Pith. sign in

REVIEW 4 major objections 6 minor 36 references

Agent-Based Decentralized Energy Management of EV Charging Station with Solar Photovoltaics via Multi-Agent Reinforcement Learning

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A decentralized multi-agent reinforcement learning controller with one LSTM-based agent per charger can cut unfinished EV charging demand to 28.9 while keeping energy costs below centralized baselines, and it keeps working chargers stable…

desk verdict Sensible MARL extension for EV charging with a plausible but under-supported evaluation: the headline numbers come from a stitched dataset and single runs. read the letter →

arxiv 2505.18750 v1 pith:Z3C3ER53 submitted 2025-05-24 eess.SY cs.AIcs.SYmath.OC

Jiarong Fan , Chenghao Huang , Hao Wang This is my paper
classification eess.SYcs.AIcs.SYmath.OC
keywords multi-agentreinforcementlearningEVchargingstationsolarPVdecentralizedenergymanagementcentralizedtrainingexecutionLSTMdenserewardchargerfaultrobustness
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to show that an EV charging station with on-site solar PV can be managed by a multi-agent reinforcement learning controller in which each charger is an independent agent, and that this decentralized design remains reliable even when some chargers fail. The authors combine LSTM networks with the MADDPG algorithm and add a dense reward signal that penalizes a charger whenever its current power falls below the completed average power needed to finish the session on time. In a one-month simulation, the proposed LSTM-MADDPG with the dense reward reaches the lowest energy cost and the lowest unfinished demand among the compared methods, with the sparse-reward version and the centralized MADQN trailing on service completion. The authors also report that centralized MADQN becomes unstable under partial charger faults while their CTDE-based controller keeps working chargers operating normally.

What carries the argument

The method is LSTM-MADDPG under centralized training with decentralized execution. Each charger is an agent whose actor policy is an LSTM followed by a fully connected layer, so the policy can read temporal patterns in PV generation and electricity prices; a centralized critic sees all agents' states and actions for training. The dense reward uses the completed average power, CAP equal to remaining demand divided by remaining time, and issues a negative reward proportional to CAP minus current charging power whenever the charger is falling behind, which turns the sparse departure-time dissatisfaction signal into a per-step training signal. A grid-capacity penalty and a battery-degradation term complete the reward.

What would settle it

Re-run the one-month evaluation with EV sessions, solar data, and wholesale prices drawn from the same region and the same time period, injecting the same two-charger fault pattern; if LSTM-MADDPG's unfinished demand rises toward the MADDPG level or its energy cost no longer beats the centralized baseline, the central claim of combined efficiency and fault reliability fails.

Watch

Extended reading notes

Core claim

The central discovery is that making each charger a locally executing agent, trained with global information and a dense charging-urgency reward, improves both cost and service completion in a simulated solar-powered charging station. The paper's own numbers show LSTM-MADDPG with the dense reward achieving an energy cost of 1021.6 and an unfinished demand of 28.9 over the evaluation month, versus 1086.2 and 349.6 for the same architecture with a sparse reward, 1102.3 and 584.5 for plain MADDPG, and 1361.6 and 82.1 for centralized MADQN. The paper argues that the centralized-training/decentralized-execution structure is what preserves stable charge and discharge behavior when two of the twenty chargers are replaced by random faulty signals, while the centralized baseline diverges.

Load-bearing premise

The load-bearing assumption is that the Los Angeles charging sessions, Belgian solar generation, and Australian wholesale prices can be stitched into a single representative environment, even though the paper does not describe how these three datasets are aligned in time or place; if that stitching is invalid, the reported costs and unfinished-demand numbers describe no real charging station.

Editorial extensions

If this is right

  • A charging station using this controller should keep delivering charge to connected vehicles when some chargers malfunction, because each working charger decides locally and does not depend on data from failed chargers.
  • The dense reward cuts unfinished demand sharply in the paper's simulation, from 349.6 to 28.9 at a slightly lower energy cost, indicating that service satisfaction and cost need not trade off against each other.
  • LSTM-encoded price and solar series let agents charge more when solar generation is high, which the paper illustrates by comparing sunny-day and cloudy-day charging decisions.
  • Because execution is decentralized, the trained policies require no communication between chargers at run time, easing deployment and reducing dependence on communication links.

Reading between the lines

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

  • The simulation mixes EV sessions from Los Angeles, solar traces from Belgium, and wholesale prices from Australia; a natural next check is whether the advantage survives with time-aligned data from a single region, since the reported absolute costs may otherwise be an artifact of dataset splicing.
  • The dense reward assumes the controller knows remaining charging time and remaining demand; in real stations these are forecasts, so the gap between simulated and deployed performance could be tested by feeding the same policies noisy arrival and departure estimates.
  • The same reward-shaping idea, penalizing a backlog of completed average power, could transfer to other multi-agent scheduling problems with sparse completion signals, such as battery storage dispatch or EV fleet coordination, but that extension is not tested here.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes a decentralized multi-agent reinforcement learning (MARL) approach for EV charging station energy management, where each charger is an agent. The method uses LSTM-augmented MADDPG with centralized training and decentralized execution (CTDE), and introduces a dense reward to improve charging completion. The authors claim the method minimizes energy cost and unfinished demand while maintaining robustness to charger faults, validated on an environment built from LA EV charging data, Belgian solar data, and Australian wholesale electricity prices. Results show LSTM-MADDPG with dense reward achieves the lowest energy cost (1021.6) and unfinished demand (28.9) among baselines, and qualitative figures suggest decentralized execution avoids instability under faults.

Significance. If the claims are substantiated, the paper would offer a useful decentralized, fault-tolerant control approach for EV charging stations with solar PV and V2G, and the LSTM-based temporal encoding plus dense reward design are reasonable contributions. The authors explicitly address system faults, a practical concern often overlooked. However, the current evidence is not sufficient to support the strong real-world validation claims: the evaluation environment is a synthetic composite of data from three different continents, the headline results lack statistical reliability, and the dense reward directly targets the main performance metric.

major comments (4)
  1. [Section IV (Simulations and Results)] The evaluation environment is assembled from three incompatible data sources: EV charging sessions in Los Angeles (ACN-Data, [34]), solar generation in Belgium (Elia, [35]), and wholesale electricity prices in Australia (AEMO, [36]). The paper does not state any temporal alignment, timezone conversion, resampling, or currency normalization among these datasets, which span UTC-8, UTC+1, and UTC+10 and use different currencies. The "Energy Cost" in Table I therefore has no defined monetary unit, and if the series are combined by row index, solar generation can be aligned to the wrong time of day relative to EV arrivals and price signals, creating artificial correlations that the agents can exploit. Consequently, the abstract's claim of "validation on a real-world dataset" is overstated; the numerical results describe a synthetic testbed, not any real charging station. This is load-bearing for the central claim and must be fixed either by providing a detailed, coherent alignment and currency conversion, or by scoping the claims to a synthetic composite environment.
  2. [Table I and Section IV] Table I reports single values for energy cost and unfinished demand for each algorithm, with no standard deviations, confidence intervals, or number of training seeds. MADDPG-style algorithms are stochastic due to random initialization, exploration noise, and replay buffer sampling, so a single run cannot support the claim that LSTM-MADDPG "achieves the best performance" or that the differences are significant. Similarly, the fault experiment (Fig. 3 and Fig. 4) is presented as two qualitative action curves with no quantitative measure of stability or degradation. The authors should run multiple seeds, report means and standard deviations, and preferably include a simple significance test or, at minimum, a clear statement of the number of runs.
  3. [Section II, Eqs. (11)-(13) and Table I] The dense reward R_user in Eq. (12) penalizes (CAP_i,t - a_i,t) when the required average power to meet the remaining demand exceeds 80% of maximum charging power. The evaluation metric "Unfinished Demand" is exactly the sum of dissatisfaction terms ds_i in Eq. (1), which is the energy shortage at departure. Thus the dense reward is a graded, per-step proxy for the headline metric, and the large reduction in unfinished demand for LSTM-MADDPG (Dense reward) versus the sparse-reward variant (28.9 versus 349.6) is substantially by construction. This is not a fatal flaw, but the paper should frame the result as an evaluation of reward shaping and justify that the shaping preserves the original objective (e.g., by citing potential-based reward shaping theory). Without this framing, the claim that the algorithm "maximizes charging service satisfaction" is circular.
  4. [Section II, near Eq. (9)] The authors state that "SOC related constraints are omitted here due to limited space," but battery state-of-charge bounds are physically essential when actions include discharging (V2G). Without SOC limits, the learned policy can discharge an EV below its minimum SOC or charge above its capacity, leading to infeasible operating points. Since battery degradation cost is explicitly modeled in Eqs. (9) and (10), the omission of SOC constraints is a modeling gap that affects the realism of the reported costs and charging satisfaction. The paper should either include SOC constraints in the formulation and simulation, or explicitly state this limitation in the abstract and conclusion and verify in the experiments that the learned policies respect plausible SOC bounds.
minor comments (6)
  1. [Abstract] The phrase "faults in faults in some chargers" contains a duplicated phrase and should be corrected to "faults in some chargers".
  2. [Section IV] The text "We set the solar capacity to32kWh" is missing a space, and the solar capacity should be a power rating in kW rather than an energy unit in kWh; please clarify the intended unit.
  3. [Figures 3 and 4] The labels "/glyph1197ormal mode" appear to contain a corrupted character; the intended label is presumably "normal mode".
  4. [Section II, Eqs. (9)-(10)] The battery degradation cost term uses κ_batt in Eq. (9) but κ_batt^t in Eq. (10); if the price is time-dependent, please use a consistent notation.
  5. [Section IV] The fault model is vaguely described as "randomly generated values" to replace faulty chargers' information. Please specify the distribution, which chargers are affected, when faults occur, and what information is corrupted (state, action, or both), so that the experiment is reproducible.
  6. [References [35] and [36]] References [35] and [36] point to general data portals rather than specific datasets; please include the exact dataset names, date ranges, and download access dates.

Circularity Check

1 steps flagged · score 6.0 of 10

The dense reward is defined as a running penalty on the exact shortfall that the headline 'Unfinished Demand' metric measures, so the dense-vs-sparse improvement is partly by construction.

  1. self definitional [Section II (dense reward, Eq. (12)); Section IV, Table I and discussion]
    "Specifically, the concept of completed average power is introduced as CAP_{i,t} = E^r_{i,t} / T^r_{i,t}, which measures the level of urgency to complete the charging task of EV i. ... Then, a negative reward is sent to guide the agent to avoid this situation, expressed as R^{user}_{i,t} = -(CAP_{i,t} - a_{i,t})."

    The dense reward is defined as minus the gap between the average power needed to finish the remaining demand by departure, E^r_{i,t}/T^r_{i,t}, and the current charging action a_{i,t}. The paper's own dissatisfaction metric in Eq. (1) is the terminal value of the same remaining-demand quantity: d^s_i = E^dem_i - E_i,t=tar whenever the final energy falls short. Thus the reward is, by construction, a potential function for the 'Unfinished Demand' reported in Table I. The large dense-reward improvement (28.9 vs. 349.6 for sparse reward) is therefore the optimizer following the objective encoded in the reward rather than an independent empirical prediction. The energy-cost and fault-tolerance results are not built into the reward in this direct way and retain independent content.

full rationale

The paper's central comparison in Table I has two components. The unfinished-demand component is partially circular: Eq. (12)'s dense reward penalizes exactly the rate at which remaining demand fails to be cleared by departure, so the dense-reward reduction of unfinished demand is the designed effect of the reward, not a separate discovery. The energy-cost component, however, is a genuine optimization result against baselines since the cost reward is the objective being optimized rather than a subsidiary claim, and the fault-tolerance/CTDE result is demonstrated through the action plots and is not encoded in the reward. The self-citations in the literature review (e.g., Refs. [8], [12], [13]) are not load-bearing, and no uniqueness theorem is imported. The stitched LA/Belgium/Australia dataset raises validity concerns but is not a circularity issue, so it is not counted in the score. Overall, partial circularity exists in one headline metric, giving a score of 6 rather than a higher score, because the cost and robustness claims are independent of the reward-shaping reduction.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central claim depends on several hand-chosen design parameters (reward threshold, penalty weights, network size) and on a set of assumptions about the simulation environment, including the validity of combining public datasets from different regions and the correctness of the battery degradation model.

free parameters (5)
  • Dense reward urgency threshold = 80% of maximum charging power (P_bar^ch)
    Hand-chosen threshold in (12) that determines when an agent receives a negative urgency reward; directly affects the unfinished-demand metric.
  • Grid violation penalty magnitude = Not reported
    R_grid in (13) enforces the station capacity constraint; its value is not given and affects the learned policy.
  • LSTM hidden size and network dimensions = Not reported
    Architecture of actor and critic networks; not specified in the paper.
  • RL hyperparameters (discount factor, learning rate, exploration noise, replay buffer size) = Not reported
    Standard MADDPG hyperparameters governing training; absent from the manuscript.
  • Reward combination weights = Implicitly 1 for cost, user, and grid terms
    Equation (13) sums the three reward terms without explicit weights; no sensitivity analysis is provided.
assumptions (5)
  • domain assumption The historical EV charging sessions from ACN-Data, solar data from Elia, and wholesale prices from AEMO can be combined into one coherent simulation environment.
    Invoked in Section IV without stating how the datasets are aligned in time or region.
  • domain assumption The energy throughput battery degradation model from [32] adequately captures cycle aging costs for V2G operation.
    Used in equation (9); the validity of this model is taken from the citation without verification in this setting.
  • standard math MADDPG with centralized critics converges to a good policy with the given reward and simulation length.
    The paper relies on the standard convergence behavior of MADDPG [33] without convergence proof or training curves.
  • ad hoc to paper SOC-related constraints omitted 'due to limited space' do not change the reported results.
    Section II states SOC constraints are omitted; the impact of this omission on cost and feasibility is not assessed.
  • ad hoc to paper The fault model, where faulty chargers broadcast randomly generated values, represents realistic charger faults.
    Section IV describes the fault simulation as random values; no real fault data or physics-based fault model is used.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Agent-Based Decentralized Energy Management of EV Charging Station with Solar Photovoltaics via Multi-Agent Reinforcement Learning." pith.science (2026). https://pith.science/paper/Z3C3ER53

@misc{pith2026250518750,
  author       = {Pith},
  title        = {Pith review of: Agent-Based Decentralized Energy Management of EV Charging Station with Solar Photovoltaics via Multi-Agent Reinforcement Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Z3C3ER53}},
  note         = {Machine review of arXiv:2505.18750}
}
read the original abstract

In the pursuit of energy net zero within smart cities, transportation electrification plays a pivotal role. The adoption of Electric Vehicles (EVs) keeps increasing, making energy management of EV charging stations critically important. While previous studies have managed to reduce energy cost of EV charging while maintaining grid stability, they often overlook the robustness of EV charging management against uncertainties of various forms, such as varying charging behaviors and possible faults in faults in some chargers. To address the gap, a novel Multi-Agent Reinforcement Learning (MARL) approach is proposed treating each charger to be an agent and coordinate all the agents in the EV charging station with solar photovoltaics in a more realistic scenario, where system faults may occur. A Long Short-Term Memory (LSTM) network is incorporated in the MARL algorithm to extract temporal features from time-series. Additionally, a dense reward mechanism is designed for training the agents in the MARL algorithm to improve EV charging experience. Through validation on a real-world dataset, we show that our approach is robust against system uncertainties and faults and also effective in minimizing EV charging costs and maximizing charging service satisfaction.

Figures

Figures reproduced from arXiv: 2505.18750 by the authors.

Figure 1
Figure 1. System model of an EV charging station with solar [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The neural network architecture (left) and the framework of MADDPG (right). [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 5
Figure 5. Solar energy under different weather conditions. [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figures from the paper (1 more)
Figure 6
Figure 6. Figure 6: Three chargers’ decisions under sunny conditions using [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 30 canonical work pages

  1. [34]

    Acn-data: Analysis and applications of an open ev charging dataset,

    Z. J. Lee, T. Li, and S. H. Low, “Acn-data: Analysis and applications of an open ev charging dataset,” inProceedings of the Tenth ACM International Conference on Future Energy Systems, 2019, pp. 139–149

    work page 2019

  2. [35]

    griddata,

    Elia, “griddata,” 2021, https://www.elia.be/en/grid-data

    work page 2021

  3. [36]

    Nem data dashboard,

    AEMO, “Nem data dashboard,” 2023, https://aemo.com.au/energy- systems/electricity/national-electricity-market-nem/data-nem/data- dashboard-nem

    work page 2023

  4. [1]

    Towards sustainable smart cities: A review of trends, architectures, components, and open challenges in smart cities,

    B. N. Silva, M. Khan, and K. Han, “Towards sustainable smart cities: A review of trends, architectures, components, and open challenges in smart cities,”Sustainable cities and society, vol. 38, pp. 697–713, 2018

    work page 2018

  5. [2]

    Smart energy systems for sustainable smart cities: Current developments, trends and future directions,

    E. O’Dwyer, I. Pan, S. Acha, and N. Shah, “Smart energy systems for sustainable smart cities: Current developments, trends and future directions,”Applied energy, vol. 237, pp. 581–597, 2019

    work page 2019

  6. [3]

    European smart cities: The role of zero energy buildings,

    A. Kylili and P. A. Fokaides, “European smart cities: The role of zero energy buildings,”Sustainable cities and society, vol. 15, pp. 86–95, 2015

    work page 2015

  7. [4]

    Global ev outlook 2020,

    C. IEAet al., “Global ev outlook 2020,”URL: https://www. iea. org/reports/global-ev-outlook-2020, 2020

    work page 2020

  8. [5]

    Iea (2024), global ev outlook 2024

    I. E. A. (IEA), “Iea (2024), global ev outlook 2024.” [Online]. Available: https://www.iea.org/reports/global-ev-outlook-2024

    work page 2024

Show all 36 references
  1. [6]

    Towards a circular and low-carbon economy: Insights from the transitioning to electric vehicles and net zero economy,

    N. O. Bonsu, “Towards a circular and low-carbon economy: Insights from the transitioning to electric vehicles and net zero economy,”Journal of Cleaner Production, vol. 256, p. 120659, 2020

    2020

  2. [7]

    Residential net-zero energy buildings: Review and perspective,

    W. Wu and H. M. Skye, “Residential net-zero energy buildings: Review and perspective,”Renewable and Sustainable Energy Reviews, vol. 142, p. 110859, 2021

    2021

  3. [8]

    Investigating state-of-the-art planning strategies for electric vehicle charging infrastructures in coupled trans- port and power networks: A comprehensive review,

    J. Li, A. Chew, and H. Wang, “Investigating state-of-the-art planning strategies for electric vehicle charging infrastructures in coupled trans- port and power networks: A comprehensive review,”Progress in Energy, vol. 6, no. 4, pp. 1–26, 2024

    2024

  4. [9]

    Optimum design of an ev/phev charging station with dc bus and storage system,

    S. Bai, D. Yu, and S. Lukic, “Optimum design of an ev/phev charging station with dc bus and storage system,” in2010 IEEE Energy Conver- sion Congress and Exposition. IEEE, 2010, pp. 1178–1184

    2010

  5. [10]

    Optimized operational cost reduction for an ev charging station integrated with battery energy storage and pv generation,

    Q. Yan, B. Zhang, and M. Kezunovic, “Optimized operational cost reduction for an ev charging station integrated with battery energy storage and pv generation,”IEEE Transactions on Smart Grid, vol. 10, no. 2, pp. 2096–2106, 2018

    2018

  6. [11]

    Online distributed mpc-based optimal scheduling for ev charging stations in distribution systems,

    Y . Zheng, Y . Song, D. J. Hill, and K. Meng, “Online distributed mpc-based optimal scheduling for ev charging stations in distribution systems,”IEEE transactions on industrial informatics, vol. 15, no. 2, pp. 638–649, 2018

    2018

  7. [12]

    Network-aware electric vehicle coordination for vehicle-to-anything value stacking considering uncer- tainties,

    C. Jiang, A. Liebman, and H. Wang, “Network-aware electric vehicle coordination for vehicle-to-anything value stacking considering uncer- tainties,” in2023 IEEE/IAS 59th Industrial and Commercial Power Systems Technical Conference (I&CPS). IEEE, 2023, pp. 1–9

    2023

  8. [13]

    Marl for decentralized electric vehicle charging coordination with v2v energy exchange,

    J. Fan, H. Wang, and A. Liebman, “Marl for decentralized electric vehicle charging coordination with v2v energy exchange,” inIECON 2023-49th Annual Conference of the IEEE Industrial Electronics Society. IEEE, 2023, pp. 1–6

    2023

  9. [14]

    Smart online charging algorithm for electric vehicles via customized actor–critic learning,

    Y . Cao, H. Wang, D. Li, and G. Zhang, “Smart online charging algorithm for electric vehicles via customized actor–critic learning,”IEEE Internet of Things Journal, vol. 9, no. 1, pp. 684–694, 2021

    2021

  10. [15]

    Optimal scheduling for charging and discharging of electric vehicles,

    Y . He, B. Venkatesh, and L. Guan, “Optimal scheduling for charging and discharging of electric vehicles,”IEEE transactions on smart grid, vol. 3, no. 3, pp. 1095–1105, 2012

    2012

  11. [16]

    Distributed hierarchi- cal coordination of networked charging stations based on peer-to-peer trading and ev charging flexibility quantification,

    J. Zhang, L. Che, X. Wan, and M. Shahidehpour, “Distributed hierarchi- cal coordination of networked charging stations based on peer-to-peer trading and ev charging flexibility quantification,”IEEE Transactions on Power Systems, vol. 37, no. 4, pp. 2961–2975, 2021

    2021

  12. [17]

    Decentralized v2g/g2v scheduling of ev charging stations by considering the conversion effi- ciency of bidirectional chargers,

    J.-T. Liao, H.-W. Huang, H.-T. Yang, and D. Li, “Decentralized v2g/g2v scheduling of ev charging stations by considering the conversion effi- ciency of bidirectional chargers,”Energies, vol. 14, no. 4, p. 962, 2021

    2021

  13. [18]

    Decentralized cooperative approach for electric vehicle charging,

    A. Paudel, S. A. Hussain, R. Sadiq, H. Zareipour, and K. Hewage, “Decentralized cooperative approach for electric vehicle charging,” Journal of Cleaner Production, vol. 364, p. 132590, 2022

    2022

  14. [19]

    Decentralized energy management strategy based on predictive controllers for a medium voltage direct current photovoltaic electric vehicle charging station,

    J. P. Torreglosa, P. Garc ´ıa-Trivi˜no, L. M. Fern ´andez-Ramirez, and F. Ju- rado, “Decentralized energy management strategy based on predictive controllers for a medium voltage direct current photovoltaic electric vehicle charging station,”Energy Conversion and Management, v...

    2016

  15. [20]

    A multi-agent system for controlled charging of a large population of electric vehicles,

    E. L. Karfopoulos and N. D. Hatziargyriou, “A multi-agent system for controlled charging of a large population of electric vehicles,”IEEE Transactions on Power Systems, vol. 28, no. 2, pp. 1196–1204, 2012

    2012

  16. [21]

    R. S. Sutton and A. G. Barto,Reinforcement learning: An introduction. MIT press, 2018

    2018

  17. [22]

    Reinforcement learning based ev charging management systems–a review,

    H. M. Abdullah, A. Gastli, and L. Ben-Brahim, “Reinforcement learning based ev charging management systems–a review,”IEEE Access, vol. 9, pp. 41 506–41 531, 2021

    2021

  18. [23]

    Smart and resilient ev charging in sdn-enhanced vehicular edge computing networks,

    J. Liu, H. Guo, J. Xiong, N. Kato, J. Zhang, and Y . Zhang, “Smart and resilient ev charging in sdn-enhanced vehicular edge computing networks,”IEEE Journal on Selected Areas in Communications, vol. 38, no. 1, pp. 217–228, 2019

    2019

  19. [24]

    Reinforcement learning for real- time pricing and scheduling control in ev charging stations,

    S. Wang, S. Bi, and Y . A. Zhang, “Reinforcement learning for real- time pricing and scheduling control in ev charging stations,”IEEE Transactions on Industrial Informatics, vol. 17, no. 2, pp. 849–859, 2019

    2019

  20. [25]

    Optimal electric vehicle charging strategy with markov decision process and reinforcement learning technique,

    T. Ding, Z. Zeng, J. Bai, B. Qin, Y . Yang, and M. Shahidehpour, “Optimal electric vehicle charging strategy with markov decision process and reinforcement learning technique,”IEEE Transactions on Industry Applications, vol. 56, no. 5, pp. 5811–5823, 2020

    2020

  21. [26]

    Intelligent electric vehicle charging recommendation based on multi- agent reinforcement learning,

    W. Zhang, H. Liu, F. Wang, T. Xu, H. Xin, D. Dou, and H. Xiong, “Intelligent electric vehicle charging recommendation based on multi- agent reinforcement learning,” inProceedings of the Web Conference 2021, 2021, pp. 1856–1867

    2021

  22. [27]

    Multistep multiagent reinforcement learning for optimal energy schedule strategy of charging stations in smart grid,

    Y . Zhang, Q. Yang, D. An, D. Li, and Z. Wu, “Multistep multiagent reinforcement learning for optimal energy schedule strategy of charging stations in smart grid,”IEEE Transactions on Cybernetics, vol. 53, no. 7, pp. 4292–4305, 2022

    2022

  23. [28]

    Electric vehicle charging scheduling control strategy for the large-scale scenario with non-cooperative game-based multi-agent reinforcement learning,

    L. Fu, T. Wang, M. Song, Y . Zhou, and S. Gao, “Electric vehicle charging scheduling control strategy for the large-scale scenario with non-cooperative game-based multi-agent reinforcement learning,”Inter- national Journal of Electrical Power & Energy Systems, vol. 153, p. 109...

    2023

  24. [29]

    Decentralized collaborative optimal scheduling for ev charging stations based on multi-agent reinforcement learning,

    H. Li, B. Han, G. Li, K. Wang, J. Xu, and M. W. Khan, “Decentralized collaborative optimal scheduling for ev charging stations based on multi-agent reinforcement learning,”IET Generation, Transmission & Distribution, vol. 18, no. 6, pp. 1172–1183, 2024

    2024

  25. [30]

    The viability of vehicle- to-grid operations from a battery technology and policy perspective,

    K. Uddin, M. Dubarry, and M. B. Glick, “The viability of vehicle- to-grid operations from a battery technology and policy perspective,” Energy Policy, vol. 113, pp. 342–347, 2018. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S0301421517307619

    2018

  26. [31]

    Long short-term memory,

    S. Hochreiter and J. Schmidhuber, “Long short-term memory,”Neural computation, vol. 9, no. 8, pp. 1735–1780, 1997

    1997

  27. [32]

    Optimal bat- tery capacity of grid-connected pv-battery systems considering battery degradation,

    Y . Wu, Z. Liu, J. Liu, H. Xiao, R. Liu, and L. Zhang, “Optimal bat- tery capacity of grid-connected pv-battery systems considering battery degradation,”Renewable Energy, vol. 181, pp. 10–23, 2022

    2022

  28. [33]

    Multi-agent actor-critic for mixed cooperative-competitive environ- ments,

    R. Lowe, Y . Wu, A. Tamar, J. Harb, P. Abbeel, and I. Mordatch, “Multi-agent actor-critic for mixed cooperative-competitive environ- ments,”arXiv preprint arXiv:1706.02275, 2017

    2017 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.