CoChat
Sign Up
Cited by 0Optimal Power Flow DistributionMicrogrid Control and OptimizationElectric Vehicles and Infrastructure

Physics-Informed AI-Based EV smart inverter control for two-stage Volt/Var coordination in renewable-rich active distribution networks

Open access
Zhong, Yuan, Qiu, Jing, Yang, Yi, Zhang, Chenxi, Tao, Yuechuan, Sun, Xianzhuo, Lai, Shuying·2026·Energy and AI
DOI 10.1016/j.egyai.2026.100869
Open this paper in CoChat — free accountOpen PDF

Abstract

High renewable penetration is reshaping the operation of distribution networks by increasing the variability, uncertainty, and spatial dispersion of power injections. In particular, photovoltaic (PV) generation introduces rapid voltage fluctuations, while electric vehicles (EVs) create flexible but highly time-varying demand. These changes fundamentally challenge conventional voltage regulation mechanisms in modern distribution systems. In modern smart grids, the integration of distributed energy resources (DERs), such as PV systems and EVs, has introduced new challenges in voltage regulation and power quality. As the penetration of inverter-based resources increases, maintaining voltage stability becomes more complex due to the variability and dynamic behavior of these resources. Traditional voltage–var control (VVC) methods, including centralized optimization, local droop control, and distributed schemes, are often insufficient to address the fast-evolving and decentralized nature of such systems. This paper proposes a two-stage coordinated VVC framework that integrates centralized day-ahead optimization with physics-informed AI-based local controllers for EV inverters in active distribution networks. Unlike conventional droop-based methods that rely on predefined low-dimensional control curves and centralized optimization approaches that require intensive communication and computation, the proposed framework enables fast local voltage regulation while preserving grid-level coordination. In Stage I, an AC branch-flow-based optimization coordinates OLTCs, capacitor banks, and PV/EV inverter set-points to obtain voltage-secure operating schedules. In Stage II, a residual multi-layer perceptron controller embedded in each EV inverter learns nonlinear reactive power correction policies from OPF-based solutions using only local measurements and scheduling information. Instead of conventional low-dimensional droop curves (Q(V), Q(P)), each EV inverter implements a residual multi-layer perceptron that maps rich local feature vectors—including voltage, active power, state-of-charge, tariff signals, and day-ahead set-points—to reactive power corrections. The network output is parameterized to satisfy inverter apparent-power limits by construction. The controller is trained offline using optimal set-points obtained from day-ahead or real-time optimal power flow, augmented with a physics-informed voltage penalty derived from the power-flow equations and admissible voltage bounds.

AI snapshot

Written from the paper’s full text.

TL;DR

The paper proposes a two-stage Volt/Var control framework combining day-ahead centralized AC branch-flow optimization with a physics-informed ResMLP local controller for EV smart inverters. On an IEEE 33-bus test system, the proposed controller approaches centralized VVC performance while reducing computation from 273 s to less than 1 s.

Key findings

  • The proposed ResMLP reactive-power correction predictor achieved RMSE = 0.355, MAE = 0.249, and R² = 0.936, outperforming LSTM (0.472, 0.331, 0.911), CNN (0.498, 0.352, 0.891), and LSTM+CNN (0.401, 0.287, 0.919).
  • In VVC case comparison, the proposed method achieved power loss = 1.779, maximum voltage deviation = 0.0069, voltage violation rate = 0.012%, and computational time < 1 s, versus centralized VVC (1.762, 0, 0, 273 s), QP curve (1.892, 0.0107, 0.23%, < 1 s), and QV curve (1.832, 0.0083, 0.16%, < 1 s).
  • The capability-constrained output layer enforces the inverter apparent power limit by construction, and the physics-informed training loss adds a voltage-security penalty to the standard regression loss.
  • Training used 10,000 operating scenarios with 24-hour profiles, resulting in 240,000 samples on an IEEE 33-bus system with aggregated EVs at buses 5, 9, 17, 18, and 33; the dataset was split 80% training and 20% testing.

Methods

  • Two-stage framework: Stage I performs day-ahead hourly AC branch-flow optimization to schedule OLTC taps, capacitor-bank statuses, and PV/EV inverter reactive power setpoints; Stage II uses real-time local EV inverter control to adjust those day-ahead references.
  • The local controller is a residual multi-layer perceptron (ResMLP) mapping local features—voltage, active power, EV state-of-charge, connection status, tariff signals, and day-ahead reactive power references—to reactive power corrections, with an output parameterization that enforces inverter capability limits.
  • Offline AC OPF simulations generate optimal real-time reactive power corrections as supervision labels; the training objective combines regression loss with a differentiable voltage-security penalty derived from power-flow equations, and the method is evaluated on the IEEE 33-bus system against centralized VVC, QP-curve, QV-curve, CNN, LSTM, and LSTM+CNN baselines.

Limitations

  • The paper states that future work will investigate scalability for large-scale distribution networks with massive EV penetration and heterogeneous distributed resources, including computational scalability, communication-efficient coordination, and distributed implementation under practical operating conditions.
  • Evident from the text: validation is simulation-only on a single IEEE 33-bus test system, with no field deployment or real-world data, so external validity to other networks is not demonstrated.
  • Evident from the text: the local controller is trained offline using OPF-derived labels and depends on the day-ahead scheduling/forecast framework; the paper does not quantify performance under severe forecast errors or communication/coordination failures.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1016/j.egyai.2026.100869

Cites

25 of 25 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 23 cited sources checked · 0 supported · 0 flagged

  • Findingreasoning

    Combining LSTM and CNN further improves accuracy.

    “Combining them (LSTM+CNN) further improves accuracy”
    Quote in the paperNot a substantial claim
  • Findingown data

    The Distribution Comparison between predicted and true values reveals strong overlap in probability density.

    “The Distribution Comparison between predicted and true values reveals strong overlap in probability density”
    Quote in the paperSubstantial claim
  • Findingown data

    The Residual Analysis plot does not exhibit a pronounced systematic trend across predicted values.

    “The Residual Analysis plot does not exhibit a pronounced systematic trend across predicted values”
    Quote in the paperSubstantial claim
  • Findingown data

    Centralized VVC yields the smoothest voltage surface, with voltages tightly clustered around 1.0 p.u. across both buses and time, reflecting strong system-wide coordination.

    “Centralized VVC yields the smoothest voltage surface, with voltages tightly clustered around 1.0 p.u. across both buses and time, reflecting strong system-wide coordination.”
    Quote in the paperSubstantial claim
  • Findingown data

    Reactive power output from EV aggregation varies throughout the day, providing additional voltage support independent of active power exchange.

    “Reactive power output varies throughout the day, providing additional voltage support independent of active power exchange.”
    Quote in the paperSubstantial claim
  • May combine two claims. 38
  • Classified as probably not a substantial claim. 7

  • May combine two claims.

    Findingown data

    The proposed method produces a noticeably smoother and more uniform voltage profile than the curve-based baselines, with fewer low-voltage regions and reduced spatio-temporal variability, approaching the centralized solution while maintaining constraint compliance.

    “The proposed method produces a noticeably smoother and more uniform voltage profile than the curve-based baselines, with fewer low-voltage regions and reduced spatio-temporal variability, approaching the centralized solution while maintaining constraint compliance.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The OLTC schedule shows stepwise voltage variation indicating discrete tap switching in response to time-varying load conditions; in early hours the OLTC lowers voltage for reduced demand, while during daytime and peak periods it increases tap position to compensate for higher loading and maintain adequate voltage levels.

    “The stepwise variation of voltage indicates discrete tap switching actions in response to time-varying load conditions.In the early hours, the OLTC lowers the voltage to accommodate reduced demand, while during daytime and peak periods, it increases the tap position to compensate for higher loading and maintain adequate voltage levels.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The proposed method achieves the best results (RMSE = 0.355, MAE = 0.249, R2 = 0.936), outperforming all baselines and demonstrating higher prediction accuracy and stronger generalization.

    “The proposed method achieves the best results (RMSE = 0.355, MAE = 0.249, R² = 0.936), outperforming all baselines and demonstrating higher prediction accuracy and stronger generalization.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Most operating points remain within the allowable capacity margin, and only a small fraction approaches constraint boundaries.

    “Most operating points remain within the allowable capacity margin, and only a small fraction approaches constraint boundaries.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The centralized VVC achieves the lowest power loss (1.762) and completely eliminates voltage deviation and violation, but requires significantly longer computational time (273 seconds).

    “The centralized VVC achieves the lowest power loss (1.762) and completely eliminates voltage deviation and violation; however, it requires significantly longer computational time (273 seconds), limiting its real-time applicability.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingreasoning

    CNN shows the weakest performance, while LSTM performs better by capturing temporal dependencies.

    “CNN shows the weakest performance, while LSTM performs better by capturing temporal dependencies.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The simulated voltages are predominantly confined within the admissible operating band, with extremely low violation rates.

    “The simulated voltages are predominantly confined within the admissible operating band, and the extremely low violation rates demonstrate that the learned control actions are not only statistically accurate but also physically effective.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    During low-demand periods, the capacitor bank remains partially or fully disconnected, resulting in low or zero reactive power output; as load increases, the bank is progressively switched in, providing higher reactive power support, with peak injection levels observed during daytime and evening high-load intervals.

    “During low-demand periods, the capacitor bank remains partially or fully disconnected, resulting in low or zero reactive power output.As load increases, the bank is progressively switched in, providing higher reactive power support, with peak injection levels observed during daytime and evening high-load intervals.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The aggregated EV operation shows positive active power in early hours indicating net charging and increasing energy storage state; around midday charging power decreases and approaches balance; in afternoon and early evening active power becomes negative, reflecting V2G discharging that supports the grid during higher demand periods.

    “The positive active power in the early hours indicates net charging behavior, during which the energy storage state increases steadily.Around midday, the charging power decreases and approaches balance, while in the afternoon and early evening, the active power becomes negative, reflecting V2G discharging that supports the grid during higher demand periods.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The QP and QV curve-based methods reduce computational burden to less than 1 second but have higher power loss and noticeable voltage deviations, with violation rates of 0.23% and 0.16%, respectively.

    “The QP and QV curve-based methods greatly reduce computational burden (<1 second), but at the expense of higher power loss and noticeable voltage deviations, with violation rates of 0.23% and 0.16%, respectively.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The proposed method achieved a power loss of 1.779, a maximum voltage deviation of 0.0069, and a voltage violation rate of 0.012%, approaching the centralized VVC solution while reducing computational time from 273 s to less than 1 s.

    “In terms of voltage regulation performance, the proposed method achieves a power loss of 1.779, a maximum voltage deviation of 0.0069, and a voltage violation rate of only 0.012%, approaching the centralized VVC solution while reducing the computational time from 273 s to less than 1 s.”
    • Also stated in: § Multi-dimension · plain
      “The proposed method achieves a favorable trade-off.It maintains low power loss (1.779), close to the centralized optimal solution, while significantly reducing voltage deviation (0.0069) and maintaining an extremely low violation rate (0.012%), all with sub-second computational time.”
    • Also stated in: § Multi-dimension · plain
      “Proposed 1.779 0.0069 0.012% <1 second”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The proposed learning algorithm achieved RMSE 0.355, MAE 0.249, and R2 0.936, outperforming LSTM, CNN, and LSTM+CNN on these metrics.

    “TABLE I PERFORMANCE COMPARISON OF DIFFERENT LEARNING ALGORITHMS RMES MAE R2 LSTM 0.472 0.331 0.911 CNN 0.498 0.352 0.891 LSTM+CNN 0.401 0.287 0.919 Proposed 0.355 0.249 0.936”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The close agreement between predicted and real EV inverter reactive power curves indicates high prediction accuracy and effective tracking performance under time-varying operating conditions.

    “The close agreement between the predicted and real curves indicates high prediction accuracy and effective tracking performance under time-varying operating conditions.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    Centralized optimization-based VVC formulations can incorporate detailed models of device capabilities and network constraints, and optimize for objectives such as loss minimization, voltage profile improvement, or conservation voltage reduction.

    “Such formulations can incorporate detailed models of device capabilities and network constraints, and optimize for objectives such as loss minimization, voltage profile improvement, or conservation voltage reduction.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [9] Hou, 2024, Voltage regulation enhanced hierarchical coordinated volt/var and volt/watt control for active distribution networks with soft open points, IEEE Transactions on Sustainable Energy source
    • cites [10] Ma, 2021, A centralized voltage regulation method for distribution networks containing high penetrations of photovoltaic power, International Journal of Electrical Power & Energy Systems source

  • May combine two claims.

    Premisecitation

    Inverter-based resources are capable of providing much faster dynamic support; however, their proper utilization requires suitable control strategies that coordinate with the rest of the network.

    “Inverter-based resources are capable of providing much faster dynamic support; however, their proper utilization requires suitable control strategies that coordinate with the rest of the network [8].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [8] Hasheminasab, 2024, A review of control techniques for inverter-based distributed energy resources applications, Energies source

  • Classified as probably not a substantial claim.

    Premisecitation

    Distributed VVC approaches aim to reduce the communication and computational burden of fully centralized OPF while still exploiting network information and coordination among devices.

    “These methods aim to reduce the communication and computational burden of fully centralized OPF while still exploiting network information and coordination among devices.”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites [15] Li, 2022, Robust data-driven and fully distributed volt/var control for active distribution networks with multiple virtual power plants, IEEE Transactions on Smart Grid source
    • cites [16] Gebbran, 2022, Multiperiod der coordination using admm-based three-block distributed ac optimal power flow considering inverter volt-var control, IEEE Transactions on Smart Grid source
    • cites [17] Li, 2023, A machine learning-assisted distributed optimization method for inverter-based volt-VAR control in active distribution networks, IEEE Transactions on Power Systems source

  • May combine two claims.

    Premisecitation

    Hybrid VVC approaches can substantially reduce computation and communication requirements compared with fully centralized OPF, while achieving better voltage regulation than fixed droop control.

    “Overall, hybrid methods can substantially reduce computation and communication requirements compared with fully centralized OPF, while achieving better voltage regulation than fixed droop control”
    Quote in the paperSubstantial claim0 of 7 sources checked
    • cites [18] Sun, 2022, Data-driven combined central and distributed volt/var control in active distribution networks, IEEE Transactions on Smart Grid source
    • cites [19] Xu, 2021, Multi-objective hierarchically-coordinated volt/var control for active distribution networks with droop-controlled PV inverters, IEEE Transactions on Smart Grid source
    • cites [20] Li, 2025, A dual time scale Voltage/Var control method for active distribution networks based on data-driven, Electrical Engineering source
    • cites [21] Sun, 2021, Optimal local volt/var control for photovoltaic inverters in active distribution networks, IEEE Transactions on Power systems source
    • cites [22] Cao, 2023, Physics-informed Graphical Representation-enabled Deep Reinforcement Learning for Robust Distribution System Voltage Control, IEEE Transactions on Smart Grid source
    • cites [23] Liu, 2022, Federated reinforcement learning for decentralized voltage control in distribution networks, IEEE Transactions on Smart Grid source
    • cites [24] Cao, 2024, Decentralized Graphical-Representation-Enabled Multi-Agent Deep Reinforcement Learning for Robust Control of Cyber-Physical Systems, IEEE Transactions on Reliability source

  • May combine two claims.

    Premisenone given

    The performance of hybrid VVC methods depends on the update rate of the supervisory layer, the fidelity of network models/sensitivities, and the ability to maintain feasibility under forecast errors and fast operational changes.

    “nevertheless, their performance depends on the update rate of the supervisory layer, the fidelity of network models/sensitivities, and the ability to maintain feasibility under forecast errors and fast operational changes.”
    Quote in the paperSubstantial claim0 of 7 sources checked
    • cites [18] Sun, 2022, Data-driven combined central and distributed volt/var control in active distribution networks, IEEE Transactions on Smart Grid source
    • cites [19] Xu, 2021, Multi-objective hierarchically-coordinated volt/var control for active distribution networks with droop-controlled PV inverters, IEEE Transactions on Smart Grid source
    • cites [20] Li, 2025, A dual time scale Voltage/Var control method for active distribution networks based on data-driven, Electrical Engineering source
    • cites [21] Sun, 2021, Optimal local volt/var control for photovoltaic inverters in active distribution networks, IEEE Transactions on Power systems source
    • cites [22] Cao, 2023, Physics-informed Graphical Representation-enabled Deep Reinforcement Learning for Robust Distribution System Voltage Control, IEEE Transactions on Smart Grid source
    • cites [23] Liu, 2022, Federated reinforcement learning for decentralized voltage control in distribution networks, IEEE Transactions on Smart Grid source
    • cites [24] Cao, 2024, Decentralized Graphical-Representation-Enabled Multi-Agent Deep Reinforcement Learning for Robust Control of Cyber-Physical Systems, IEEE Transactions on Reliability source

  • May combine two claims.

    Interpretationreasoning

    During real-time implementation, each EV inverter only performs local inference using the trained ResMLP controller, without requiring repeated OPF calculations or extensive communication with a central controller.

    “During real-time implementation, each EV inverter only performs local inference using the trained ResMLP controller.The online procedure consists of feature acquisition, neural network forward propagation, and capability-constrained reactive power adjustment, without requiring repeated OPF calculations or extensive communication with a central controller.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    The proposed AI-based local controller considers both real-time system states and day-ahead scheduling information, unlike conventional local controllers that only rely on instantaneous voltage or active power measurements.

    “Different from conventional local controllers that only rely on instantaneous voltage or active power measurements, the proposed AI-based local controller considers both real-time system states and day-ahead scheduling information.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    The proposed approach effectively balances optimality, voltage regulation performance, and computational efficiency, making it suitable for practical real-time voltage control applications.

    “These results demonstrate that the proposed approach effectively balances optimality, voltage regulation performance, and computational efficiency, making it suitable for practical real-time voltage control applications.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    The learned control actions are not only statistically accurate but also physically effective.

    “The simulated voltages are predominantly confined within the admissible operating band, and the extremely low violation rates demonstrate that the learned control actions are not only statistically accurate but also physically effective.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationnone given

    The controller is implemented by a residual neural network that maps rich local feature vectors to real-time reactive power corrections, while respecting inverter capability limits by construction.

    “The controller is implemented by a residual neural network that maps rich local feature vectors to real-time reactive power corrections, while respecting inverter capability limits by construction.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    The learned policy is not solely based on statistical correlations but follows the fundamental Volt/Var regulation principle, where reactive power is adjusted according to system voltage conditions while respecting inverter limitations.

    “Therefore, the learned policy is not solely based on statistical correlations but follows the fundamental Volt/Var regulation principle, where reactive power is adjusted according to system voltage conditions while respecting inverter limitations.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Interpretationnone given

    The residual structure improves training stability and enables effective representation of the nonlinear relationships among voltage, power, EV state-of-charge, and control references.

    “The residual structure improves training stability and enables effective representation of the nonlinear relationships among voltage, power, EV state-of-charge, and control references.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    The ResMLP is selected because the proposed controller aims to learn a local nonlinear mapping between inverter operating conditions and optimal reactive power corrections, rather than reconstructing network-wide states.

    “The residual multi-layer perceptron (ResMLP) is selected because the proposed controller aims to learn a local nonlinear mapping between inverter operating conditions and optimal reactive power corrections, rather than reconstructing network-wide states.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    The model reproduces not only pointwise values but also the statistical characteristics of reactive power behavior.

    “confirming that the model reproduces not only pointwise values but also the statistical characteristics of reactive power behavior.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Interpretationnone given

    Such relationships are difficult to represent using fixed low-dimensional control curves, especially under uncertainties caused by renewable generation fluctuations, load variations, and EV availability.

    “Such relationships are difficult to represent using fixed low-dimensional control curves, especially under the uncertainties caused by renewable generation fluctuations, load variations, and EV availability.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    Predefined local curves are less adaptive to changing network conditions and may leave residual undervoltage pockets.

    “indicating that predefined local curves are less adaptive to changing network conditions and may leave residual undervoltage pockets.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    The learning process respects physical limits and avoids generating infeasible reactive power commands, reflecting good integration between data-driven learning and engineering constraints.

    “This indicates that the learning process respects physical limits and avoids generating infeasible reactive power commands, reflecting good integration between data-driven learning and engineering constraints.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    The voltage is consistently maintained within the acceptable operational range, demonstrating that the OLTC effectively performs discrete voltage regulation and ensures system voltage stability under dynamic conditions.

    “The voltage is consistently maintained within the acceptable operational range, demonstrating that the OLTC effectively performs discrete voltage regulation and ensures system voltage stability under dynamic conditions.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    The Stage II real-time controller treats day-ahead scheduling results as reference values rather than fixed constraints and compensates for deviations caused by forecast errors and real-time uncertainties.

    “The Stage II real-time controller treats the day-ahead scheduling results as reference values rather than fixed constraints and compensates for deviations caused by forecast errors and real-time uncertainties.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    Because the voltage penalty is constructed via the power-flow operator, the resulting policy remains consistent with the network equations and can generalize to unseen operating conditions where the interplay between voltage magnitudes, reactive power, and network topology is nontrivial.

    “Second, because the voltage penalty is constructed via the power-flow operator, the resulting policy remains consistent with the network equations and can generalize to unseen operating conditions where the interplay between voltage magnitudes, reactive power, and network topology is nontrivial.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    The proposed controllers can be interpreted as learned multi-dimensional control hypersurfaces that generalize and significantly enrich traditional Q(V)/Q(P) droop curves, while remaining fully local and communication-free at run time.

    “Thus, the proposed controllers can be interpreted as learned multi-dimensional control hypersurfaces that generalize and significantly enrich traditional 𝑄(𝑉)/𝑄(𝑃) droop curves, while remaining fully local and communication-free at run time.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    Overall, the multi-faceted evaluation demonstrates that the proposed AI learning framework achieves high regression accuracy, stable error characteristics, operational constraint compliance, and strong temporal generalization, validating its suitability for data-driven reactive power correction in EV-integrated distribution systems.

    “Overall, the multi-faceted evaluation demonstrates that the proposed AI learning framework achieves high regression accuracy, stable error characteristics, operational constraint compliance, and strong temporal generalization, validating its suitability for data-driven reactive power correction in EV-integrated distribution systems.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    Considering the uncertainty of EV participation, the proposed controller incorporates EV connection status and state-of-charge information into the local feature vector, so plug-in/plug-out behaviors are considered by adapting reactive power regulation according to actual availability of EV resources.

    “Considering the uncertainty of EV participation, the proposed controller incorporates EV connection status and state-of-charge information into the local feature vector.Therefore, plug-in/plug-out behaviors are considered by adapting the reactive power regulation according to the actual availability of EV resources.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    The proposed framework is designed with a hierarchical architecture to satisfy practical real-time control requirements.

    “The proposed framework is designed with a hierarchical architecture to satisfy practical real-time control requirements.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Interpretationown data

    The figure highlights the bidirectional power capability and temporal coordination of EV aggregation, enabling dynamic energy balancing and grid support while maintaining acceptable energy state levels.

    “Overall, the figure highlights the bidirectional power capability and temporal coordination of EV aggregation, enabling dynamic energy balancing and grid support while maintaining acceptable energy state levels.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    The high R2 value (0.936) confirms that most of the output variance is explained by the learned model, indicating strong representational capacity and effective feature extraction.

    “The high R² value (0.936) confirms that most of the output variance is explained by the learned model, indicating strong representational capacity and effective feature extraction.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    The proposed framework integrates day-ahead centralized scheduling with physics-informed AI-based local control, where the optimization layer provides coordinated operating references and the AI controller enables fast adaptive voltage regulation under uncertain operating conditions.

    “The proposed framework integrates day-ahead centralized scheduling with physics-informed AI-based local control, where the optimization layer provides coordinated operating references and the AI controller enables fast adaptive voltage regulation under uncertain operating conditions.”
    • Also stated in: § Multi-dimension · plain
      “This paper proposes a two-stage coordinated VVC framework that integrates centralized optimization with physics-informed, AI-based local control to manage the active distribution network, shown in Fig. 1.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    The model possesses strong dynamic tracking capability and generalizes well over time, rather than merely fitting static data patterns.

    “This suggests that the model possesses strong dynamic tracking capability and generalizes well over time, rather than merely fitting static data patterns.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    The capacitor bank operates as a staged reactive compensation device, effectively supporting voltage regulation and reducing reactive power deficits under varying load conditions.

    “The discrete multi-level pattern confirms that the capacitor bank operates as a staged reactive compensation device, effectively supporting voltage regulation and reducing reactive power deficits under varying load conditions.”
    Quote in the paperSubstantial claim
42 of 42

  • The paper relies on this without support.

    EV charging can lead to rapidly varying and spatially clustered loads, aggravating voltage variations and thermal loading in low-and medium-voltage feeders.

    “EV charging, in particular, can lead to rapidly varying and spatially clustered loads, aggravating voltage variations and thermal loading in low-and medium-voltage feeders.”

  • The paper relies on this without support.

    OLTCs and CBs operate on relatively slow time scales (minutes to hours) and are primarily scheduled based on anticipated loading patterns.

    “OLTCs and CBs operate on relatively slow time scales (minutes to hours) and are primarily scheduled based on anticipated loading patterns.”

  • The paper relies on this without support.

    With increasing DER penetration and uncertainty, classical rule-based VVC mechanisms become insufficient to ensure secure and efficient operation under a wide range of operating conditions.

    “With increasing DER penetration and uncertainty, classical rule-based VVC mechanisms become insufficient to ensure secure and efficient operation under a wide range of operating conditions.”

  • The paper relies on this without support.

    ResMLP offers good expressive power and stable training for tabular inputs.

    “We implement the mapping from local inputs to real-time reactive power correction using a residual multi-layer perceptron (ResMLP), which has been shown to offer good expressive power and stable training for tabular inputs.”

  • The authors state this as an assumption.

    Each real-time interval τ belongs to a coarse day-ahead period t(τ) in the day-ahead scheduling set T_DA.

    “We assume that each real-time interval 𝜏 belongs to a coarse day-ahead period 𝑡(𝜏) ∈ 𝒯 𝐷𝐴 .”
5 of 5

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • Case 3 is assigned two different control curves

    Minor

    Case 3 is described as using a Q(Q) curve, but the results identify Case 3 as the QV-curve method.

  • PV inverter rating is described as an EV rating

    Minor

    The EV-inverter constraint defines the symbol PV i S as the capacity of the EV, while another passage defines that same symbol as the capacity of the PV.

2 of 2

Cites

Flagged first, then by citations

Authors
Zhong, Y., Qiu, J., Yang, Y., Zhang, C., Tao, Y., Sun, X., & Lai, S.
Affiliations
  • The University of Sydney (AU)
  • State Grid Fujian Electric Power Company (China) (CN)
  • Fuzhou University (CN)
  • City University of Hong Kong (HK)
  • Shandong University (CN)
  • City University of Macau (MO)
Corresponding author
Tao, Yuechuan
ORCID
Venue
Energy and AI
Publisher
Elsevier BV
Published
2026-12
Volume
26
Pages
100869
ISSN
2666-5468 (ISSN-L)
DOI
10.1016/j.egyai.2026.100869
OpenAlex
W7202398842
Cited by
0
References
23
FWCI
0.00
Type
journal article
Language
en
Open access
Licence
Locations
  • Energy and AI (publishedVersion, cc-by-nc-nd)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion)
Funders
  • Australian Research Council (LP200100056, IC200100023, DP220103881, IH180100020)
  • Huawei Technologies Co Ltd (9229217)
Topics
  • Physical Sciences › Engineering › Electrical and Electronic Engineering › Optimal Power Flow Distribution
  • Physical Sciences › Engineering › Control and Systems Engineering › Microgrid Control and Optimization
  • Physical Sciences › Engineering › Electrical and Electronic Engineering › Electric Vehicles and Infrastructure
Keywords
physics-informed neural networks, multilayer perceptron, day-ahead optimization, optimal power flow, voltage regulation, photovoltaic generation, electric vehicles, active distribution networks, reactive power control, droop control, capacitor bank, state of charge, distributed energy resources

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-28.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-28.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review