{"id":"36e66e24-4999-4104-8684-841f24099c4e","arxiv_id":"2411.10431","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Jointly conditioning a diffusion-based inverse solver on multiple fault-event trajectories narrows the posterior over non-identifiable composite-load-model parameters and cuts mean parameter error by 58.6% in simulation.","lead":"This paper trains a conditional diffusion model to estimate the parameters of a power-grid composite load model from measured power trajectories, conditioning on several fault events at once. The authors report a 58.6% cut in parameter error and accurate response reproduction, as a way to attack the non-uniqueness of such inverse problems.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Selection of the 30 estimated parameters is informed by the same three fault classes used for evaluation, so the headline accuracy gains may partially reflect test leakage through the sensitivity screen.","rationale":"The Reader's weakest_assumption concerns fidelity of the synthetic pipeline (noise, model mismatch, prior mismatch). That is a legitimate and important limitation, explicitly acknowledged by the authors, and it justifies a CONDITIONAL rather than ACCEPT verdict. My stress-test pass surfaces a different, more proximal concern: the parameter-selection protocol uses the same fault events as the evaluation, creating a risk of selection leakage that is not acknowledged in the paper. This is not a fabrication or misconduct claim; it is a methodological circularity. The sensitivity screen is a standard dimensionality reduction technique, but using test-fault-specific Sobol rankings to choose which parameters to estimate means the reported error metrics measure the method on a favorably curated subset of the inverse problem. A holdout-fault experiment would settle whether the central 58.6% claim and OOD generalization results survive when the fault classes are not used to preselect the parameters. I therefore maintain CONDITIONAL, with the concrete test specified above. I partially agree with the reader because they correctly identify the sim-to-real gap as a key limitation, but I do not think that is the single most load-bearing concern for the central claim as stated.","tokens_in":23254,"tokens_out":1417,"duration_ms":12759,"concrete_test":"Hold out one entire fault class (e.g., stall) before sensitivity analysis and parameter selection: compute Sobol indices and select 30 parameters using only the ordinary and trip faults; train JCDI and CDI on the retained fault conditions; then evaluate MARPE and trajectory RMSE on the held-out stall fault. If the 58.6% reduction and RMSE gains persist under that protocol, the selection-leakage concern is resolved. If not, the headline gains are partially artifacts of parameter preselection tuned to the test faults.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central quantitative claim is the 58.6% MARPE reduction (18.0% to 7.46%) and the trajectory RMSE gains from joint conditioning. A load-bearing concern is that the task itself is partly selected using the test labels. In Section 2 and Supplementary Section 3, the authors compute Sobol total-effect indices under the three fault events (ordinary, trip, stall) and then select 30 parameters with 'relatively high sensitivity' using rankings from those same three events. The evaluation then measures MARPE and RMSE under exactly these three fault classes. This creates a circularity risk: the 30-parameter estimation problem is defined by what is identifiable under the test conditions, so the favorable results may be inflated relative to a deployment setting where the fault types are not known in advance. Concretely, parameters with low sensitivity under all three chosen faults are excluded from the estimation problem entirely; including them would change MARPE and trajectory RMSE. The 'universal framework' and OOD generalization claims are also weakened because the setup presupposes knowledge of the fault classes used for evaluation. This is not an internal inconsistency, but it is a correctness risk for the generalization claims. A clean test would hold out one fault class from the sensitivity screen and selection step, then evaluate the trained model on that held-out fault class.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes JCDI, a conditional diffusion-model framework for parameter estimation in ill-posed inverse problems, and demonstrates it on WECC CMPLDWG composite load model parameterization in an IEEE 39-bus ePHASORSIM setup. The method trains a transformer-based denoiser to generate parameter samples conditioned on observed active/reactive power trajectories; joint conditioning on multiple fault-event trajectories is claimed to narrow the posterior of non-identifiable parameters. The headline results are a reduction in mean absolute range percentage error (MARPE) from 18.0% (single-condition CDI) to 7.46% (JCDI), a 58.6% relative improvement, and trajectory re-simulation RMSEs below 4e-3 for ordinary, trip, and stall faults. The paper also reports an out-of-distribution generalization test with randomized fault locations and clearing times, comparisons against DQN and a ResNet-transformer supervised baseline, and discusses scalability, inference speed, and real-world deployment limitations.","tokens_in":23479,"tokens_out":7872,"duration_ms":78385,"significance":"If the central claim holds, the paper makes a useful empirical contribution: it shows that a diffusion model can serve as an amortized probabilistic inverse solver for a high-dimensional, non-identifiable dynamical-system parameterization problem, and that conditioning on multiple fault-event observations sharpens the estimated parameter distribution. The strengths of the empirical design are the controlled CDI-versus-JCDI comparison, the 1000-sample posterior analysis, the use of re-simulated trajectory RMSE rather than only parameter error, and a separate generalization test with randomized fault parameters. The diffusion training objective and the joint-conditioning factorization are standard, and the quantitative reduction is an empirical measurement rather than an algebraic artifact. The main caveats are that all data are synthetic from the same simulator, the parameter set is selected using the same fault classes used for evaluation, and real-PMU validation is explicitly left to future work, which the authors acknowledge in Section 3.","major_comments":[{"comment":"The 30-parameter estimation task is defined by a Sobol sensitivity screen performed under exactly the three fault events (ordinary, trip, stall) that are later used for evaluation (Supplementary Fig. 6 and Supplementary Table 2). Parameters with low sensitivity under all three of these events are excluded from the estimation problem, so the reported 58.6% MARPE reduction and the trajectory RMSE gains measure performance on a task that is partly selected using the test events. This is not an internal contradiction, but it is a correctness risk for the out-of-distribution and 'universal framework' claims: a deployment setting may contain fault classes not represented in the sensitivity screen. A concrete remedy is to hold out one fault class from the sensitivity analysis and parameter-selection step and evaluate on that class; at minimum, the authors should report results on a version of the problem that includes some low-sensitivity parameters.","section":"Section 2, Supplementary Sections 3-4"},{"comment":"The manuscript does not state whether CDI was trained on ordinary-fault trajectories only or on all three fault events. If CDI was trained only on ordinary-fault data, the comparison in Table 1 and the 58.6% reduction conflate the joint-conditioning mechanism with a difference in the size and diversity of the training set. The authors should specify the training data for each model and add an ablation trained on all three fault events but conditioned on a single event at inference, so that the effect of joint conditioning is isolated from the effect of additional training data.","section":"Section 2, Table 1, Supplementary Section 4"},{"comment":"The out-of-distribution test randomizes bus locations and fault clearing times, but every test scenario is assigned to one of the same three fault classes (ordinary, trip, stall) used in training and sensitivity selection. No truly unseen fault mechanism is tested, so the claim of 'favorable OOD generalization performance' in Section 3 is stronger than the evidence; this concern is closely related to the first major comment.","section":"Section 2, Fig. 6, Section 3"}],"minor_comments":[{"comment":"The factorization in Eq. (9) assumes conditional independence of observations given the parameters, and the text attributes this to independence of fault events; the authors should also state the assumed observation-noise model, since in a deterministic simulator the likelihood is a delta measure and the learned distribution is an amortized approximation of the posterior rather than a calibrated Bayesian posterior.","section":"Section 4, Eq. (9)"},{"comment":"There is a typo in 'Due to high computational burdern'; it should read 'burden'.","section":"Section 2, Table 1 paragraph"},{"comment":"References [25] and [26] repeat references [22] and [20], respectively; the duplicates should be consolidated.","section":"References"},{"comment":"The text says fault clearing times are selected from the empirical range of 3-6 cycles, but the listed values are 135 ms (about 8.1 cycles) and 44 ms (about 2.6 cycles); the statement and the table should be reconciled.","section":"Supplementary Section 2, Supplementary Table 1"},{"comment":"Both statements say the data and code 'will be made available prior to publication'; for review purposes, the authors should provide an anonymized repository or a clear timeline, since all results are synthetic and cannot otherwise be reproduced.","section":"Data availability / Code availability"}],"recommendation":"major_revision","confidential_remarks":"The paper is better scoped as a method demonstration on a synthetic power-system testbed than as a validated universal inverse-problem solver. The two most important fixes before acceptance are (i) clarifying and isolating the training-data confound in the CDI versus JCDI comparison, and (ii) addressing the selection-screen leakage by holding out a fault class or by including low-sensitivity parameters in an auxiliary evaluation. The 'universal framework' framing in the abstract and introduction is broader than a single-domain synthetic demonstration supports."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: this is a real and reasonably careful application of conditional diffusion models to a genuinely hard inverse problem in power systems — estimating ~200 CMPLDWG parameters from disturbance trajectories. The multi-event joint conditioning idea is the actual new bit, and the paper shows clearly, with 1000 posterior samples and re-simulated trajectories, that conditioning on three fault types narrows the posterior and cuts MARPE from 18.0% to 7.46%. The method comparison (DQN, supervised Res-TFR, CDI, JCDI) is internally consistent, and the OOD test with randomized faults is a good-faith attempt at generalization. The transformer-based denoiser and attention over parameter tokens is also a sensible architectural choice.\n\nThe soft spots are real but not fatal. The stress-test note is right: the 30 estimated parameters are picked by Sobol sensitivity ranking under the exact same three fault classes used for evaluation. So the headline numbers partly reflect the fact that the task was defined by what is identifiable under the test conditions. That is a leakage, not a contradiction — the paper is transparent about the selection, but the 'universal framework' and OOD claims should be read with that in mind. A clean test would hold out one fault class from both the sensitivity screen and the training, then evaluate on it.\n\nThe other weaknesses are the ones the authors themselves flag: all data come from one commercial simulator (ePHASORSIM), the prior is uniform over hand-picked ranges, and there is no real PMU validation. The scalability claims (100K+ parameters via diffusion) are optimistic extrapolation from image generation and are not demonstrated here. The discussion section is honest about these, which I respect.\n\nBottom line: within its stated synthetic setup, the central mechanism — joint conditioning narrows the posterior of non-identifiable parameters — is supported. The study is a credible proof-of-concept, but the absolute error numbers and the 'universal' framing should be treated as provisional until a held-out-fault and real-data test appear. No code or data released yet, which limits immediate reproducibility.\n\nFor review: yes, send it to a competent referee. It deserves a serious look, with the sensitivity-screen leakage addressed or acknowledged in revision. I would cite it as a solid reference for diffusion-based inverse solvers in power systems, and I would bring it to a reading group — the method is topical and the evaluation is unusually thorough for this area.\n\nBest,\n[You]","headline":"A solid, well-scoped simulation study of joint-conditioned diffusion for power-system load parameter estimation, with a real leak in the sensitivity screen and no real-data validation, but worth refereeing.","tokens_in":24051,"tokens_out":624,"would_cite":true,"duration_ms":8003,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A diffusion model conditioned on several fault-event observations at once cuts composite load-model parameter error by 58.6%.","keywords":["composite load model","CMPLDWG","diffusion model","inverse problem","parameter estimation","parameter non-uniqueness","joint conditioning","dynamic power systems"],"falsifier":"Record PMU data from real grid disturbances, feed the measured active and reactive power trajectories into the trained JCDI, and compare the estimated CMPLDWG parameters against independently verified values: if the parameter MARPE degrades toward the single-condition level of about 18% or the reconstructed trajectories exceed RMSE of roughly $10^{-2}$ under measurement noise and model mismatch, the synthetic-data premise fails. A complementary check that stays fully in simulation is to perturb the fault-clearing time or bus location of the three conditioning events and test whether the joint-conditioning benefit degrades continuously; if trajectory RMSE jumps to the CDI level for any perturbation, the posterior narrowing is specific to the three trained fault classes rather than a general resolution of non-identifiability.","tokens_in":23046,"feed_emoji":"⚡","tokens_out":7496,"duration_ms":62087,"temperature":0.7,"pith_summary":"The paper claims that the ill-posedness of inverse problems—different parameter sets producing identical observations—can be tamed by conditioning a generative diffusion model on several observations of the same system simultaneously. It introduces JCDI, a conditional denoising diffusion model that, instead of predicting parameters directly, generates samples from the posterior parameter distribution given measured active and reactive power trajectories. Tested on CMPLDWG composite load model parameterization in an IEEE 39-bus simulation, joint conditioning on three electrical fault events reduces the mean absolute range percentage error from 18.0% (single-event conditioning) to 7.46%, a 58.6% reduction. The same model reproduces post-fault power trajectories with root mean square errors below $4\\times10^{-3}$ and generalizes to out-of-distribution faults with randomized bus locations and clearing times. The authors further claim that, being purely data-driven, the same joint-conditioning recipe applies to parameter estimation beyond power systems.","feed_headline":"Diffusion model cuts load-parameter error by 58.6%","feed_subtitle":"Conditioning on several fault measurements at once shrinks the ambiguous parameter space of grid load models.","key_machinery":"The load-bearing object is the Joint Conditional Diffusion Model-based Inverse Problem Solver (JCDI): a conditional denoising diffusion probabilistic model run in parameter space rather than data space, with observed power trajectories as conditions. The denoising network is the Inverse Grid Transformer (IGT), a transformer encoder whose tokens are the 30 parameters being denoised, ResNet-encoded power trajectories, and a sinusoidally embedded diffusion time step; its attention mechanism is what lets the model learn correlations among parameters and between parameters and observations. The key innovation is multi-event joint conditioning: the reverse process is conditioned simultaneously on several fault-event trajectories, so the generated posterior is proportional to the product of the individual likelihoods, narrowing the set of non-identifiable parameter combinations. The training objective is standard DDPM noise prediction, minimizing $\\mathbb{E}\\|\\varepsilon - \\varepsilon_\\omega(x_t, t, y)\\|^2$, with inference by iterated denoising starting from Gaussian noise.","core_discovery":"The central discovery is that parameter non-identifiability in a complex dynamical system need not be removed by regularization or optimization; it is an uncertainty structure that a diffusion model can represent, and that additional observations can shrink. Working in the parameter space of the WECC composite load model CMPLDWG, JCDI learns the distribution of 30 sensitive parameters conditioned on observed power trajectories, so the posterior's spread encodes exactly which parameter combinations are interchangeable under a given disturbance. Conditioning the same model jointly on ordinary, trip, and stall fault trajectories multiplies the likelihood constraints, collapsing posterior width for correlated pairs such as Fmb/Fmc and CompPF/Qel0 and for parameters that are insensitive under any single fault, such as Rstall, Xstall, and frcel. The measured consequence is a drop in mean MARPE from 18.0% to 7.46%, and trajectory RMSEs of $7.84\\times10^{-4}$, $8.61\\times10^{-4}$, and $3.75\\times10^{-3}$ under ordinary, trip, and stall faults, where single-condition estimation degrades to $4.44\\times10^{-2}$ and $7.97\\times10^{-1}$. The paper also reports that the joint-conditioned estimates generalize to held-out fault scenarios with randomized locations and clearing times.","pith_inferences":["A natural stress test the paper does not run is a sweep over the number of conditioning events: if posterior variance keeps shrinking as a fourth or fifth fault is added, the method would behave like an online sensor-fusion scheme, and the power-system results would be a special case of a general 'more constraints, tighter posterior' principle.","The method's practical value depends on how many distinct informative disturbances a utility can actually record, which suggests a natural extension: actively selecting which fault or operating condition to record next so as to maximally reduce the remaining parameter uncertainty.","The framework assumes the forward simulator is the correct model class; a direct extension would be to train it on simulator output corrupted by realistic PMU noise, or to force generated parameters to produce trajectories matching real measurements. The authors flag real-data validation as future work, and the transfer of the reported gains likely hinges on it.","Because the 30 estimated parameters were chosen by Sobol sensitivity ranking across the three chosen faults, the framework could be combined with an adaptive sensitivity analysis that re-selects parameters as new fault classes are added during deployment."],"forward_implications":["Joint conditioning on multiple disturbance events removes the need for iterative forward simulation during training and inference, since the forward operator is only used once to build the dataset.","Parameters estimated under joint conditioning reproduce system dynamics under untrained fault events with randomized bus locations and clearing times, giving cross-event generalization that single-event posteriors lack.","Because the posterior is a distribution rather than a point estimate, JCDI makes parameter correlations and identifiability limitations visible: mutual information between correlated parameter pairs drops sharply when multiple conditions are imposed.","For the CMPLDWG case, trajectory RMSEs of $7.84\\times10^{-4}$, $8.61\\times10^{-4}$, and $3.75\\times10^{-3}$ under ordinary, trip, and stall faults are the paper's evidence that parameter accuracy translates directly into dynamic-response fidelity.","The same framework transfers, by design, to other electrical systems such as power electronic converters and energy storage, and to any dynamical system for which several observation records of the same unknown parameters are available."],"supporting_citations":[{"why":"Ho et al., denoising diffusion probabilistic models — supplies the forward-noising and reverse-denoising machinery that JCDI adapts from data space to parameter space.","marker":"[27]"},{"why":"WECC composite load model with DG specification — defines CMPLDWG, the system whose 30 sensitive parameters are the estimation targets.","marker":"[34]"},{"why":"Vaswani et al., 'Attention is all you need' — the transformer encoder that the Inverse Grid Transformer is built on, providing attention over parameter and trajectory tokens.","marker":"[57]"},{"why":"He et al., deep residual learning — the ResNet trajectory encoder that converts raw power trajectories into conditioning tokens.","marker":"[58]"},{"why":"Afrasiabi et al., multi-residual deep neural network — the supervised-learning baseline (Res-TFR) that JCDI outperforms in parameter MARPE and trajectory RMSE.","marker":"[21]"},{"why":"Wang et al., double deep Q-learning for WECC composite load modeling — the reinforcement-learning baseline (DQN) against which JCDI's accuracy and simulation-free efficiency are compared.","marker":"[18]"},{"why":"Tarantola, inverse problem theory — frames parameter estimation as an ill-posed inverse problem with non-unique solutions, the problem JCDI targets.","marker":"[1]"},{"why":"Kawar et al., denoising diffusion restoration models — the demonstration that diffusion models can solve inverse problems, which JCDI extends to dynamical-system parameter estimation.","marker":"[33]"}],"fun_headline_variants":["Joint fault data shrinks ambiguous grid parameters","Diffusion model learns all valid load-parameter combos","Multi-fault conditioning cuts parameter error 58.6%","Diffusion model with joint fault data cuts error 58.6%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the synthetic pipeline—uniformly sampling 30 hand-selected parameters, simulating CMPLDWG on the IEEE 39-bus system under three fault classes, and defining the posterior against a uniform prior—faithfully stands in for the real-world inverse problem; if real PMU measurements bring noise, model mismatch, or a different prior over plausible parameters, the measured 58.6% error reduction and reported RMSE gains may not transfer.","fun_headline_variants_meta":{"raw":{"variants":["Joint fault data shrinks ambiguous grid parameters","Diffusion model learns all valid load-parameter combos","Multi-fault conditioning cuts parameter error 58.6%","Diffusion model with joint fault data cuts error 58.6%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000865,"raw_usage":{"total_tokens":3772,"prompt_tokens":986,"completion_tokens":2786,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":602,"completion_tokens_details":{"reasoning_tokens":2718}},"tokens_in":602,"tokens_out":2786,"duration_ms":19736,"temperature":1.0,"reasoning_tokens":2718,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T19:38:04.301875+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Record PMU data from real grid disturbances, feed the measured active and reactive power trajectories into the trained JCDI, and compare the estimated CMPLDWG parameters against independently verified values: if the parameter MARPE degrades toward the single-condition level of about 18% or the reconstructed trajectories exceed RMSE of roughly $10^{-2}$ under measurement noise and model mismatch, the synthetic-data premise fails. A complementary check that stays fully in simulation is to perturb the fault-clearing time or bus location of the three conditioning events and test whether the joint-conditioning benefit degrades continuously; if trajectory RMSE jumps to the CDI level for any perturbation, the posterior narrowing is specific to the three trained fault classes rather than a general resolution of non-identifiability.","supporting_citations":[{"cited_title":"Technical report, Western Electricity Coordinating Council (WECC) (2015)","cited_arxiv_id":null,"evidence_quote":"Tarantola, inverse problem theory — frames parameter estimation as an ill-posed inverse problem with non-unique solutions, the problem JCDI targets."}],"review_version":1}