiv preprint arXiv:2308.02921, 2023

  15. [23]

    Towards an open-source solution using modelica for time-domain simulation of power systems,

    A. Guironnet, M. Saugier, S. Petitrenaud, F. Xavier, and P. Panciatici, “Towards an open-source solution using modelica for time-domain simulation of power systems,” in Proc. IEEE PES Innovative Smart Grid Technol. Conf. Europe , 2018, pp. 1–6

  16. [24]

    Small-signal sta- bility impacts of load and network dynamics on grid-forming inverters,

    R. Henriquez-Auba, J. D. Lara, and D. S. Callaway, “Small-signal sta- bility impacts of load and network dynamics on grid-forming inverters,” in IEEE Power Energy Soc. Innov. Smart Grid Technol. Conf. , 2024, pp. 1–5

  17. [25]

    Abiola-Ellison, Getting Started with Lazarus and Free Pascal: a beginners and intermediate guide to Free Pascal using Lazarus IDE

    M. Abiola-Ellison, Getting Started with Lazarus and Free Pascal: a beginners and intermediate guide to Free Pascal using Lazarus IDE . CreateSpace Independent Publishing Platform, 2015

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.