CoChat
Sign Up
Cited by 15Optimal Power Flow DistributionMicrogrid Control and OptimizationPower Quality and Harmonics

Adaptive Volt–Var Control in Smart PV Inverter for Mitigating Voltage Unbalance at PCC Using Multiagent Deep Reinforcement Learning

Open access
Jung, Yoongun, Han, Changhee, Lee, Dongwon, Song, Sungyoon, Jang, Gilsoo·2021·Applied Sciences
DOI 10.3390/app11198979
Open this paper in CoChat — free accountOpen PDF

Abstract

Modern distribution networks face an increasing number of challenges in maintaining balanced grid voltages because of the rapid increase in single-phase distributed generators. Because of the proliferation of inverter-based resources, such as photovoltaic (PV) resources, in distribution networks, a novel method is proposed for mitigating voltage unbalance at the point of common coupling by tuning the volt–var curve of each PV inverter through a day-ahead deep reinforcement learning training platform with forecast data in a digital twin grid. The proposed strategy uses proximal policy optimization, which can effectively search for a global optimal solution. Deep reinforcement learning has a major advantage in that the calculation time required to derive an optimal action in the smart inverter can be significantly reduced. In the proposed framework, multiple agents with multiple inverters require information on the load consumption and active power output of each PV inverter. The results demonstrate the effectiveness of the proposed control strategy on the modified IEEE 13 standard bus systems with time-varying load and PV profiles. A comparison of the effect on voltage unbalance mitigation shows that the proposed inverter can address voltage unbalance issues more efficiently than a fixed droop inverter.

AI snapshot

Written from the paper’s full text.

TL;DR

The paper proposes a multiagent deep reinforcement learning method using proximal policy optimization to adaptively tune each PV inverter’s volt–var curve for mitigating voltage unbalance at the point of common coupling. Simulations on modified IEEE 13- and 34-bus systems show the proposed adaptive control reduces average voltage unbalance factor compared with fixed volt–var and no reactive-power compensation.

Key findings

  • The proposed method (Case 1) achieved lower average VUF than fixed volt–var control (Case 2) and no reactive-power compensation (Case 3) across sunny and cloudy conditions. On a sunny day, average VUF was 1.23% (IEEE 13) and 0.05% (IEEE 34) for Case 1, versus 1.38% and 0.0514% for Case 2, and 1.37% and 0.52% for Case 3.
  • On a cloudy day, average VUF was 1.29% (IEEE 13) and 0.053% (IEEE 34) for Case 1, versus 1.40% and 0.054% for Case 2, and 1.42% and 0.055% for Case 3.
  • The reward profile converged at approximately 2000 epochs, indicating training completion and little further neural-network variable updating.
  • Fixed volt–var control was either insignificant or worse than no reactive-power compensation in some conditions, while the proposed strategy performed well even when actual PV/load profiles differed from forecasts due to cloudiness.

Methods

  • Multiagent DRL with proximal policy optimization: each agent has an actor–critic neural network. State is the i-th PV active power and normalized load apparent power; action is the set {V1, V2, V3, V4}, with V1,V2 selected from {0.92, 0.94, 0.96, 0.98} and V3,V4 from {1.02, 1.04, 1.06, 1.08}, giving 36 action spaces. Reward is −{VUF_proposed − VUF_fixed}. Gaussian noise is added each episode for forecast and communication errors.
  • Training used 3000 episodes, each comprising 50,400 1-s intervals from 6:00 to 20:00. PV and load data were selected from historical yearly PV generation data of the Korea Power Exchange. Power flow was solved with OpenDSS via a Python COM interface, with PyTorch for reinforcement learning.
  • Case study compared three cases on modified IEEE 13- and IEEE 34-bus systems with single-phase PV systems: Case 1 proposed adaptive volt–var, Case 2 fixed volt–var values of 0.92, 0.98, 1.02, and 1.08, and Case 3 no reactive-power compensation.

Limitations

  • The paper states that it did not consider the size and connected position of PV systems, focusing instead on operation strategy after the distribution network topology was already determined.
  • The paper states that the proposed strategy compensates reactive power based on its volt–var curve rather than a mathematically calculated value, so it is impossible to make VUF zero. It also notes that inverters cannot provide or absorb reactive power when active power generation is less than 0.05 × P_rated.
  • Evident from the text: validation is simulation-based on modified IEEE 13- and 34-bus systems using OpenDSS and historical Korean PV/load data, with no field deployment or hardware-in-the-loop validation reported.

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

Integrity signals

Registry

Crossref · 10.3390/app11198979

Cites

28 of 28 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

5 of 11 cited sources checked · 0 supported · 0 flagged · accuracy not yet measured

  • Premisecitation

    Multiagent proximal policy optimization methods have stable and fast convergence ability.

    “Multiagent proximal policy optimization methods [16] were used to train multiple inverters in a distribution network owing to the stable and fast convergence ability.”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [16] Schulman, J., Wolski, F., Dhariwal, P., Radford, A., and Klimov, O. (2017). Proximal policy optimization algorithms. arXiv. source

      The Conclusion (p. 8) describes PPO as stable and reliable, and the Abstract (p. 1) reports favorable empirical sample complexity and wall-time. The paper does not establish fast convergence for multiagent PPO or study training multiple inverters in a distribution network. (LLM, conf=0.86)

      Experimental — accuracy not yet measured
  • Premisecitation

    Most strategies proposed for using PV inverters to improve phase unbalance rely on centralized cooperative control.

    “Most strategies proposed for using PV inverters to improve phase unbalance rely on centralized cooperative control to solve the economic problem.”
    Quote in the paperSubstantial claim2 of 3 sources checked
    • cites [5] Jha, 2019, Bi-level volt-var optimization to coordinate smart inverters with voltage control devices, IEEE Trans. Power Syst. source

      The Introduction (p. 2) and Conclusion (p. 11) describe this paper’s coordinated optimization of smart inverters and other voltage-control devices in a three-phase unbalanced system. They do not establish that most PV-inverter strategies use centralized cooperative control, or that their purpose is to improve phase unbalance; this paper’s stated objective is conservation voltage reduction. (LLM, conf=0.84)

      Experimental — accuracy not yet measured
    • cites [6] Nájera, J., Mendonça, H., de Castro, R.M., and Arribas, J.R. (2019). Strategies comparison for voltage unbalance mitigation in LV distribution networks using EV chargers. Electronics, 8. source

      The Introduction discusses centralized and decentralized control of EV charging, while §3.2.4 describes a decentralized inverter control for voltage unbalance. Neither passage addresses whether most PV-inverter strategies use centralized cooperative control or do so to solve an economic problem; searches for those terms found no supporting passage. (LLM, conf=0.76)

      Experimental — accuracy not yet measured
    • cites [7] Nejabatkhah, 2019, Flexible unbalanced compensation of three-phase distribution system using single-phase distributed generation inverters, IEEE Trans. Smart Grid source
  • Findingreasoning

    In the time periods 6-7 h and 19-20 h, there is no apparent difference between the VUF(%) of the three cases.

    “In Figure 11, in the time period of 6~7 h and 19~20 h, there doesn't seem to be any difference between VUF(%) of each case.”
    Quote in the paperNot a substantial claim
  • Findingown data

    The proposed inverter can address voltage unbalance issues more efficiently than a fixed droop inverter.

    “A comparison of the effect on voltage unbalance mitigation shows that the proposed inverter can address voltage unbalance issues more efficiently than a fixed droop inverter.”
    Quote in the paperSubstantial claim
  • Findingown data

    The VUF(%) of case 1 is lower than that of the other cases for each day.

    “Additionally, the VUF(%) of case 1 is lower than others for each day.”
    Quote in the paperSubstantial claim
  • Classified as probably not a substantial claim. 11
  • May combine two claims. 6

  • Classified as probably not a substantial claim.

    Premisecitation

    The digital twin provides an environment for exploration and exploitation for the learning of agents that converges to the optimal solution through various experiences.

    “The digital twin provides an environment for exploration and exploitation for the learning of agents that converges to the optimal solution through various experiences.”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [17] Fuller, 2020, Digital twin: Enabling technologies, challenges and open research, IEEE Access source

      The Manufacturing section says a digital twin can learn and monitor simultaneously while applying machine-learning algorithms. The Simulation and Artificial Intelligence section describes testing and optimizing processes in a virtual environment. Neither passage says that agents explore and exploit that environment or converge to an optimal solution through experience. (LLM, conf=0.83)

      Experimental — accuracy not yet measured

  • May combine two claims.

    Premisecitation

    Neural networks take little time (less than 1 ms) to derive the output, giving neural networks an advantage in real-time use.

    “The Neural networks (NNs) take little time (less than 1ms) to derive the output, so NN has an advantage in using real time [9].”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [9] Schmidhuber, 2015, Deep learning in neural networks: An overview, Neural Netw. source

      Section 4.5 (p. 8) discusses GPUs speeding neural-network training, not the time needed to produce an output. Searches for sub-millisecond output latency and real-time inference found no passage in our copy supporting either the less-than-1-ms figure or the claimed real-time advantage. (LLM, conf=0.75)

      Experimental — accuracy not yet measured

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

    Findingnone given

    After day-ahead training, the neural network has an optimal policy to act properly.

    “After day-ahead training, NN has an optimal policy to act properly.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingnone given

    The effect of case 2 was either insignificant or made matters worse than case 3, despite the compensation of reactive power based on the fixed volt-var curve.

    “The effect of case 2 was either insignificant or made matters worse than case 3, despite the compensation of reactive power based on the fixed volt-var curve.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingreasoning

    The reward almost converges at 2000 epochs, meaning the training is almost complete and the neural network variables are rarely updated.

    “With epochs at 2000, the reward almost converges, which means that the training is almost complete; therefore, the variables of the neural nets are rarely updated.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Findingnone given

    After training, the neural network assigns a large part of the action probability to the optimal action.

    “After the training, the NN assigns a large part of the probability to the optimal action.”
    Quote in the paperNot a substantial claim

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

    Interpretationnone given

    The -VUF_fixed term in the reward function makes the proposed strategy more effective than fixed volt-var curve inverters.

    “In addition, -VUF f ixed makes the proposed strategy more effective than fixed volt-var curve inverters.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Study design and conductnone given

    The proposed framework uses a probability model based on actual PV generations and load profiles to ensure that the inverter acts properly under the influence of uncertainty in PV outputs.

    “In the proposed framework, a probability model based on actual PV generations and load profiles is adopted to ensure that the inverter acts properly under the influence of uncertainty in PV outputs.”
    Quote in the paperSubstantial claim

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

    Interpretationnone given

    Deep reinforcement learning has a major advantage in that the calculation time required to derive an optimal action in the smart inverter can be significantly reduced.

    “Deep reinforcement learning has a major advantage in that the calculation time required to derive an optimal action in the smart inverter can be significantly reduced.”
    Quote in the paperNot a substantial claim

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

    Interpretationnone given

    The proposed strategy uses proximal policy optimization, which can effectively search for a global optimal solution.

    “The proposed strategy uses proximal policy optimization, which can effectively search for a global optimal solution.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    The fluctuating output patterns of single-phase DG and the irregularity of EV charging patterns aggravate voltage unbalance.

    “The fluctuating characteristic output patterns of single-phase DG and the irregularity of EV charging patterns [1,2] aggravated the voltage unbalance.”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [1] Putrus, G.A., Suwanapingkarl, P., Johnston, D., Bentley, E.C., and Narayana, M. (2009, January 7–10). Impact of electric vehicles on power distribution networks. Proceedings of the 2009 IEEE Vehicle Power and Propulsion Conference, Dearborn, MI, USA. source
    • cites [2] Balamurugan, 2012, Impact of distributed generation on power distribution system, Energy Procedia source

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

    Premisenone given

    Once a PV system is installed by its owner in the distribution system, its size and connected position cannot change at all.

    “Once the PV system installed by its owner is in the distribution system, the size and the connected position of the PV system cannot change at all.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    The minus sign at the beginning of the reward function was implemented to minimize VUF at the PCC, because the agents' policy is trained to maximize the reward.

    “Because the policy of agents is trained to maximize the reward, the minus was implemented at the beginning of Equation (11) to minimize VUF at the PCC.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Study design and conductnone given

    Each training episode includes Gaussian noise to address forecast errors of load consumption and PV production or errors occurring because of communication.

    “To address such errors as the forecast error of load consumption and PV production or errors occurring because of communication, each episode has Gaussian noise.”
    Quote in the paperSubstantial claim

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

    Study design and conductnone given

    In this study, the size and the connected position of the PV systems were not considered.

    “In this study, we did not consider the size and the connected position of PV systems.”
    Quote in the paperNot a substantial claim
15 of 15

  • The paper relies on this without support.

    Power-electronics-based equipment such as static synchronous compensators or passive devices such as shunt capacitors can mitigate unbalanced voltage by compensating reactive power.

    “Power-electronics-based equipment, such as static synchronous compensators or passive devices, such as shunt capacitors, can mitigate the unbalanced voltage by compensating for the reactive power.”

  • The paper relies on this without support.

    These methods have low economic feasibility because they require additional equipment and maintenance costs.

    “However, these methods have low economic feasibility because they require additional equipment and maintenance”

  • The authors state this as an assumption.

    A voltage-passive device such as the on-load tap-changer (OLTC) and static voltage compensator is present in the grid.

    “Then, a voltagepassive device, such as the on-load tap-changer (OLTC) and static voltage compensator, is considered as a realistic assumption for a grid.”

  • The paper relies on this without support.

    More inverter-based resources will be added to the grid in the future.

    “Furthermore, because more inverter-based resources will be added to the grid in the future, an adaptive learning method is needed by using the latest learning models, which are trained using the digital twin model [17].”

  • The authors state this as an assumption.

    The magnitude of kvarMax is 0.44 × S_rated_i.

    “According to the IEEE 1547 standard, it is assumed that the magnitude of kvarMax is 0.44 × S rated i .”

  • The paper relies on this without support.

    The DRL method calculates optimal parameters based on offline analysis results and uses them for the next day's online operation.

    “On the other hand, the Deep Reinforcement Learning (DRL) method calculates the optimal parameters based on offline analysis results and uses them for the next day's online operation.”

  • The paper relies on this without support.

    The PG algorithm does not guarantee adequate convergence.

    “However, the PG algorithm does not guarantee adequate convergence”

  • The paper relies on this without support.

    The digital twin grid used for day-ahead training faithfully represents the actual distribution grid into which the trained policy is deployed.

    “Information flow of the day-ahead training in the digital twin and implementation in the actual grid.”
    “Update trained NN into Inverter in Actual Grid”
8 of 8

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.

  • Fixed control substantially improves one reported result

    Minor

    The paper says fixed volt-var control was insignificant or worse than no compensation. For the sunny-day IEEE 34 bus system, Table 2 instead reports VUF of 0.0514% for fixed control versus 0.52% for no compensation.

  • Reference [25] is attributed the subject of reference [26]

    Minor

    The text attributes work on ill-conditioned power flow using current injection to [25], but the bibliography gives that title for [26]. Reference [25] is listed as a marine-predators optimal-power-flow paper.

  • Reference [26] is attributed the subject of reference [25]

    Minor

    The text attributes a marine-predators optimal-power-flow method to [26], while the bibliography lists the marine-predators paper as [25].

3 of 3

Cites

Flagged first, then by citations

Cited by

Sorted by citations

Related

Authors
Jung, Y., Han, C., Lee, D., Song, S., & Jang, G.
Affiliations
  • Korea University (KR)
  • Korea Electrotechnology Research Institute (KR)
Corresponding author
Jang, Gilsoo
ORCID
Venue
Applied Sciences
Publisher
MDPI AG
Published
2021-09-26
Volume
11
Issue
19
Pages
8979
ISSN
2076-3417 (ISSN-L)
DOI
10.3390/app11198979
OpenAlex
W3203529343
Cited by
15
References
26
Citations by year
2026: 2 · 2025: 2 · 2024: 4 · 2023: 4 · 2022: 3
FWCI
0.79
Type
journal article
Language
en
Open access
Licence
Locations
  • Applied Sciences (publishedVersion, cc-by)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
  • Korea Institute of Energy Technology Evaluation and Planning (20191210301890)
  • Korea Electric Power Corporation (R20XO02-4)
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 › Power Quality and Harmonics
Keywords
photovoltaic inverter, voltage unbalance mitigation, point of common coupling, multi-agent deep reinforcement learning, proximal policy optimization, inverter-based resources, IEEE 13-bus system

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.

Cited sources are checked by gpt-6-sol.

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