EV-Centric Technical Virtual Power Plants in Active Distribution Networks: An Integrative Review of Physical Constraints, Bidding, and Control

Abstract

The accelerated low-carbon transition of power systems and the widespread integration of Electric Vehicles (EVs) present both severe operational challenges and substantial flexible regulation potential for Active Distribution Networks (ADNs). This paper provides an integrative review of the coordinated control and multi-market bidding mechanisms for EV-centric Technical Virtual Power Plants (TVPPs). Moving beyond descriptive surveys, this review systematically synthesizes the fragmented literature across three critical dimensions: (1) the physical-economic bidirectional mapping, which considers nonlinear power flow constraints and node voltage limits within the TVPP framework; (2) multi-market coupling mechanisms, evolving from unilateral energy bidding to coordinated participation in carbon trading and ancillary services; and (3) real-time control strategies, critically evaluating the trade-offs between optimization techniques (e.g., Model Predictive Control) and cutting-edge artificial intelligence approaches (e.g., Deep Reinforcement Learning) in mitigating battery degradation. Furthermore, a transparent review methodology is adopted to ensure literature rigor. By explicitly outlining the boundaries between TVPPs, Commercial VPPs (CVPPs), and EV aggregators, this paper identifies core unresolved trade-offs among aggregation fidelity, market complexity, and communication latency, providing evidence-backed pathways for future engineering demonstrations and V2G applications.

1. Introduction

Guided by the “carbon peaking and carbon neutrality” (Dual-Carbon) goals, China is continuously advancing energy structure reforms and the construction of novel power systems and is vigorously developing green energy systems with an emphasis on increasing the penetration rate of renewable energy sources such as photovoltaics (PV) and wind power. According to data released by the National Energy Administration, by the end of 2024, the total proportion of renewable energy in China reached 56%, with wind and PV power accounting for over 70% of this share. Faced with the integration of high-proportion distributed renewable energy, traditional distribution network planning and scheduling methods encounter severe challenges of inadequate coordination among source-network-load-storage flexible resources [1,2]. How to generate temporal scenarios that accurately characterize renewable energy uncertainties during the expansion and planning stages of Active Distribution Networks (ADNs), and achieve the optimal configuration of multi-energy complementary systems (e.g., electricity-hydrogen hybrid energy storage), has become the core foundation for achieving the Dual-Carbon goals [3].

However, the rapid growth in the penetration rate of renewable energy has significantly exacerbated the difficulty of power system operation and control. Due to the influence of natural meteorological conditions, PV and wind power outputs exhibit strong volatility and randomness. The rapid changes in generation power inevitably cause severe impacts on the active distribution network, leading to issues such as local node voltage violations, increased network losses, and bidirectional power flow reversals. To this end, researchers have introduced probabilistic power flow calculation methods incorporating Hybrid Distribution Transformers (HDTs) [4], as well as hierarchical decentralized coordination mechanisms based on the Alternating Direction Method of Multipliers (ADMM) to alleviate distribution network congestion [5]. Moreover, optimizing the allocation of intelligent switching devices, such as Soft Open Points (SOPs), in the distribution network can effectively improve the operational flexibility of the network topology [6]. However, relying solely on grid-side hardware retrofits or the traditional expansion of thermal power units to provide peak-shaving reserves incurs exorbitant economic costs and fundamentally contradicts the original intention of low-carbon transition.

Simultaneously, with the accelerating pace of transportation electrification, the ownership of electric vehicles (EVs) has surged exponentially. The uncoordinated charging behavior of massive EVs creates a severe “peak-on-peak” superposition effect with the grid’s basic daily peak load, leading to accelerated transformer aging and severe line congestion in the distribution network [7]. Studies have shown that the different distributions of public charging stations and residential charging piles exert vastly different spatial pressures on the distribution network, urgently necessitating day-ahead charging scheduling and peak shaving via heuristic optimization algorithms such as multi-objective particle swarm optimization [8,9]. On the flip side of the coin, through Vehicle-to-Grid (V2G) technology, EVs equipped with large-capacity power batteries can be regarded as highly premium distributed mobile energy storage units, possessing immense physical potential to rapidly respond to millisecond-level grid frequency fluctuations (e.g., Fast Frequency Response, FFR) [10] and exhibiting extremely high value in suppressing microgrid voltage deviations [11].

To further enhance the resilience of such systems, recent research has explored intelligent adaptive controllers to manage environmental stochastics. For instance, ref. [12] proposed a fuzzy-logic-based PID control strategy for PV-powered EV charging, which dynamically tunes controller gains to maintain voltage stability under fluctuating solar irradiance and dynamic load demands. This approach demonstrates that integrating heuristic reasoning with classical control can significantly improve energy efficiency and battery protection in renewable-integrated infrastructures.

Therefore, breaking the spatio-temporal dispersion barriers of massive flexible resources, such as EVs, distributed generators, and community energy storage, has become an urgent pain point for both academia and industry [13,14]. Virtual Power Plant (VPP) technology emerged to address this challenge. By leveraging advanced Internet of Things communication technologies and sophisticated software algorithms, VPP aggregates massive heterogeneous flexible resources in the distribution network into a generalized frequency regulation/peak shaving unit with “collective certainty” to participate in upper-level grid bidding. This has become a key pathway to break the deadlock of flexibility scarcity in novel power systems.

Although research on the aggregated control of VPPs and EVs has witnessed explosive growth in recent years, with the evolution of novel power system configurations, most existing review studies (e.g., Wang et al. [15]; Yi et al. [16]) have gradually exposed obvious limitations in the following dimensions:

First, regarding the integration of underlying physical topologies, existing studies mostly focus on traditional Commercial VPPs (CVPPs), often neglecting the underlying physical constraints of the distribution network. As astutely pointed out by Gough et al. in recent research, the so-called “pure economic dispatch” lacking distribution network congestion and voltage control easily incurs the risk of boundary violations during actual physical execution. Therefore, the construction of Technical VPP (TVPP) models in active distribution networks, considering diverse distributed resources, appears particularly crucial [17,18]. Accurately assessing the operational flexibility of TVPPs [19] and precisely depicting the Feasible Operation Region (FOR) of VPPs through polytope equivalence or dynamic aggregation strategies represent the physical baseline for ensuring dispatch security [20,21], yet existing reviews still lack a systematic summary of this underlying physical architecture.

Second, in terms of the cross-dimensional coupling of market mechanisms, the participation pathway of VPPs is accelerating its evolution from a single day-ahead energy market towards a multi-market coupling of “energy-reserve-carbon trading” [22]. Especially in the context of vigorously promoting carbon emission trading, “low-carbon orientation” and “electricity-carbon coupling” have become core variables determining the economic benefits of VPP multi-microgrid aggregators [23,24]. The latest research has even begun to explore multi-objective scheduling models combining carbon trading and green certificate trading [25], as well as incentive-based tiered carbon pricing mechanisms [26]. However, extant literature mostly explores energy and carbon trading in isolation, failing to deeply analyze the endogenous impact of stepped carbon trading on the differentiated bidding behaviors between power generators and VPPs [27].

Finally, concerning the handling of multiple uncertainties and high-frequency control, due to the extreme randomness of EV users’ travel willingness and external market price jumps, a single risk-averse strategy may lead to the loss of massive “windfall profits”. Therefore, introducing hybrid stochastic scheduling strategies that integrate risk-seeking and risk-averse preferences has become a cutting-edge direction [28,29]. Currently, risk-averse bidding models based on Conditional Value at Risk (CVaR) are widely applied to quantify financial losses caused by extreme price volatility [30,31]. In addition, facing the “curse of dimensionality” induced by massive discrete variables, traditional mathematical programming algorithms encounter computational bottlenecks. Cutting-edge research has begun to introduce Deep Reinforcement Learning (DRL) algorithms to construct Stackelberg game models [32], but existing reviews often fail to provide systematic theoretical concatenations and cross-evaluations between CVaR risk metrics and advanced AI black-box algorithms.

To bridge these research gaps, this paper aims to provide an integrative review of the coordinated control technologies and bidding mechanisms for multi-type flexible resources—particularly EV aggregators—operating within ADNs under the Technical VPP (TVPP) framework. Rather than presenting a uniquely panoramic or “first” survey, this paper strictly positions itself as a critical synthesis of EV-centric TVPP physics, multi-market participation, and real-time control. The core contributions of this integrative review are as follows:

• Synthesizing a “physics-economics” bidirectional mapping framework: We systematically outline the Feasible Operation Region (FOR) equivalence and optimal scheduling methods for large-scale EV clusters, strictly considering nonlinear power flow constraints and network congestion within the TVPP framework.
• Evaluating multi-market coupling mechanisms: We analyze recent multi-agent coordinated architectures, tracking the evolution from unilateral energy market bidding to coupled participation involving carbon trading, green certificates, and Stackelberg game pricing.
• Critically comparing real-time control and AI algorithms: We cross-evaluate various uncertainty management approaches (e.g., CVaR) and real-time execution methods, objectively discussing the specific capabilities of Deep Reinforcement Learning (DRL) in circumventing high-dimensional state spaces and addressing battery degradation.

To provide an intuitive overview of the methodological landscape, Figure 1 illustrates the distribution of core algorithms applied in recent VPP-EV coordination studies. Furthermore,

Table 1. Innovation comparison of the current review with recent literature.

Reference/Review Paper	Year	Focus on TVPP & Physical Constraints	Multi-Market & Carbon Coupling	High-Frequency Control & AI	Uncertainty & Risk Modeling	Comprehensive TEA & Engineering
[15]	2024	Ignored (Focuses on commercial aggregation)	Energy & ancillary services only	Traditional MILP/Heuristics	Stochastic programming	Simulation only
[16]	2024	Considered (Node voltage limits included)	Unconsidered (Ignores carbon trading)	Model Predictive Control (MPC)	Information Gap Decision Theory (IGDT)	Lacks real-world validation
[33]	2024	Ignored (Network congestion omitted)	Day-ahead energy market only	Deep Q-Network (DQN)	Not explicitly modeled	LCC analysis included
Proposed Review	2025	Comprehensive (Nonlinear power flow & TVPP FOR equivalence)	Coupled (Energy, carbon trading & Stackelberg pricing)	Cross-evaluated (DRL vs. MPC vs. Exact Polytope)	Synthesized (CVaR & Hybrid Risk Theories)	Evaluated (Micro-bilateral benefits & empirical data)

explicitly summarizes the key innovations of this review compared with existing literature.

To provide physical context for the application scenarios shown in Figure 2, it is crucial to understand that these market mechanisms are deeply coupled with the physical operational limits of Active Distribution Networks (ADNs). For instance, “Voltage & Congestion Management” directly addresses the physical bottlenecks of ADNs, such as distribution transformer overloading and feeder voltage deviations caused by uncoordinated EV charging. By participating in these scenarios, EV aggregators do not merely pursue financial arbitrage; they physically alter the local power flow by providing localized active power curtailment or reactive power support, thereby extending the thermal lifespan of grid assets. Similarly, “Frequency Regulation” requires EVs to physically adjust their charging power at sub-second intervals to balance the real-time active power mismatch, highlighting the transition of EVs from passive loads to active physical regulating assets.

2. Review Methodology

To substantiate this paper as a systematic and integrative review rather than a narrative survey, a transparent methodology was adopted to identify, screen, and select the relevant literature.

Databases: A comprehensive literature search was conducted across major academic databases, primarily Web of Science, IEEE Xplore, Scopus, and MDPI.

Keywords: The search string utilized combinations of primary terms: (“Virtual Power Plant” OR “Technical VPP” OR “CVPP”) AND (“Electric Vehicle” OR “V2G” OR “EV Aggregator”) AND (“Active Distribution Network” OR “Multi-market Bidding” OR “Real-time Control”).

Time Span: The retrieval focused heavily on articles published between 2019 and 2024 to capture the latest advancements in AI applications and multi-market structures, with seminal foundational papers included where necessary.

Inclusion and Exclusion Criteria: Articles were included if they provided quantitative models, peer-reviewed algorithmic validations, or techno-economic assessments of EV/VPP integration. Conceptual conference abstracts, non-English publications, and studies lacking auditable simulation baselines were explicitly excluded.

Screening Procedure: The initial search yielded over 800 records. After removing duplicates and screening titles/abstracts for direct relevance to the TVPP physical-economic framework, 150 full-text articles were critically assessed, resulting in the final selection synthesized in this review.

3. Taxonomy and Conceptual Boundaries of Key Entities

A recurring challenge in the existing literature is the blurred terminology regarding architectural frameworks, market roles, and control layers. To establish a rigorous foundation for the subsequent analysis,

Table 2. Taxonomy and Conceptual Boundaries of Key Entities in the Integration of EVs.

Concept	Category/Layer	Core Function & Definition	Treatment of Physical Grid Constraints
CVPP (Commercial VPP)	Market Role/Economic Layer	Aggregates portfolio capacity to bid in wholesale energy or ancillary markets to maximize financial profit.	Typically ignores local network topology, nodal voltages, and line congestion.
TVPP (Technical VPP)	Architecture/Physical Layer	Manages the actual dispatch of DERs within a specific geographical grid area, ensuring safe operation limits.	Strictly adheres to Active Distribution Network (ADN) power flow limits and thermal constraints.
EV Aggregator	Market Entity/Intermediary	Acts as an intermediary agent that clusters dispersed EVs to offer required capacity to CVPPs, TVPPs, or Grid Operators.	Relies on upper-level TVPP signals or independent heuristic rules for internal physical dispatch.
V2G/V2X	Control Layer/Interface Technology	The bi-directional charging/discharging hardware and protocol layer enabling power exchange between vehicles and the grid/buildings.	Focuses on battery-level physics (e.g., SOC, degradation, charging current) and local converter limits.

explicitly clarifies the formal taxonomy and conceptual boundaries among Technical VPPs (TVPPs), Commercial VPPs (CVPPs), EV Aggregators, and V2G/V2X concepts.

4. Load Forecasting and Aggregation Modeling of EVs

In the paradigm of novel power systems, the uncoordinated integration of massive electric vehicles (EVs) inevitably exacerbates the spatio-temporal load fluctuations of the distribution network, even triggering severe “peak-on-peak” effects at local nodes. To integrate highly heterogeneous, spatio-temporally dispersed, and stochastic discrete vehicles into controllable virtual energy storage units—achieving a fundamental shift from “individual randomness” to “collective certainty”—high-precision load forecasting and scientifically grounded underlying aggregation modeling have become prerequisites for Virtual Power Plants (VPPs) to securely participate in grid interactions. Current research is gradually moving away from macroscopic deterministic models, advancing deeply along two technical pathways: AI-driven microscopic spatio-temporal precise forecasting and bi-level aggregation evaluation balancing physical constraints and subjective willingness.

The physical significance of the aggregation architecture depicted in Figure 3 lies in its ability to map the highly stochastic, microscopic physical states of individual EVs into a deterministic macroscopic boundary. In a physical ADN, an EV is not just a numerical data point, but a mobile battery subject to rigorous electrochemical constraints and vehicular mobility. The spatio-temporal forecasting layer captures the physical movement of EVs across urban traffic networks, while the aggregation layer constructs the feasible physical power and energy boundaries (i.e., the Polytope) of the EV cluster. This architecture is physically critical because it ensures that the aggregated dispatch signals from the upper-level VPP will never violate the actual electrochemical safety limits (State-of-Charge) of individual EV batteries, nor will they exceed the localized topological capacities of the distribution feeders where the EVs are currently plugged in.

4.1. Data-Driven and Spatio-Temporal Charging Load Forecasting

Addressing the complex load coupling characteristics in novel power systems, traditional physical statistics-based models (e.g., single Monte Carlo sampling) can no longer meet the requirements of high-frequency scheduling. Currently, academia widely adopts data-driven AI methods for high-precision forecasting. To further capture the spatial transfer patterns and temporal fluctuation coupling relationships of EV charging loads, researchers have introduced multi-source information fusion and advanced deep learning architectures:

Multi-dimensional Spatio-temporal Feature Extraction and Graph Neural Networks (GNN): EV loads are highly bound to urban traffic flows. Authentic literature points out that considering dynamic traffic flow changes and the coupling influence of multiple origin-destinations in complex road networks, the spatio-temporal distribution of charging loads can be obtained via the Monte Carlo algorithm, and a Bi-GRU-Seq2Seq model can be used to establish the mapping between travel information and charging loads [34]. Furthermore, by integrating historical loads, charging pile occupancy, and dynamic electricity prices, improved Spatio-Temporal Graph Convolutional Networks (STGCN) exhibit outstanding performance in hourly forecasting. Empirical studies show that a reasonable multi-scale temporal window strategy (hour-day-week) can significantly reduce the Mean Absolute Error (MAE) by 9% to 16% [35].

Generative Diffusion Models and Data-driven Uncoordinated Baseline Estimation: Prior to calculating VPP regulation potential, accurately obtaining the uncoordinated charging baseline is a crucial starting point. To profoundly reveal the internal physical relationships among charging factors, cutting-edge research proposed a physically-based diffusion model for the charging power of plug-in EVs, successfully applying it to efficient Area Control Error (ACE) correction [36]. At the macro load evolution level, by comparing different diffusion models like Gompertz and Bass, researchers accurately simulated the EV load evolution trends across various countries under the influence of external environments (e.g., charging infrastructure) [37]. Meanwhile, at the micro data foundation level, fully open-source, data-driven EV charging simulation environments like ACN-Sim are widely introduced. It embeds realistic battery charging behaviors and unbalanced three-phase infrastructure constraints, providing the most realistic simulation scenarios and uncoordinated charging baselines for reinforcement learning agents (e.g., OpenAI Gym) [38].

Probabilistic Forecasting and Meta-learning Architectures: Facing strong load uncertainty, single deterministic forecasting lacks risk consideration. To solve the “cold start” pain point of newly built charging stations lacking historical data, the latest research proposes MetaProbformer, a probabilistic forecasting model based on Transformer and meta-learning architectures, effectively enveloping extreme fluctuations without requiring massive training data [39]. Additionally, at the distribution network planning level, researchers have introduced the spherical simplex unscented transformation, an analytical uncertainty modeling method, to precisely quantify the impact of green technology forecasting errors on system flexibility [40].

Data Privacy Protection and Underlying Cryptographic Architectures: Precise forecasting relies heavily on user privacy trajectories. Cutting-edge research proposes a three-stage deep reinforcement learning framework integrated with differential privacy, where charging stations locally encrypt net load data, entirely protecting data privacy while ensuring zero voltage violations [41]. Even more rigorously, some studies introduce SWIFFT hash functions and signcryption based on lattice cryptography. While effectively balancing loads, this fundamentally avoids the heavy computational overhead caused by traditional encryption [42].

4.2. Dimensionality Reduction, Psychological Quantification, and Capacity Modeling of Heterogeneous EV Clusters

The fundamental prerequisite for VPPs to engage in grid interactions is the accurate quantification of the aggregate flexibility of heterogeneous EV clusters. As emphasized in recent comprehensive reviews [15], the aggregation characteristics of VPPs must be evaluated through a multi-dimensional lens, considering not only the power-energy capacity but also the response speed and duration of the aggregated resources. Furthermore, the systematic categorization of flexible resource control within the VPP framework [16] suggests that bi-level optimization and hierarchical aggregation are essential to bridge the gap between microscopic stochastic behaviors of individual EVs and macroscopic grid requirements.

After accurately obtaining the load baseline, the core of EV aggregation lies in precisely depicting its physically adjustable boundaries and transforming massive discrete nonlinear variables into low-dimensional continuous variables. Current mainstream methods deeply integrate precise geometric reconstruction, dynamic clustering, and subjective willingness games:

Dynamic Clustering Algorithms and Generalized Virtual Battery (VB) Equivalents: Facing the “curse of dimensionality,” advanced clustering and equivalent models provide excellent dimensionality reduction solutions for engineering dispatch. Cutting-edge literature points out that utilizing hierarchical clustering and probabilistic mixture models can accurately group users based on their historical spatio-temporal charging characteristics, thereby efficiently deriving the load profiles of millions of EVs in just 45 s [43]. When formulating joint scheduling strategies for large-scale EVs and thermal units, the K-means clustering algorithm is widely used to group vehicle populations, drastically reducing the computational burden [44]. At the modeling equivalence level, the “Virtual Battery” concept is widely adopted. Researchers innovatively proposed the dual metrics of Dynamic Equivalent Battery (DEB) and Value of DEB (VoDEB) to dynamically quantify the adjustable capacity boundaries of EV fleets in demand response [45]. Moreover, integrating EV flexibility based on the virtual battery model optimally facilitates the low-carbon economic dispatch of integrated energy systems, achieving V2G revenues that offset battery aging costs by up to 91% [46].

Polyhedron Projection and Minkowski Sum of Feasible Operation Region (FOR): Compared to simplified virtual batteries, advanced mathematical topological models provide strictly precise physical boundaries. Ref. [47] proposes an iterative method based on the high-dimensional polyhedron projection principle, transforming the exact Aggregation Feasible Region (AFR) into a set of linear inequalities whose complexity grows linearly with the number of EVs. Targeting heterogeneous EV populations, ref. [48] deeply investigates computing maximum-volume inner approximations of the exact Minkowski sum of individual flexibility sets through affine transformations of convex polytopes, equivalencing hundreds of thousands of EVs into a single compliant generalized polytope.

Nonlinear Quantification of Multi-attribute Willingness and Psychological Effects: Traditional aggregation models mostly assume users are “optimally rational economic men.” Ref. [49] innovatively introduces the psychological effect of range anxiety based on the Weber-Fechner law, truly quantifying users’ scheduling costs and subjective willingness. Additionally, facing distribution network voltage violation risks, the latest Bi-level Safe SAC algorithm combines K-means clustering to assess EV scheduling potential. This willingness modeling, combining multi-attribute attitudes and price incentives, drastically enhances robustness against real-world user default risks [50].The detailed technical comparison is shown in

Table 3. Comparison of advanced algorithms for EV load forecasting and aggregation modeling.

Core Technical Dimension	Representative Algorithm/Model	Core Advantages & Breakthroughs	Limitations/Applicable Scenarios
Spatio-temporal Forecasting	STGCN + Bi-GRU-Seq2Seq	Efficiently extracts spatial spillover effects and long-term temporal dependencies; optimally reduces MAE by 9–16%.	Highly dependent on multi-source data; suitable for city-level public charging station clusters.
Probabilistic & Scenario Gen.	Diffusion Models + MetaProbformer	Breaks historical data limits to solve “cold start” issues; precisely envelopes extreme fluctuations without massive training data.	Generation process is time-consuming; ideal for large-scale Monte Carlo substitution in day-ahead dispatch.
Privacy-preserving Crypto.	Lattice Cryptography + $ϵ$-DP	Fundamentally protects privacy and mitigates FDIA; hash and signcryption avoid heavy computational overhead.	Higher hardware requirements for edge gateways; suitable for decentralized, multi-aggregator environments.
Physical Topology Reduction	Minkowski Sum + Polyhedron	Mathematically rigorous; aggregates hundreds of thousands of EVs into a single compliant generalized polytope without physical violations.	Suffers from a severe “curse of dimensionality”; relies on maximum-volume inner approximation for acceleration.
Clustering & Simulation	Hierarchical/K-means + ACN-Sim	Extremely high reduction efficiency (deriving millions of loads in 45 s); open-source baseline lowers the data barrier.	Engineering equivalent method with approximation errors; suitable for ultra-fast dispatch of heterogeneous fleets.
Psychological Quantification	Weber-Fechner Law + DEB/VoDEB	Breaks the “rational economic man” assumption; truly quantifies users’ “loss aversion” towards battery range anxiety.	Psychological parameters vary subjectively; primarily used for designing V2G micro-compensation mechanisms.

.

Through advanced spatio-temporal graph network forecasting, physical polytope reconstruction based on the Minkowski sum, and psychological willingness quantification accounting for bounded rationality, VPPs can now clearly grasp the spatio-temporal distribution, true capacity boundaries, and credit default risks of their internal controllable resources with extremely high fidelity. However, how to push these highly complex “generalized deterministic energy storage” units into diversified electricity markets and maximize economic benefits in energy, ancillary services, and even carbon trading markets requires more sophisticated cross-market bidding games and coordinated clearing strategies. The relevant mechanisms will be deeply discussed in Section 5.

5. Multi-Market Co-Optimization and Bidding Strategies

In the process of advancing the construction of novel power systems, the profitability models of Virtual Power Plants (VPPs) and Electric Vehicle (EV) aggregators are no longer confined to a single day-ahead energy market. With the deregulation of electricity markets and the tightening of carbon emission quota mechanisms, aggregating dispersed flexible loads to collaboratively participate in the multi-market coupled bidding of “energy—peak shaving/frequency regulation ancillary services—carbon trading” has become the fundamental driving force for enhancing overall operational efficiency. However, bidding decisions in a multi-market environment face multiple challenges, including cross-market capacity allocation coupling, extreme volatility of market prices, and conflicting interests among diverse entities. Current research conducts in-depth optimal designs primarily from three core dimensions: the positioning of bidding game roles and electricity-carbon/flexibility joint trading, multi-level games and internal P2P profit allocation, and hybrid uncertainty risk aversion encompassing distributionally robust optimization (DRO).

5.1. Market Role Positioning, Differentiated Bidding, and Electricity-Carbon Joint Clearing Mechanism

In the evolving market environment, the coordinated bidding of VPPs across coupled markets—such as energy, spinning reserve, and peak-shaving—has become a focal point of research. Recent optimization models for VPPs participating in joint energy and peak-shaving markets [27] demonstrate that considering the differentiated bidding strategies of various power supply and flexible load equipment can significantly enhance the overall market competitiveness and regulation benefits of the aggregator. This aligns with the broader goal of enhancing system response capability through market-based mechanisms [1].

Based on the aggregator’s market share and influence, traditional business models unconstrained by physics are gradually evolving into Technical Virtual Power Plants (TVPPs). Cutting-edge literature proposes a bi-level stochastic mixed-integer linear programming (MILP) trading model for TVPPs. This model maximizes VPP profit at the upper level and accurately simulates market clearing, incorporating Locational Marginal Prices (LMP) at the lower level, ensuring day-ahead energy bids are strictly physically feasible on the underlying network [18].

Quantitative Benefits of Energy-Carbon-Reserve Joint Game and Stepped Carbon Pricing: A single electricity price driver can no longer meet low-carbon requirements; the latest research constructs a coordinated optimization model targeting the joint “energy-reserve-carbon” market [22]. To further excavate low-carbon potential, studies designed incentive-based stepped carbon price constraints and bi-level configuration models. Quantitative empirical evidence shows that introducing electricity-carbon coupled pricing significantly pushes up the bidding costs of high-carbon units, highlighting the arbitrage advantage of EVs as zero-carbon resources. Data confirm that this mechanism drastically reduces overall system carbon emissions by 32.2%, while the VPP aggregator’s total revenue contrarily increases by 11.69%, optimally validating the dual environmental and economic dividends of the joint market [46].

Emerging Local Flexibility Market and Capacity Bidding: As high-proportion renewable energy integration triggers local distribution network congestion, recent research proposes a flexibility-oriented bi-level stochastic-robust bidding strategy for VPPs. The upper-level VPP maximizes profit by exploiting locational flexibility, while the lower-level Local Energy Communities (LECs) minimize costs based on dynamic retail prices, truly completing the commercial closed loop of “energy-local flexibility” cross-market arbitrage [51].

5.2. Multi-Level Games, Nash Bargaining, and Internal P2P Allocation Between Aggregators and Users

After winning bids in external markets, the VPP still needs to resolve profit allocation issues among internal heterogeneous resources (especially private car owners). Because the aggregator pursues overall profit maximization while users pursue individual cost minimization, a typical non-cooperative game relationship is formed.

Generalized Nash Bargaining and Peer-to-Peer (P2P) Energy Sharing: Traditional centralized allocation faces immense unfairness. Ref. [52] proposes a contribution-driven peer-to-peer cooperative trading framework for multi-energy VPPs. This framework constructs an Asymmetric Nash bargaining optimization model based on contributions, where the bargaining power of each sub-entity is quantitatively determined by its “marginal economic gains” and “emission-reduction contributions,” achieving incentive-compatible Pareto optimal allocation.

Distributed Ultra-fast Solving via ADMM and Plummeting Costs: To protect privacy and break computing bottlenecks during P2P profit allocation, the Alternating Direction Method of Multipliers (ADMM) is introduced to solve the game model [53]. Combined with a differentiated bidding mechanism tailored to each sub-device (discarding traditional uniform pricing), the latest simulation data reveal that this dynamic mechanism comprehensively stimulates the enthusiasm of underlying entities and effectively resolves internal interest conflicts. It enables a notable maximum reduction of 50.34% in the joint optimal operating cost, providing a highly attractive financial prospect for the internal coordination of VPPs [28].

5.3. Distributionally Robust and Hybrid Uncertainty Modeling, Risk Aversion, and Behavior Quantification

In real-time markets, EV load demands and extremely volatile clearing prices constitute dual uncertainties. If the aggregator only adopts a risk-neutral strategy, it is highly vulnerable to massive financial losses. Researchers have therefore introduced top-tier risk decision-making tools:

Conditional Value-at-Risk (CVaR) and Hybrid “Risk-Seeking/Averse” Dual Preferences: Targeting extreme price jumps, cutting-edge research comprehensively introduces Conditional Value-at-Risk (CVaR) into bidding models, successfully averting the “tail risks” of real-time electricity prices. Even more groundbreaking, regarding the arbitrage needs of radical investors, ref. [27] proposes a hybrid uncertainty scheduling strategy integrating “risk-seeking and risk-averse preferences.” This strategy utilizes CVaR to control extreme losses in worst-case scenarios while introducing the Value at Best (VaB) metric to precisely capture “windfall profits,” achieving a perfect financial balance between extreme losses and potential excess profits.

Distributionally Robust Optimization (DRO) Based on Wasserstein Distance: Although CVaR measures risk, it relies on preset precise probability distributions. When facing extreme weather, the true distribution is highly uncertain. Therefore, Data-driven Distributionally Robust Optimization (DRO) is introduced into the bidding game. Authentic literature [54] constructs a two-stage Distributionally Robust Chance Constrained (DRCC) bidding model based on the Wasserstein distance, utilizing real samples to construct an Ambiguity Set and guaranteeing strict satisfaction of two-sided chance constraints. Additionally, targeting energy-constrained VPPs, researchers combined moment information to propose a two-sided distributionally robust convex reformulation method. This top-tier method strikes a perfect mathematical balance between the “over-optimism” of stochastic programming and the “over-conservatism” of pure robust optimization, strictly guaranteeing the expected profit lower bound under the worst-case probability distribution [55].The detailed technical comparison is shown in

Table 4. Quantitative summary of core literature on VPP and EV coordinated control.

Dimension	Refs.	Core Algorithm/Model	Objective/Scenario	Quantitative Highlights (with Baseline & Scenario Context)
1. Forecasting	[34,41]	Bi-GRU-Seq2Seq/FedAvg + LDP	Traffic-power coupled network (TPCN), Privacy-preserving	Finding: Reduces MAE by 9–16%. Scenario: Hourly forecasting using a multi-scale temporal window strategy (hour-day-week). Baseline: Single-scale temporal forecasting models.
2. Aggregation	[47,48]	Minkowski Sum + Polyhedron Projection	Exact Feasible Operation Region (FOR) extraction	Finding: Translates complex heterogeneous constraints into low-dimensional polytopes, avoiding the “curse of dimensionality”. Assumption: Aggregation of hundreds of thousands of EVs without physical violations
3. Joint Market	[22,46]	Bi-level Optimization + Stepped Carbon Trading	Energy-Carbon-Reserve market co-optimization	Finding: Reduces system carbon emissions by 32.2%; increases overall aggregator revenue by 11.69%. Scenario: Integrated energy system with P2G-CCS and EV participation. Baseline: High-carbon units bidding without the incentive-based stepped carbon price constraints
4. Profit Allocation	[52,56]	Asymmetric Nash Bargaining + ADMM	P2P internal trading & privacy-preserving allocation	Finding: Reduces joint optimal costs by a maximum of 50.34%. Scenario: Contribution-driven P2P trading with dynamic bidding tailored to each sub-device. Baseline: Traditional uniform pricing and centralized allocation frameworks.
5. Risk Aversion	[28,54]	Distributionally Robust Optimization (DRO)	Day-ahead bidding under extreme price volatility	Finding: Strictly guarantees expected profit lower bounds under worst-case scenarios. Assumption: Employs Wasserstein metric to build ambiguity sets from real samples.
6. Real-time Control	[57,58]	Stochastic Model Predictive Control (SMPC)	AGC tracking and distribution network physical security	Finding: Reduces maximum tracking error of AGC power by 53.2%; total network loss by 10%. Scenario: EV cluster tracking high-frequency AGC commands.
7. AI Control	[59,60]	Multi-Agent Deep Reinforcement Learning (MADRL)	Fully decentralized control & continuous-discrete actions	Finding: Increases frequency regulation aggregator’s total revenue by 16.25%. Baseline: Traditional Mixed-Integer Linear Programming (MILP) limits
8. Degradation Mgmt.	[61,62]	High-fidelity Electrochemical + Multi-objective	Cycle & calendar aging mitigation during V2G	Finding: Reduces battery cycle aging by 67%, improves grid congestion by 90%, and reduces end-user’s energy cost by 88.2%. Scenario: Multi-objective techno-economic-environmental (TEA) scheduling. Baseline: Uncontrolled charging behaviors.

.

When evaluating the economic benefits of multi-market bidding strategies, it is crucial to recognize that the revenue metrics reported in various studies are heavily dependent on their specific evaluation frameworks. The quantitative profitability claims are not absolute; they vary drastically based on the chosen baseline (e.g., unmanaged charging versus time-of-use pricing), the specific market scenario (e.g., day-ahead energy versus regulation reserve), and critical assumptions (e.g., regional electricity tariffs and battery wear models). Therefore, direct cross-study numerical comparisons should be avoided, and results must be interpreted within their unique contextual boundaries.

6. Hierarchical Control Strategies for AGC

After successfully bidding in the day-ahead market, the core task of a Virtual Power Plant (VPP) is to accurately track Automatic Generation Control (AGC) or Load Frequency Control (LFC) commands issued by the upper-level dispatch center on a seconds or minutes timescale. Unlike traditional units, VPPs face massive discrete and micro-capacity heterogeneous EVs. They must break through the “curse of dimensionality” to achieve ultra-fast allocation of high-frequency commands and optimal control of battery degradation, while ensuring system frequency stability and cyber-physical security.

The hierarchical control strategies for AGC discussed in this section are intrinsically coupled with the multi-timescale framework depicted in Figure 4. The transition from “bidding” to “execution” is governed by a Reference Trajectory Mapping mechanism. Specifically, the day-ahead and intraday stages (discussed in Section 3) determine the optimal energy baseline and regulation capacity reservation. These upper-level decisions act as the “feasible corridor” for the real-time AGC layer. The hierarchical controller must then decompose the grid-level frequency deviation signal into sub-second power set-points for individual EVs, ensuring that the cumulative real-time energy exchange remains consistent with the pre-cleared market schedules while managing the stochasticity of EV plug-in behaviors.

6.1. Breaking the Curse of Dimensionality: From ADMM to MADRL

Upon receiving the Area Control Error (ACE) signal, aggregators must rapidly decompose the total frequency regulation command down to massive internal nodes. The massive concurrent coupling of continuous and discrete variables easily triggers the computational collapse of centralized solving (the curse of dimensionality).

Distributed Optimization and ADMM Architecture: To eliminate excessive reliance on centralized controllers, cutting-edge research transforms VPP resource deployment into a fully distributed cooperative optimization problem. Authentic literature indicates that by utilizing the Alternating Direction Method of Multipliers (ADMM) and consensus optimization theory, each distributed energy resource can be treated as an independent agent. This architecture not only performs computations locally and communicates only with neighboring nodes—drastically improving system scalability—but also fundamentally achieves user data privacy protection [56].

MADRL and Mixed-Action Space Dimensionality Reduction: Facing the complex strong coupling of continuous power allocation and discrete charge/discharge states, traditional Deep Reinforcement Learning (DRL) often struggles to balance scalability and stability. The latest research proposes a mixed-action DRL algorithm based on the Twin Delayed Deep Deterministic Policy Gradient (TD3). This method innovatively introduces a latent representation mechanism based on an Encoder–Decoder structure, optimally embedding the dependency between discrete and continuous actions, and providing a dimensionality-reduction solving scheme with the lowest operating cost for VPPs with extreme uncertainty [60]. Furthermore, addressing strong random disturbances in large-grid multi-region collaboration, MADRL integrating Deep Q-Networks and an improved Actor-Critic strategy (DDQN-CDP strategy) achieves ultra-fast convergence of frequency regulation among distributed nodes without sharing underlying private data. Its control performance comprehensively surpasses traditional centralized heuristic algorithms [63].

6.2. Communication Delays and Decentralized Security in Cyber-Physical Systems (CPS)

Participating in secondary frequency regulation services like AGC demands extremely low communication latency (typically sub-second or second-level). In real Cyber-Physical Systems (CPS), packet loss and time delays severely undermine system frequency stability.

Time-Delay Robust Control and Filter Reconstruction: Authentic literature points out that ignoring communication delays in the EV feedback link easily leads to instability in frequency regulation actions. Therefore, researchers designed an auxiliary frequency regulation control strategy based on vehicle information feedback considering communication delays. By introducing specialized filters, high-frequency regulation commands are safely and precisely allocated to EV clusters [57]. Furthermore, in closed-loop LFC system design, researchers constructed a robust PID-type load frequency control strategy considering time delays and uncertainties based on Linear Matrix Inequalities (LMI) and Particle Swarm Optimization (PSO) algorithms. Empirical evidence shows that this scheme effectively suppresses severe frequency fluctuations caused by time delays and provides powerful physical robustness against the uncertainties brought by EV cluster integration [58].

Decentralization via Blockchain and Inference Networks: To comprehensively resolve data privacy and single-point failure issues under existing aggregation frameworks, recent research constructed a fully decentralized aggregation framework based on a consortium blockchain. Based on this, researchers introduced an inference network and proposed a decentralized interaction algorithm based on Multi-Agent Soft Actor-Critic (IN-MASAC). This method enables EVs to autonomously optimize frequency regulation mileage without exposing private data, optimally balancing the timeliness of frequency response with the absolute security of underlying data [59].

6.3. High-Fidelity Battery Degradation Management and Multi-Objective Control

Frequently responding to AGC high-frequency charge/discharge commands inevitably accelerates the degradation of onboard lithium-ion batteries. How to endogenize actual physical battery wear and integrate it into the high-frequency control loop is the core challenge for maintaining the long-term healthy operation of VPPs.

Calendar/Cycle Aging Separation and High-Fidelity Predictive Control: Actual aging is a complex electrochemical process encompassing calendar aging and cycle aging. Ref. [62] adopted empirical models to strictly separate these two aging mechanisms, accurately evaluating the true degradation of V2G participating in primary frequency regulation based on the initial State of Charge (SOC), providing a quantitative basis for enhancing the system’s cost–benefit. Meanwhile, to achieve Fast Frequency Response (FFR) in microgrids, the latest Improved Model Predictive Control (IMPC) strategy was proposed. This strategy innovatively embeds deep learning-based battery degradation prediction directly into the closed-loop control, achieving optimal power sharing that minimizes battery aging degradation while coordinating local stationary energy storage and mobile EVs [64].

“Techno-Economic-Environmental” Multi-Objective Optimization Empirical Evidence: Transforming battery anti-aging from “theoretical derivation” into “real money” revenue is the focus of the engineering community. When defining different power system operation states, the optimization objective in normal states is directly anchored to minimizing frequency regulation costs, including battery degradation penalties [65]. More representatively, ref. [61] constructed a multi-objective coordinated optimization model accounting for end-user electricity bills, deep battery degradation, grid interaction efficiency, and system carbon emissions. Its real-world case studies strikingly demonstrate that advanced multi-objective optimization, while providing frequency regulation services, can drastically reduce battery aging by 67%, improve grid congestion by 90%, and reduce CO2 emissions by 34%. These quantitative data conclusively proves that under rational algorithmic dispatch, EV high-frequency regulation and battery health maintenance can strictly achieve a win-win situation.

7. Economic Assessments and Engineering Demonstrations

The commercial viability of VPP-aggregated EVs hinges on a rigorous Techno-Economic Assessment (TEA). Rather than a simple revenue calculation, a comprehensive TEA must capture the complex trade-offs between market compensation and the hidden costs of physical hardware execution.

7.1. Comprehensive Economic Evaluation Models

The foundation of EV aggregation economics lies in accurately modeling the Net Present Value (NPV) and Return on Investment (ROI) across the project lifecycle. The total operational costs are predominantly driven by battery degradation (both calendar and cycle aging) and the initial Capital Expenditure (CapEx) of bi-directional chargers. It is imperative to note that the quantitative economic metrics reported in the existing literature, as summarized in Table 3, are highly sensitive to their specific baselines (e.g., comparing smart charging against unmanaged “dumb” charging) and underlying assumptions (e.g., battery replacement costs and localized electricity tariffs). Therefore, these performance claims should be interpreted within their specific scenario boundaries rather than as absolute universal values.

7.2. Cost–Benefit Analysis Across Different Markets

Current economic evaluations reveal a critical consensus regarding market selection. Participating solely in day-ahead energy arbitrage often yields marginal or even negative net profits, as the revenue generated from peak-valley price spreads is frequently outweighed by the severe cycle aging costs induced by deep discharging. Conversely, multi-market coupling—particularly providing high-frequency ancillary services (e.g., AGC regulation)—significantly enhances profitability. Frequency regulation requires rapid but shallow cycling (narrow State-of-Charge fluctuations), which generates substantial capacity reservation fees while minimizing electrochemical battery wear, thus accelerating the payback period of V2G infrastructure.

7.3. Insights from Global Wide-Area Demonstrations

Theoretical economic models have been robustly validated by pioneering global Vehicle-to-Grid (V2G) demonstrations. Noteworthy initiatives, such as the Parker Project in Denmark and the Sciurus Project in the UK, have provided invaluable real-world operational datasets.

Table 5. Global V2G Demonstration Projects.

Demonstration Project/Region	Target Aggregated Resources	Core Market/Service	Measured Economic Benefits (with Baseline & Scenario)	Current Limitations & Assumptions
Parker Project (Denmark, Europe)	Branded vehicle fleets (e.g., Nissan Leaf)	Primary Frequency Regulation	Finding: Successfully validated technical feasibility across multiple brands.	Limitation: Relatively small scale; highly dependent on underlying communication protocol support from specific automakers.
Sciurus Project (UK, Europe)	Residential smart charge/discharge infrastructure	Routine grid services and Frequency Response	Finding: Annualized revenue per vehicle reaches £513. Scenario: Deep participation in high-end ancillary services (Grid Frequency Response) via an aggregator. Baseline: Limited to £340/year when only participating in conventional routine grid services.	Limitation: Long-term actual battery life degradation (5–10 years) under extreme high-frequency dispatch remains unverified.
INVENT Project (USA, North America)	Pure Electric School Buses	Peak Shaving & Capacity Market	Finding: Explored a novel business model utilizing idle periods as a massive flexible resource pool.	Limitation: Vehicle operation is strictly restricted by school schedules, resulting in poor flexibility for real-time spot markets.
Provincial V2G Platforms (Shanghai, etc., China)	Public transit, logistics, and private car fleets	Demand Response (DR) and microgrid interaction	Finding: Net revenue of approx. 4000 RMB/year for private car owners. Scenario: Deep participation in provincial vehicle-grid interaction platforms for V2G demand response.	Assumption: Profitability models currently still heavily rely on unilateral government subsidies; cross-market free arbitrage is immature.

synthesizes the key configurations and economic findings of these leading global demonstrations. However, it must be emphasized that the figures drawn from these heterogeneous case studies are not directly comparable. The quantitative returns vary drastically based on the specific fleet size, the maturity of local power markets, and the presence of governmental subsidies during the pilot phase.

It is crucial to distinguish between demonstrated outcomes and prospective speculations within the current economic assessments of TVPPs. The micro-bilateral economic benefits and local ROI metrics (discussed in Section 5.1) are established results, having been verified through small-scale pilot projects and hardware-in-the-loop tests. However, the macro-level diffusion models and the projections of profitability at a multi-million wide-area scale remain largely prospective. These future projections are highly sensitive to unconfirmed policy subsidies and evolving cross-market arbitrage rules, which have not yet been fully realized in current industrial operations.

7.4. Commercialization Barriers and Policy Implications

While the unit economics and technical feasibility of V2G are proven in localized pilots, transitioning to million-scale EV aggregation faces substantial hurdles. The high CapEx for V2G chargers, combined with the lack of standardized communication protocols (such as unified OCPP 2.0.1 and ISO 15118 adoption), remains a primary barrier. Additionally, many wholesale markets still impose strict minimum capacity thresholds that effectively exclude distributed EV aggregators. Future mechanism designs must shift from purely subsidy-driven pilot models to self-sustaining business paradigms, facilitated by dynamic pricing mechanisms and verifiable carbon credit trading.

8. Techno-Economic Challenges

Although the theoretical framework of Virtual Power Plants (VPPs) aggregating massive electric vehicles (EVs) for grid interactions has matured in small-scale empirical studies, realizing wide-area commercial applications that scale from pilot projects to a multi-million level still faces multiple challenges. These span underlying computing bottlenecks, Cyber-Physical System (CPS) communication security, physical battery degradation, and external macro-policy barriers. Clarifying these obstacles is a necessary path to drive the evolution of next-generation smart grids.

8.1. Scalability Bottlenecks and the “Curse of Dimensionality” in Large-Scale Aggregation

When VPPs attempt to simultaneously incorporate distribution network power flow constraints and the charge/discharge states of massive EVs into an optimization framework, computational complexity explodes exponentially. Authentic literature points out that when formulating joint scheduling strategies for large-scale EVs and traditional thermal power units, adopting traditional centralized dispatching directly triggers a severe “dimension disaster.” This causes day-ahead and intraday market clearing to be entirely unfinishable within the specified timeframe, necessitating forced dimensionality reduction using algorithms like K-means clustering [44].

Solving Limits of Continuous-Discrete Mixed Action Spaces: Furthermore, in highly uncertain renewable energy environments, VPPs need to simultaneously control continuous variables (e.g., unit output, power allocation) and discrete variables (e.g., EV charge/discharge state switching). Cutting-edge research [60] profoundly reveals this pain point: in traditional mixed-action Deep Reinforcement Learning (DRL) methods, the algorithm’s scalability, control stability, and the deep dependency between discrete and continuous actions cannot be simultaneously accommodated at the mathematical foundation. Without introducing extremely complex latent representation mechanisms (such as an Encoder–decoder structure), traditional DRL agents quickly face the dead ends of “policy failure” and “non-convergence” when confronted with tens of thousands of nodes.

8.2. Communication Delays and Underlying Protocol Vulnerabilities in Cyber-Physical Systems

In real-time interactive secondary frequency regulation (e.g., LFC) scenarios, extremely high communication rates and ultra-low information latency are the baselines for ensuring the security of the physical grid. However, massive concurrent data is highly susceptible to packet loss and congestion in real network environments.

Physical Instability Triggered by Time Delays: Authentic literature [58] points out that in a deregulated power market environment, due to the inherent limitations of communication networks, there are inevitable communication delays when control signals are transmitted from the dispatch center to EV terminals. If the control system lacks delay margin robustness, these delays will couple with dynamic demand response, causing EVs—originally intended to smooth fluctuations—to instead become interference sources that excite severe oscillations in system frequency.

Security Black Holes in the OCPP Communication Protocol: Beyond physical time delays, data interception and malicious attacks at the network layer are the “Sword of Damocles” hanging over the commercialization of V2G. Currently, the Open Charge Point Protocol (OCPP), widely adopted globally, is used to coordinate communication between charging stations and central systems. However, top-tier review literature [65] issues a severe warning after in-depth investigation: early OCPP versions were almost entirely unprotected, and even the latest OCPP 2.0.1 version includes only certain limited security features. Specifically, the lack of mandatory end-to-end payload encryption makes VPPs highly vulnerable to False Data Injection Attacks (FDIA) and Man-in-the-Middle (MitM) interceptions. If attackers spoof State of Charge (SoC) data or manipulate high-frequency AGC dispatch commands, they can directly “blind” the VPP control center, triggering severe reverse power flows and physical grid instability. When facing such complex cyber-threats and private data theft, OCPP still possesses a massive number of unresolved open security issues. Furthermore, simply patching these vulnerabilities with traditional heavy Public Key Infrastructure (PKI) introduces severe cryptographic latency, directly conflicting with the sub-second response requirements of real-time AGC. The fragility of this underlying protocol directly leads to the stagnation of cross-regional, multi-brand, scaled V2G interconnection.

8.3. Disconnect Between Battery Physical Degradation Models and Real-Time Incentive Mechanisms

The accelerated consumption of battery cycle life is the greatest subjective barrier preventing private car owners from participating in VPP scheduling. However, in current engineering practices, there is a severe disconnect between the physical evaluation of battery degradation and the scheduling compensation mechanism.

Pricing Absence Under High Battery Costs: Although lithium-ion battery technology continues to advance, replacement costs remain exorbitantly high. Authentic literature explicitly states that battery aging is an extremely complex nonlinear process where time-driven calendar aging and depth-of-discharge-driven cycle aging must be strictly distinguished. Crucially, providing high-frequency V2G services inherently demands fast charging and discharging (high C-rates) to meet rapid grid commands. While fast charging enables rapid energy storage and grid support, it simultaneously generates severe internal thermal stress and unfavorable temperature increases, which exponentially accelerate battery cycle aging. However, to pursue solving speed, most existing VPP bidding models adopt oversimplified linear depreciation constants, entirely ignoring these vital charging speed and temperature variables. This causes the true physical wear of V2G to be severely underestimated when providing high-frequency services like primary frequency regulation. If high-fidelity empirical aging models based on the initial State of Charge (SoC), charging current ratios, and thermal dynamics cannot be accurately translated into real-time economic compensation metrics, it is fundamentally impossible to design incentive schemes that truly increase cost–benefits, consequently leading to a sustained slump in user response willingness.

8.4. Fragility of Business Models and Shocks from External Macro Environments

The vast majority of current V2G demonstration projects still heavily rely on government policy subsidies and have not yet formed independent business models with strong risk-resistance capabilities in a free-competition electricity spot market.

Testing Policy Sensitivity Under Extreme Uncertain Environments: When simulating the long-term global EV penetration rate using various diffusion models like Gompertz, Bass, and Generalized Bass, authentic literature [37] discovered a highly critical conclusion: external environmental variables (such as the construction speed of charging infrastructure) have a decisive impact on the evolution of EV load demand. More importantly, when facing shocks from extreme uncertain macro environments, such as sudden pandemics (e.g., COVID-19), the sensitivity exhibited by different policy measures varies drastically across different countries and clustered groups. This implies that VPP aggregators cannot solely rely on static subsidy expectations; instead, they must develop dynamic, stress-resistant business models capable of coping with sudden “black swan” events.

8.5. Algorithmic Paradigm Shift and Trade-Offs: A Critical Comparison of CVaR, MPC, and DRL

Although individual algorithms have achieved breakthroughs in specific scenarios, a critical cross-evaluation reveals fundamental trade-offs among Conditional Value-at-Risk (CVaR) optimization, Model Predictive Control (MPC), and Deep Reinforcement Learning (DRL) in VPP scheduling.

In the day-ahead multi-market bidding phase, CVaR-based optimization serves as an irreplaceable white-box tool. It provides rigorous mathematical guarantees for bounding extreme financial losses (tail risks) under price volatility. However, its heavy reliance on precise scenario generation and explicit formulation makes it computationally prohibitive for real-time dispatch when confronted with the discrete constraints of millions of EVs.

Conversely, to manage micro-physical constraints such as high-fidelity battery cycle aging and local voltage limits, MPC demonstrates exceptional robustness. Through receding horizon optimization, MPC optimally embeds nonlinear physical models into the control loop. Yet, its inherent requirement to solve complex constrained optimization problems online at every time step results in an exponential surge in computational latency, severely limiting its scalability in ultra-fast Automatic Generation Control (AGC) tracking.

To break this “curse of dimensionality,” DRL (and MADRL) paradigm shifts the computational burden offline. Its model-free, sub-second online inference capability makes it the optimal choice for high-frequency AGC command allocation among massive heterogeneous nodes. Nevertheless, the black-box nature of DRL constitutes its fatal flaw: it fundamentally lacks the strict physical safety bounds of MPC and the theoretical economic bottom lines of CVaR. When confronted with unseen out-of-distribution market shocks or extreme grid contingencies, purely data-driven DRL agents are highly susceptible to catastrophic policy failures and physical constraint violations.

Therefore, no single algorithm can conquer the entire V2G operational spectrum. The future evolution must point toward a “Physics-Informed Neuro-Symbolic” architecture: utilizing CVaR to establish macro-economic boundaries, embedding MPC to outline physical safety envelopes, and deploying DRL within these strictly defined safe regions for ultra-fast continuous-discrete action execution.The detailed technical comparison is shown in

Table 6. Critical comparison of dominant optimization and control methodologies.

Evaluation Dimension	CVaR-Based Optimization	Model Predictive Control (MPC)	Deep Reinforcement Learning (DRL)
Primary Domain/Horizon	Day-ahead multi-market bidding	Intraday/Minute-level execution	Real-time/Sub-second AGC tracking
Uncertainty Handling	Explicit risk-bounding (Tail risks)	Receding horizon feedback correction	Implicit adaptation via environment exploration
Interpretability & Safety	High: Strict mathematical proofs for economic bounds	High: Explicit physical constraint formulation	Low: Black-box nature; poor out-of-distribution generalization
Online Computation	Heavy: Often involves complex MILP solving	Moderate: Solves optimization per time step	Extremely Light: O(1) neural network inference
Scalability (Millions of EVs)	Poor: Susceptible to the curse of dimensionality	Moderate: Relies on decomposition methods (e.g., ADMM)	Excellent: Decentralized execution via MADRL

.

8.6. Synthesis of Unresolved System-Level Trade-Offs: The “Impossible Triangles” of V2G

Transitioning from theoretical models to multi-million-scale commercialization requires confronting the harsh reality of system-level trade-offs. The challenges detailed in Section 4 through Section 6 are not isolated engineering hurdles; rather, they form deeply intertwined structural conflicts. Synthesizing these reveals several core unresolved trade-offs—effectively, the “impossible triangles” of current VPP coordination:

Trade-off 1: Aggregation Fidelity vs. Market Complexity. There is a fundamental computational tension between accurately representing physical realities and participating in advanced electricity markets. High-fidelity aggregation methods, such as extracting the exact Feasible Operation Region (FOR) using Minkowski sums, ensure absolute physical safety. However, when these rigorous, high-dimensional topological constraints are injected into a multi-coupled market clearing model (spanning energy, reserve, and carbon trading), the computational burden explodes. Currently, aggregators are forced into a rigid compromise: either sacrifice aggregation fidelity (using simplified virtual battery models) to participate in complex markets, or abandon multi-market arbitrage to ensure micro-level physical compliance.

Trade-off 2: Communication Latency vs. Data Privacy (The CPS Paradox). In the Cyber-Physical System (CPS) layer, real-time Secondary Frequency Regulation (e.g., AGC) demands sub-second communication latency to maintain grid stability. Simultaneously, user data must be strictly shielded from False Data Injection Attacks (FDIA) and unauthorized profiling. The trade-off lies in the fact that advanced privacy-preserving architectures—such as fully decentralized blockchain consensus or lattice cryptography (signcryption)—require heavy local computational overhead and multi-round peer-to-peer verification. This inherently introduces communication delays that can easily destabilize the physical frequency control loop. Balancing ultra-low latency with zero-trust privacy remains a severe architectural deadlock.

Trade-off 3: Battery Degradation vs. Multi-Market Arbitrage Revenue. At the micro-economic level, the VPP aggregator’s incentive to maximize profits directly contradicts the private EV owner’s desire to preserve battery health. To capture “windfall profits” in volatile real-time and frequency response markets, EVs are subjected to high-frequency, deep charge–discharge cycles. This aggressive arbitrage behavior accelerates nonlinear electrochemical degradation (cycle aging). Current uniform pricing and simplified linear compensation mechanisms are mathematically insufficient to cover these true, dynamic degradation costs. If the VPP restricts its dispatch frequency to protect the battery (prioritizing State of Health), it loses the capacity to compete in lucrative multi-market bidding, collapsing the underlying business model.

9. Future Research Directions

Based on the critical synthesis of current literature, future research on VPP-EV coordination can be strictly categorized into evidence-backed near-term advancements and prospective, speculative explorations.

9.1. Evidence-Backed Near-Term Directions

Current literature firmly establishes that addressing the communication latency and nonlinear degradation of batteries is an immediate priority. Near-term research should focus on edge-cloud collaborative control architectures. By shifting high-frequency regulation tasks (e.g., sub-second frequency response) to edge computing nodes directly embedded within EV charging stations, the burden on centralized VPP communication channels can be significantly alleviated, which is a mathematically validated trajectory in recent MPC-based studies.

9.2. Speculative Themes: Blockchain and Large Language Models (LLMs)

Beyond established results, emerging technologies present forward-looking speculation. While some conceptual studies propose LLMs and blockchain as future panaceas, these remain largely prospective. Rather than acting as immediate “core driving forces,” LLMs could specifically assist in processing unstructured market data or generating automated bidding codes under complex regulations. However, their high inference latency and hallucination risks make them currently unsuitable for real-time physical dispatch. Similarly, while blockchain offers decentralized trust for P2P trading, its massive computational overhead contradicts the ultra-fast control requirements of active distribution networks, necessitating lightweight consensus mechanisms in future explorations.

9.3. Applications in Industrial and Power Electronics

To bridge the gap between theoretical algorithms and real-world deployment, future frameworks must integrate deeply with the industrial and power electronics sectors. On the electronics front, the adoption of Wide-Bandgap (WBG) semiconductors, such as Silicon Carbide (SiC) and Gallium Nitride (GaN), in bi-directional V2G inverters is a critical direction. These materials can drastically reduce switching losses and thermal accumulation during high-frequency EV discharging, directly mitigating the battery degradation issues discussed in Section 6. On the industrial front, integrating EV aggregators into industrial microgrids—where factory electric fleets participate in demand response alongside heavy machinery—presents a highly practical application scenario that can substantially lower peak tariff penalties for manufacturers.

9.4. Deep Integration of Cross-Domain Trading and Blockchain Smart Contracts

With the popularization of P2P and V2X trading, achieving extremely high-frequency, tamper-proof profit allocation in a trustless environment is the ultimate test for mechanism design. Recent research comprehensively introduces underlying Blockchain architectures. Ref. [66] proposed a fully decentralized V2V energy trading system prototype and utilized DRL models for dynamic trading matches. In particular, its embedded sharding consensus mechanism demonstrated extremely strong throughput and anti-attack capabilities in empirical tests. Furthermore, targeting massive concurrent vehicular trading, ref. [67] proposed a novel scheme combining spatio-temporal network optimization with blockchain. By converting dynamic routing into parallelizable static flow allocation and embedding real-time constraints via distributed matrix operations, this smart contract system achieved a significant 23.6% increase in computational throughput and a 31.2% rise in transaction success rates. In the future, highly scalable blockchain technology will undoubtedly become the sole credit cornerstone supporting wide-area, free arbitrage for tens of millions of EVs.

9.5. Limitations of This Review

While this paper provides an integrative review of EV-centric TVPPs, several limitations must be acknowledged. First, the systematic search was restricted to English-language databases (e.g., IEEE Xplore, Web of Science, Scopus), which poses a language restriction risk, potentially omitting significant regional developments and non-English technical reports from major EV markets like China and Europe. Second, the review’s coverage boundaries strictly focus on the grid-side and market-side impacts of V2G, without delving into the chemical engineering aspects of battery cell manufacturing. Finally, given the rapid industrial evolution of charging protocols (e.g., OCPP), there is an inherent omission risk regarding the very latest unpublished industry white papers and proprietary corporate test data that have not yet undergone formal academic peer review.

10. Conclusions

Against the backdrop of the accelerated low-carbon transition, the uncoordinated integration of massive electric vehicles poses severe challenges to distribution networks, yet offers substantial flexible regulation potential. This review systematically synthesizes the theoretical framework of EV-centric Technical Virtual Power Plants (TVPPs) participating in multi-coupled markets. The core conclusions of this review are summarized as follows:

• Physical-Economic Bidirectional Mapping: Traditional Commercial VPPs fail to guarantee local grid security. Incorporating nonlinear power flow constraints and exact Feasible Operation Region (FOR) extraction methods is necessary to safely translate physical EV flexibility into market capacity.
• Multi-Market Coupling Mechanisms: Bidding strategies have evolved from single energy markets to complex multi-agent architectures, incorporating stepped carbon trading and green certificates. However, aligning multi-timescale optimization from day-ahead bidding to real-time execution remains mathematically challenging.
• Cross-Evaluation of Real-Time Control: While Deep Reinforcement Learning (DRL) demonstrates advantages in breaking the “curse of dimensionality” and handling unknown parameters, it still lacks the absolute stability guarantees provided by Model Predictive Control (MPC) during extreme grid contingencies.
• Core Unresolved Trade-offs: The industrial deployment of V2G through TVPPs remains hindered by critical trade-offs, most notably the tension between aggregation fidelity, communication latency, and the accelerated battery degradation caused by high-frequency market participation.