{"id":"81f76050-57e0-45b6-a7df-17ddc1dbace8","arxiv_id":"2506.23304","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"An artificial neural network estimates grid impedance online and feeds an adaptive gain scheduler that keeps virtual synchronous generator settling time and overshoot consistent across weak, strong, and stiff grids in simulation.","lead":"This paper trains an artificial neural network to read one cycle of voltage and current from the grid connection point, estimate the grid's impedance, and then automatically retune a virtual synchronous generator controller to keep it stable as the grid strengthens or weakens. The appeal is the prospect of inverter-based renewables that adapt to changing grid conditions without manual redesign.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Gain-scheduling equations (10)-(11) are inconsistent with the gains in Table II: the designed products DpKip=8 and Kiq(D+Dq)=4 appear as 16 and ~8 in the table, so the reported simulations may not implement the described adaptive law.","rationale":"In stress-testing the paper, I focused on whether the reported simulation results can be traced to the proposed gain-scheduling law. The reader's weakest assumption concerned the online update of Vpcc0 and delta0. That is a real gap, but the active-loop gain A is only weakly sensitive to delta0 at small power angles, so it may not be the deciding issue. A more decisive problem is that the numerical gains in Table II do not satisfy the design equations (10)-(11) even by the paper's own formulas. The products DpKip and Kiq(D+Dq) are determined by the desired settling time alone; the table gives values twice as large. The text explicitly says both schemes use the same gains in the weak-grid segment, so these are the scheduled gains. This internal inconsistency means the manuscript does not demonstrate that the proposed adaptive law generates the gains used in the simulation. It is an addressable issue, so the verdict remains conditional, but it is more load-bearing than the operating-point gap because it can be checked directly from the text and would require only a factor-of-2 correction or a table revision. If the discrepancy is real, the central claim lacks support; if it is a typo, the authors must correct it and re-verify. I therefore keep the reader's CONDITIONAL verdict but flag this as the first thing to resolve.","tokens_in":7422,"tokens_out":21992,"duration_ms":214776,"concrete_test":"Recompute the scheduled gains from equations (10)-(11) at the SCR=2 operating point using the actual grid impedance, Vpcc0, and delta0 from the Simulink model, and compare with Table II. If the tabulated DpKip is 16 rather than 8, the model did not implement (10)-(11); equivalently, rerun the simulation with DpKip = 8 (e.g., keep Kip = 0.00767 and set Dp = 1043) and verify whether the settling time is still 1 s and whether the AVSG and CVSG responses are still identical in the first 20 s. This factor-of-2 check directly decides whether the reported results are produced by the described adaptive gain scheduler.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III-B and the Appendix derive the adaptive gains from a desired 1 s settling time with critical damping in the active loop, giving DpKip = 2*xi*omega_n = 8 and Kiq(D + Dq) = 4 from the reactive first-order time constant. Yet Table II lists Dp = 2.087e3 and Kip = 0.00767, so DpKip = 16.0, and with Dq = 0.687 (so D = 100*Dq = 68.7) the tabulated Kiq = 0.115 gives Kiq(D + Dq) = 7.98. The text states that in the SCR=2 interval the adaptive and conventional VSGs use the same values of Dp, Kip, Dq, and Kiq, so these are the scheduled gains at the baseline condition. Since the products DpKip and Kiq(D + Dq) are fixed independently of A and D by the settling-time specification, no operating point or impedance can reconcile Table II with equations (10)-(11). If the table reflects the actual simulation, the implemented scheduler is not the one described; if the equations are correct, the reported settling time and overshoot values are not those designed. Either way the central claim that the proposed gain-scheduling law maintains performance is not supported by the manuscript as written.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes an adaptive gain-scheduling scheme for a virtual synchronous generator (VSG). A feedforward ANN is trained offline on Simulink-generated data to estimate the fundamental-frequency grid resistance and inductance from one cycle of PCC voltage and current samples. The estimated impedance is then used to recompute the power-flow Jacobian terms A and D, and the VSG power-loop gains are scheduled so that, according to the Appendix, the active loop has a settling time of 1 s with critical damping and the reactive loop uses a first-order time constant of 0.25 s. The scheme is validated in a 60 s Simulink simulation with SCR values of 2, 8, and 20, and is compared with a fixed-gain VSG. The authors report that the adaptive VSG maintains settling time and overshoot across grid conditions while the fixed-gain VSG degrades, and that the ANN estimates previously unseen impedances with small delay.","tokens_in":7696,"tokens_out":9003,"duration_ms":97487,"significance":"If the consistency issues below are resolved, the paper offers a practical engineering contribution: it combines online impedance estimation with analytically derived gain-scheduling rules, and the design targets (Ts=1 s, xi=1) are explicit rather than fitted to the validation outputs. The use of holdout SCR values, the public code link, and the direct comparison against a fixed-gain VSG are strengths. However, the contribution is incremental relative to existing adaptive VSG and ANN impedance-estimation work, the validation is entirely simulation-based, and the current manuscript contains an internal inconsistency between the derived gain law and the tabulated gains that undermines the central claim as written.","major_comments":[{"comment":"Equations (10)-(11) and the Appendix are inconsistent with the gains reported in Table II. The Appendix design fixes DpKip=8 and Kiq(D+Dq)=4 regardless of the operating point. With the Table II values Dp=2.087e3 and Kip=0.00767, the active product is 16.0; with Dq=0.687 (so D=68.7), the reactive product Kiq(D+Dq) is 7.98. Section IV-B states that at SCR=2 the adaptive and conventional VSGs use the same four gains, so these are the scheduled gains at the baseline condition. No choice of A or D can reconcile these numbers because the designed products are independent of A and D. Either the implemented scheduler is not the law described by (10)-(11), or the tabulated parameters do not correspond to the simulated system. In both cases the claim that the reported 1 s settling time and critical damping follow from the proposed law is not supported. The authors should correct one side and re-run or re-report the validation.","section":"Section III-B / Table II / Appendix"},{"comment":"The gain-scheduling law requires online values of the power-flow Jacobian elements A and D, which depend on Vpcc0, delta0, Rg, and Xg. The manuscript states that the Jacobian components are 'first recalculated' after an impedance update, but it never specifies how Vpcc0 and delta0 are obtained or updated during operation. If fixed nominal values are used, the scheduled gains will not realize the intended Ts=1 s and critical damping once the operating point changes; if measured values are used, the measurement and update mechanism should be described. Without this step the scheme cannot be reproduced or assessed under the 'dynamic grid conditions' claimed in the paper.","section":"Section III-B, Eqs. (4)-(5)"}],"minor_comments":[{"comment":"Equation (11) reads 'Dp = D/100', which appears to be a typo for 'Dq = D/100'; as written it reassigns the active droop coefficient instead of the reactive droop coefficient.","section":"Section III-B, Eq. (11)"},{"comment":"SCR=8 is between the trained values 7 and 9.5, so describing it as 'previously unseen' is imprecise; only SCR=20 is an extrapolation outside the training range.","section":"Section IV-B / Fig. 9"},{"comment":"The IAE values in Fig. 8 are not tied to a specific trace; clarify whether they apply to the conventional VSG, the adaptive VSG, or both.","section":"Section IV-B / Fig. 8"},{"comment":"The settling-time and overshoot claims are made qualitatively; a table listing settling time and percentage overshoot for each SCR and for both AVSG and CVSG would make the central comparison quantitative and easier to verify.","section":"Section IV-B"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is readable and the small-signal derivation is coherent, but the mismatch between Table II and the gain formulas in Section III-B/Appendix is a serious internal inconsistency that must be resolved before the validation can be trusted. I would ask the authors to provide the actual scheduled gains for SCR=2, 8, and 20 and to confirm which values were used in the simulations. If the figures were produced with a scheduler different from the text, the contribution and conclusions change. The contribution is incremental relative to prior work on VSG adaptive control and ANN impedance estimation, and the validation is simulation-only; nonetheless, the proposed approach is plausible and the claimed results are testable, so a careful major revision is appropriate rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea is clean: estimate grid impedance with an ANN, feed it into an analytic gain-scheduling law for a VSG, and keep settling time and overshoot fixed as SCR changes. The small-signal derivation in Section II and the Appendix is internally consistent, and the gain formulas (10)-(13) follow from choosing Ts=1 s and critical damping. The ANN is also tested on holdout SCR values (8 and 20), which is a genuine generalization check. That part is worth building on.\n\nThe problem is that the numbers don't line up. The derived law gives Dp*Kip = 8 and Kiq*(D + Dq) = 4. Table II lists Dp = 2.087e3, Kip = 0.00767, so Dp*Kip = 16.0, and with Dq = 0.687, D = 68.7, Kiq*(D + Dq) = 7.98. The text says the AVSG uses the same values as the CVSG in the SCR=2 interval, so these are the scheduled gains at the baseline condition. No operating point reconciles those products with the design equations. If Table II reflects the simulation, the implemented scheduler is not the one described; if the equations are correct, the reported settling times and overshoot values are not the ones designed. Either way, the central claim that the proposed gain law maintains performance is not supported by the manuscript as written. This is a load-bearing inconsistency, not a typo-level detail.\n\nThere are also smaller soft spots. The online computation of A and D requires Vpcc0 and delta0, but the paper does not explain how these are obtained or updated; equation (10)-(11) assume they are known. The evidence is entirely simulation, with the ANN trained on data from the same Simulink environment, so real-world validity is untested. The claims about oscillation reduction are a bit stronger than what the figures show, but that is minor. The code link is also malformed.\n\nThe reader's take was more optimistic than mine: the conditional verdict is fair, but I'd push the inconsistency harder. It is not a minor revision issue; the authors need to either correct the table or revise the design equations. That said, the paper deserves a serious referee because the idea is relevant and the derivation is otherwise careful. I'd send it to review, but the referee should be asked to check the parameter consistency explicitly.\n\nFor whom: VSG and grid-impedance estimation researchers. A reading group could learn a lot about why verifying claimed designs against reported parameters matters. I wouldn't cite it until the numbers are fixed.","headline":"Useful idea, but the reported gains don't match the derived scheduling law, so the validation is on shaky ground.","tokens_in":8266,"tokens_out":2520,"would_cite":false,"duration_ms":26997,"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":"This paper claims that a virtual synchronous generator can keep a fixed 1-second settling time across weak, strong, and stiff grids when a neural network estimates grid impedance online and a Jacobian-based gain-scheduling law retunes the…","keywords":["virtual synchronous generator","grid impedance estimation","artificial neural network","gain scheduling","weak grid stability","small-signal analysis","grid strength","inverter control"],"falsifier":"Run the same 60-second scenario with the gain-scheduling block fed the true grid impedance instead of the ANN estimates; if settling time or overshoot still drifts from the designed 1-second and zero-overshoot targets, the scheduling law itself is the cause. Separately, freeze $V_{pcc0}$ and $\\delta_0$ at their initial values while the ANN tracks the actual impedance changes; if the closed-loop response then degrades, the Jacobian's dependence on the operating point is the limiting assumption.","tokens_in":7197,"feed_emoji":"⚡","tokens_out":5190,"duration_ms":55485,"temperature":0.7,"pith_summary":"The paper tries to establish that the main weakness of virtual synchronous generators, degrading damping as the grid gets stronger, can be removed by retuning the controller gains online using an artificial neural network that estimates grid impedance from one cycle of PCC voltage and current measurements. If true, the same VSG hardware could keep its designed transient response without re-tuning when grid conditions change, because the control parameters would track impedance automatically. The authors support this with a small-signal derivation tying a 1-second settling time and critical damping to the power-flow Jacobian elements, and with a 60-second simulation that steps the grid from weak to strong to stiff while the fixed-gain VSG oscillates and the adaptive VSG does not. The load-bearing premise is that the scheduling formulas can be evaluated online, but the paper does not fully explain how the operating-point voltage and angle are obtained.","feed_headline":"ANN retunes VSG gains to stay stable as grid stiffens","feed_subtitle":"Adaptive scheduling holds a 1-second settling time from weak to stiff grids while fixed-gain VSGs oscillate.","key_machinery":"The matching identity that carries the argument is the Jacobian-based gain-scheduling law (10)-(11), derived from the closed-loop transfer functions. The active-power closed loop is a second-order system $G^P_{cl}(s) = K_{ip}A/(s^2 + D_p K_{ip} s + K_{ip} A)$; choosing $\\zeta=1$ and $\\omega_n=4$ forces $D_p = A/2$ and $K_{ip}=16/A$. The reactive loop is first-order with steady-state error $D_q/(D_q + D)$, giving $D_q=D/100$ and $K_{iq}=4/(D+D_q)$. The ANN supplies the estimated grid resistance and inductance, which enter $A$ and $D$ together with the operating-point quantities $V_{pcc0}$ and $\\delta_0$, and the control architecture buffers 200 samples of PCC voltage and current per cycle before feeding the normalized vector to the network.","core_discovery":"The central claim is that online grid-impedance estimates, produced by a feedforward ANN fed with buffered samples of $v_{pcc}$ and $i_{pcc}$, can drive an adaptive gain-scheduling law that keeps the VSG's active- and reactive-power loops at a designated damping ratio and settling time regardless of grid strength. The gains $D_p$, $K_{ip}$, $D_q$, and $K_{iq}$ are recomputed from the Jacobian entries $A=\\partial P_{pcc}/\\partial\\delta$ and $D=\\partial Q_{pcc}/\\partial V_{pcc}$ evaluated at the current operating point and estimated grid impedance, so that the closed-loop active-power denominator has damping ratio $\\zeta=1$ and natural frequency $\\omega_n=4$, and the reactive loop has a 1-second settling time with negligible steady-state error. Simulation shows the fixed-gain VSG's phase margin falling from 52.12 degrees to 20.10 degrees as SCR rises from 2 to 20, while the scheduling law holds the response near a 1-second settling time across all three scenarios. The paper also reports that the ANN estimates previously unseen grid strengths with about 0.02 seconds of delay, and notes that a nonzero overshoot remains because active-reactive coupling terms are omitted in the controller design.","pith_inferences":["Editorial extension: if $V_{pcc0}$ and $\\delta_0$ are tracked online rather than treated as constants, the same scheduling law could also adapt to load changes and operating-point shifts, not just to grid impedance changes.","Editorial extension: the transient peak errors the paper reports for previously unseen SCR values could likely be reduced by adding VSG control states or parameters to the ANN input vector, since the paper itself identifies their omission as the cause of those errors.","Editorial extension: a hardware-in-the-loop or field test on a scaled inverter would be the natural next verification, because the simulated validation does not exercise measurement noise, converter nonlinearities, or communication delays.","Editorial extension: training the ANN over a denser grid-strength range, or augmenting its inputs with physics-derived features, would test whether the claimed generalization to unseen impedance values holds beyond the specific SCR values studied."],"forward_implications":["A VSG can be tuned once and remain correctly tuned as grid strength changes, because the controller gains follow the estimated impedance.","Previously unseen grid strengths can be handled by the estimator with high accuracy and minimal delay, so the training data need not cover every possible operating point.","The approach avoids the disturbance and delay of injection-based impedance estimation while still supporting real-time gain scheduling.","The active- and reactive-power responses keep the intended 1-second settling time across weak, strong, and stiff grids, whereas fixed-gain operation becomes increasingly oscillatory as SCR increases."],"supporting_citations":[{"why":"Provides the Jacobian linearization method for the power-flow equations and the phase-margin analysis showing that a stronger grid reduces VSG phase margin.","marker":"[2]"},{"why":"Supplies the VSG topology and the small-signal power-loop parameter design procedure that the adaptive gain-scheduling law builds on.","marker":"[5]"},{"why":"Defines the synchronverter/VSG concept whose dynamic behavior the controller is designed to preserve.","marker":"[6]"},{"why":"The prior online impedance-estimation adaptive control for VSGs, used as the comparison baseline for estimation delay and performance.","marker":"[7]"},{"why":"The extended-Kalman-filter impedance estimator whose tuning sensitivity motivates the switch to an ANN approach.","marker":"[8]"},{"why":"The ANN grid-impedance identification method that this paper extends from grid-following inverters to VSG gain scheduling.","marker":"[9]"}],"fun_headline_variants":["ANN-based VSG gain scheduler adapts to grid impedance changes","Adaptive VSG gain scheduling via ANN grid impedance estimates","Neural network tunes VSG gains for weak and strong grids","ANN-driven gain scheduling keeps VSG stable as grid changes","VSG adaptive gains from ANN impedance estimates for dynamic grids"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The scheduling formulas depend on the steady-state PCC voltage and angle being known and correct at every retune, but the paper never explains how those operating-point values are measured, estimated, or updated online.","fun_headline_variants_meta":{"raw":{"variants":["ANN-based VSG gain scheduler adapts to grid impedance changes","Adaptive VSG gain scheduling via ANN grid impedance estimates","Neural network tunes VSG gains for weak and strong grids","ANN-driven gain scheduling keeps VSG stable as grid changes","VSG adaptive gains from ANN impedance estimates for dynamic grids"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000327,"raw_usage":{"total_tokens":1840,"prompt_tokens":966,"completion_tokens":874,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":582,"completion_tokens_details":{"reasoning_tokens":791}},"tokens_in":582,"tokens_out":874,"duration_ms":7381,"temperature":1.0,"reasoning_tokens":791,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:47:06.135638+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same 60-second scenario with the gain-scheduling block fed the true grid impedance instead of the ANN estimates; if settling time or overshoot still drifts from the designed 1-second and zero-overshoot targets, the scheduling law itself is the cause. Separately, freeze $V_{pcc0}$ and $\\delta_0$ at their initial values while the ANN tracks the actual impedance changes; if the closed-loop response then degrades, the Jacobian's dependence on the operating point is the limiting assumption.","supporting_citations":[{"cited_title":"Full-state feedback power decoupling control for grid forming converter with improved stability and inertia response,","cited_arxiv_id":null,"evidence_quote":"Provides the Jacobian linearization method for the power-flow equations and the phase-margin analysis showing that a stronger grid reduces VSG phase margin."},{"cited_title":"Synchronverters: Inverters that mimic synchronous generators,","cited_arxiv_id":null,"evidence_quote":"Defines the synchronverter/VSG concept whose dynamic behavior the controller is designed to preserve."},{"cited_title":"Online grid impedance estimation-based adaptive control of virtual synchronous gen- erators considering strong and weak grid conditions,","cited_arxiv_id":null,"evidence_quote":"The prior online impedance-estimation adaptive control for VSGs, used as the comparison baseline for estimation delay and performance."},{"cited_title":"Minimal invasive equivalent grid impedance estimation in inductive–resistive power networks using ex- tended kalman filter,","cited_arxiv_id":null,"evidence_quote":"The extended-Kalman-filter impedance estimator whose tuning sensitivity motivates the switch to an ANN approach."},{"cited_title":"Artificial neural network-based intelligent grid impedance identification method for grid-connected inverter,","cited_arxiv_id":null,"evidence_quote":"The ANN grid-impedance identification method that this paper extends from grid-following inverters to VSG gain scheduling."}],"review_version":1}