I develop an aggregated frequency regulation trading mechanism for heterogeneous electric vehicle resources. The central motivation is straightforward: a large number of electric vehicle users can provide valuable frequency regulation capacity, but the actual available capacity is not equal to the sum of nameplate charging and discharging powers. Electric vehicle users differ in travel patterns, initial state of charge, parking duration, battery anxiety, incentive sensitivity, and willingness to accept battery degradation. If an aggregator ignores these differences, it will overestimate the regulation potential of an electric vehicle cluster, submit aggressive bids, and expose itself to real-time deviation penalties. I therefore treat user response, physical aggregation, market bidding, real-time power allocation, and revenue sharing as one connected problem rather than as separate modules.
My mechanism has five connected layers. First, I quantify the response willingness of heterogeneous electric vehicle users through a membership-function and mental-accounting model. Second, I aggregate individual electric vehicle feasible regions into charging-station-level virtual battery regions by using the Minkowski sum. Third, I embed power allocation into a stochastic bidding model so that the aggregator co-optimizes energy and frequency regulation bids. Fourth, I design a real-time power allocation model that includes the marginal future value of stored energy through a Lagrange multiplier. Fifth, I evaluate each charging station along four contribution dimensions and allocate coalition revenue through an improved Shapley value model. The entire design is intended to keep the electric vehicle aggregator profitable while preserving fairness among heterogeneous electric vehicle charging stations.

Mechanism Overview and Notation
I consider an electric vehicle aggregator that coordinates multiple electric vehicle charging stations. Each station contains heterogeneous electric vehicle users. The aggregator collects battery parameters, state of charge, plug-in and plug-out times, expected departure state of charge, and user response information. It then evaluates the feasible regulation region of each station and submits energy and frequency regulation bids to the market. When regulation signals arrive, the aggregator allocates the required power among stations. After settlement, the aggregator deducts its operating cost and distributes the remaining coalition value according to multi-dimensional contributions. The key notation is summarized in Table 1.
| Symbol | Meaning |
|---|---|
| \(p_r\) | Incentive price offered to an electric vehicle user |
| \(D\) | Depth of discharge during regulation response |
| \(C\) | Current rate during regulation response |
| \(S_0\) | Initial state of charge of an electric vehicle battery |
| \(T_{on}\) | Parking duration or dwell time of an electric vehicle |
| \(K_e\) | Perceived intensity of the incentive price |
| \(\mathbf{U}_a\) | Membership vector of influencing factor \(a\) |
| \(B_a(\cdot)\) | Mental-account value function of factor \(a\) |
| \(V\) | Comprehensive psychological utility of response |
| \(\Lambda_{n,t}\) | Aggregated feasible region of electric vehicle charging station \(n\) |
| \(P_t\) | Energy bid or baseline power at time \(t\) |
| \(R_t\) | Frequency regulation capacity bid at time \(t\) |
| \(\pi_s\) | Normalized regulation signal in scenario \(s\) |
| \(\lambda_{n,t}^{S}\) | Lagrange multiplier of the station energy constraint |
| \(\varphi_n\) | Comprehensive contribution coefficient of station \(n\) |
| \(\phi_n(v)\) | Traditional Shapley value of station \(n\) |
| \(\phi_n^{*}(v)\) | Improved Shapley value of station \(n\) |
I use the phrase electric vehicle repeatedly because the physical and economic behavior of an electric vehicle is fundamentally different from that of a stationary battery. An electric vehicle is mobile, user-constrained, and socially embedded in travel behavior. The aggregator cannot command an electric vehicle as if it were a purely technical asset. The response decision of an electric vehicle user is a bounded-rational decision under economic incentive, battery degradation anxiety, and departure-time anxiety. This is why I begin with user response rather than with power capacity.
User Response Decision for Heterogeneous Electric Vehicle Users
I define five influencing factors that jointly describe the decision boundary of an electric vehicle user: incentive price \(p_r\), depth of discharge \(D\), current rate \(C\), initial state of charge \(S_0\), and parking duration \(T_{on}\). These five factors cover three psychological dimensions. The incentive price represents economic gain. Depth of discharge and current rate represent battery degradation cost. Initial state of charge and parking duration represent energy security and departure anxiety. I summarize the judgment sets in Table 2.
| Influencing factor | Judgment set | Interpretation |
|---|---|---|
| Incentive price \(p_r\) | \(\{p_r^{L},p_r^{M},p_r^{H}\}\) | Low, medium, and high incentive price |
| Depth of discharge \(D\) | \(\{D^{ad},D^{inad}\}\) | Reasonable and unreasonable discharge depth |
| Current rate \(C\) | \(\{C^{ad},C^{inad}\}\) | Reasonable and unreasonable current rate |
| Initial state of charge \(S_0\) | \(\{S_0^{L},S_0^{M},S_0^{H}\}\) | Low, medium, and high initial state of charge |
| Parking duration \(T_{on}\) | \(\{T_{on}^{L},T_{on}^{H}\}\) | Insufficient and abundant parking duration |
I use a trapezoidal membership function to represent the fuzzy boundary of each judgment level. For a general variable \(x\), the membership function is
$$
\mu(x;a,b,c,d)=
\begin{cases}
0, & x<a,\\ &="" 0,="" 1,="" \frac{d-x}{d-c},="" \frac{x-a}{b-a},="" a\le="" b\le="" c,\\="" c<x\le="" d,\\="" x="" xd.
\end{cases}
$$
For the incentive price, I use the Weber-Fechner law because the marginal psychological impact of an electric vehicle incentive decreases as the price rises. I define the perceived intensity as
$$
K_e=k\lg p_r+c,
$$
and normalize it as
$$
K_e=\frac{\lg p_r-\lg p_r^{\min}}{\lg p_r^{\max}-\lg p_r^{\min}}.
$$
Because electric vehicle users are heterogeneous, I introduce a perception deviation \(\Delta K\) and obtain a perception band
$$
K_e^{s}=K_e\pm\Delta K.
$$
This band produces fuzzy price boundaries. I map the lower, middle, and upper perception zones into trapezoidal membership functions for low, medium, and high incentive prices. The same logic applies to initial state of charge, because range anxiety is a convex function of remaining energy for many electric vehicle users. For depth of discharge and current rate, I use acceptable thresholds \(D_b\) and \(C_b\), together with acceptance coefficients \(r_D\in(0,1)\) and \(r_C\in(0,1)\), to construct a transition interval from acceptable to unacceptable battery stress. For parking duration, I compute the minimum charging time required to reach the expected departure state of charge:
$$
t_{need}=\frac{(S_{exp}-S_0)E}{P_{\max}\eta_{ch}}.
$$
I then define a tight threshold and an abundant threshold as
$$
t_1=t_{need},\qquad t_2=r_t t_{need},\qquad r_t>1.
$$
These thresholds allow me to distinguish an electric vehicle user who is forced to charge at maximum power from an electric vehicle user who has a comfortable charging buffer. The membership output for factor \(a\) is a vector
$$
\mathbf{U}_a=\{u_{a1},u_{a2},\ldots,u_{ab}\},
$$
where \(u_{ab}\) is the degree to which factor \(a\) belongs to judgment level \(b\). I convert the membership vector into a single score through
$$
x_a=\sum_{i=1}^{b}u_{ai}s_i,
$$
where \(s_i\) is the score of level \(i\). A higher score means a stronger positive satisfaction for gain factors or a stronger anxiety for cost factors, depending on the account.
I then apply mental accounting theory. An electric vehicle user does not merge all attributes into one linear utility. Instead, the user evaluates economic gain, battery degradation cost, and energy security in separate psychological accounts. The gain account uses the expected incentive price as a reference point:
$$
B_1(x_1)=
\begin{cases}
(x_1-x_1′)^{\chi}, & x_1\ge x_1′,\\
-\lambda(x_1′-x_1)^{\delta}, & x_1<x_1'. $$=""
The cost account represents battery degradation anxiety caused by depth of discharge and current rate. When the stress exceeds the tolerance reference point, the user experiences loss aversion:
$$
B_a(x_a)=
\begin{cases}
-\theta(x_a-x_a’)^{\alpha}, & x_a\ge x_a’,\\
(x_a’-x_a)^{\beta}, & x_a<x_a', $$=""
The security account represents departure anxiety caused by low initial state of charge and short parking duration. When the security margin is below the reference point, the user experiences a negative utility:
$$
B_a(x_a)=
\begin{cases}
(x_a-x_a’)^{\sigma}, & x_a\ge x_a’,\\
-\upsilon(x_a’-x_a)^{\psi}, & x_a<x_a', $$=""
The comprehensive utility of an electric vehicle user is
$$
V=\omega_1 B_1(x_1)+\omega_2\left[B_2(x_2)+B_3(x_3)\right]+\omega_3\left[B_4(x_4)+B_5(x_5)\right].
$$
I set the response threshold to zero. If \(V>0\), the electric vehicle user decides to respond to the aggregator. The account weights \(\omega_1,\omega_2,\omega_3\) are heterogeneous across electric vehicle user types. I report the weights in Table 3. These weights reflect different electric vehicle usage patterns: a private electric vehicle user may care about gain and security in a balanced way, a ride-hailing electric vehicle user faces severe time pressure and high battery cycling, and a commuter electric vehicle user has relatively regular parking behavior.
| Electric vehicle user type | Gain account weight | Cost account weight | Security account weight |
|---|---|---|---|
| Private electric vehicle | 0.4 | 0.3 | 0.3 |
| Ride-hailing electric vehicle | 0.3 | 0.3 | 0.4 |
| Commuter electric vehicle | 0.4 | 0.2 | 0.4 |
At the cluster level, I define the user response rate of an electric vehicle charging station as
$$
\gamma=\frac{M_w}{M_{total}},
$$
where \(M_w\) is the number of electric vehicle users who are willing to respond and \(M_{total}\) is the total number of electric vehicle users in the station. This response rate is not a cosmetic indicator. It directly scales the feasible regulation region of the station. In my simulation, the response rate differs sharply across electric vehicle user types. I summarize the results in Table 4.
| Electric vehicle user type | Response rate | Main behavioral interpretation |
|---|---|---|
| Private electric vehicle | 65.86% | Balanced account weights and long parking duration support moderate response |
| Ride-hailing electric vehicle | 27.48% | High operating intensity, low initial state of charge, and high departure anxiety suppress response |
| Commuter electric vehicle | 74.85% | Regular daytime plug-in pattern, abundant dwell time, and low battery-cost weight encourage response |
I also compute station-level response rates. The station heterogeneity is a direct mapping of its internal electric vehicle user composition. The results are shown in Table 5. A station dominated by private electric vehicle users has a different response rate from a station dominated by ride-hailing electric vehicle users, even if the number of chargers is similar. This is a central reason why I do not use a single average response rate for the entire electric vehicle aggregator.
| Electric vehicle charging station | Response rate |
|---|---|
| Station 1 | 45.07% |
| Station 2 | 55.59% |
| Station 3 | 50.75% |
| Station 4 | 70.17% |
Aggregated Regulation Potential of Electric Vehicle Clusters
After I determine which electric vehicle users are willing to respond, I aggregate their feasible power-energy regions. A single electric vehicle is too small to participate in the frequency regulation market. The aggregator must form a cluster-level virtual battery. A naive linear superposition of charging powers is fast but inaccurate because it ignores the coupling between power limits, energy limits, and plug-in state. I therefore use the Minkowski sum to preserve the feasible region of each electric vehicle.
For the \(m\)-th electric vehicle in station \(n\), the individual feasible region is described by
$$
0\le p_{m,t}^{ch}\le p_m^{ch,\max}Q_{m,t},
$$
$$
0\le p_{m,t}^{dis}\le p_m^{dis,\max}Q_{m,t},
$$
$$
s_{m,t}=s_{m,t-1}+\left(\eta_m^{ch}p_{m,t}^{ch}-\frac{p_{m,t}^{dis}}{\eta_m^{dis}}\right)\Delta t,
$$
$$
s_m^{\min}\le s_{m,t}\le s_m^{\max},
$$
$$
p_{m,t}^{ch}p_{m,t}^{dis}=0,
$$
where \(Q_{m,t}\in\{0,1\}\) is the plug-in state of the electric vehicle. The mutual exclusion constraint prevents simultaneous charging and discharging. The individual feasible region is a polytope in the power-energy space. I aggregate it through
$$
\Lambda_{n,t}=\bigoplus_{m=1}^{M_n}\Lambda_{nm,t},
$$
where \(\bigoplus\) denotes the Minkowski sum. The aggregated station-level virtual battery parameters are
$$
P_{n,t}^{ch,\max}=\sum_{m=1}^{M_n}p_{nm}^{ch,\max}Q_{nm,t},
$$
$$
P_{n,t}^{dis,\max}=\sum_{m=1}^{M_n}p_{nm}^{dis,\max}Q_{nm,t},
$$
$$
S_{n,t}^{\max}=\sum_{m=1}^{M_n}s_{nm}^{\max}Q_{nm,t}-\Delta S_{n,t},
$$
$$
S_{n,t}^{\min}=\sum_{m=1}^{M_n}s_{nm}^{\min}Q_{nm,t}-\Delta S_{n,t},
$$
where \(\Delta S_{n,t}\) represents the equivalent energy change caused by electric vehicle plug-in and plug-out events. The aggregated variables \(\{P_{n,t}^{ch,\max},P_{n,t}^{dis,\max},S_{n,t}^{\max},S_{n,t}^{\min},\Delta S_{n,t}\}\) form the regulation potential of the electric vehicle charging station. This potential is not a fixed number. It changes with electric vehicle arrival, departure, initial state of charge, and user response. The virtual battery of an electric vehicle cluster is therefore fundamentally different from a stationary battery with a fixed energy capacity.
I want to emphasize that the Minkowski sum is not merely a mathematical convenience. It gives the aggregator a precise boundary for bidding. When the aggregator submits a frequency regulation capacity, it must know whether the electric vehicle cluster can sustain that capacity for the required duration under both upward and downward regulation signals. The Minkowski sum retains the internal power-energy coupling and therefore reduces the risk that a high bid cannot be delivered by heterogeneous electric vehicle users.
Bidding and Power Allocation Co-Optimization
I formulate the aggregator bidding problem as a stochastic optimization model. The aggregator is a price taker in the energy and frequency regulation markets. It chooses a baseline power \(P_t\) and a regulation capacity \(R_t\) for each trading interval. The objective is to maximize total net revenue over the scheduling horizon:
$$
\max F=\sum_{t\in T}\left(F_t^{e}+F_t^{cap}+F_t^{mil}+F_t^{dep}-\sum_{s\in\Omega}\omega_{s,t}\left(C_{s,t}^{deg}+C_{s,t}^{mot}\right)\right),
$$
where the revenue components are
$$
F_t^{e}=p_t^{e}P_t\Delta t,
$$
$$
F_t^{cap}=p_t^{cap}R_t s_{reg}\Delta t,
$$
$$
F_t^{mil}=p_t^{mil}R_t a_t^{mil}s_{reg}\Delta t,
$$
$$
F_t^{dep}=p_t^{e}R_t\pi_s\Delta t.
$$
The cost term is
$$
C_{s,t}^{deg}+C_{s,t}^{mot}=\sum_{n=1}^{N}c_n^{cost}P_{s,t,n}^{dis}\Delta t,
$$
where \(c_n^{cost}\) includes the incentive payment to electric vehicle users and the battery degradation cost coefficient. The decision variables are \(P_t\) and \(R_t\), but the model embeds the power allocation variables \(P_{s,t,n}^{ch}\) and \(P_{s,t,n}^{dis}\) for each electric vehicle charging station and each scenario. This embedding is important because the bidding decision changes the feasible region of future power allocation, and the power allocation result changes the stored energy and therefore the future bidding capability. I therefore co-optimize bidding and allocation instead of solving them separately.
The constraints include nonnegative bids, power balance, charging and discharging limits, energy dynamics, minimum regulation duration, and energy security. The nonnegative bid constraints are
$$
P_t\ge0,\qquad R_t\ge0.
$$
The scenario power balance is
$$
P_t+R_t\pi_s=\sum_{n=1}^{N}\left(P_{s,t,n}^{dis}-P_{s,t,n}^{ch}\right).
$$
The station-level power limits are
$$
0\le P_{s,t,n}^{dis}\le P_{n,t}^{dis,\max},
$$
$$
0\le P_{s,t,n}^{ch}\le P_{n,t}^{ch,\max}.
$$
The energy dynamics of station \(n\) are
$$
S_{n,t}=S_{n,t-1}+\left(\eta_n^{ch}P_{s,t,n}^{ch}-\frac{P_{s,t,n}^{dis}}{\eta_n^{dis}}\right)\Delta t.
$$
The energy security constraints are
$$
S_{n,t}^{\min}\le S_{n,t}\le S_{n,t}^{\max}.
$$
I also enforce a minimum duration capability constraint so that the electric vehicle cluster can sustain regulation for a required time:
$$
\sum_{n=1}^{N}\left(S_{n,t}-S_{n,t}^{\min}\right)\ge P_t\Delta t+R_t\Delta t.
$$
This constraint prevents the aggregator from bidding a regulation capacity that can be delivered only for a few seconds. It is especially relevant for electric vehicle clusters because the energy buffer of an electric vehicle is constrained by the user’s departure time and battery capacity.
The real-time power allocation model is triggered when a regulation signal arrives. I define the allocation objective as the sum of real-time cost, future opportunity cost, and response deviation penalty:
$$
\min \sum_{n=1}^{N}\left(c_n^{cost}P_{n,t}^{dis}+\lambda_{n,t}^{S}\Delta S_{n,t}+M\left(\Delta P_t^{ch}+\Delta P_t^{dis}\right)\right),
$$
where \(\lambda_{n,t}^{S}\) is the Lagrange multiplier associated with the station energy constraint in the bidding model. This multiplier quantifies the marginal impact of a change in station energy on the remaining-horizon profit. If station \(n\) has a high future value of energy, \(\lambda_{n,t}^{S}\) induces the allocator to use less energy from that station, even if its current incentive cost is low. The term \(M\) is a large positive penalty coefficient that forces the aggregator to track the regulation signal as closely as possible.
The power allocation constraints include the signal-tracking balance
$$
\sum_{n=1}^{N}\left(P_{n,t}^{dis}-P_{n,t}^{ch}\right)+\Delta P_t^{dis}-\Delta P_t^{ch}=P_t+R_t\hat{\pi}_t,
$$
the charging and discharging limits
$$
0\le P_{n,t}^{dis}\le P_{n,t}^{dis,\max},
\qquad
0\le P_{n,t}^{ch}\le P_{n,t}^{ch,\max},
$$
the nonnegative deviation constraints
$$
\Delta P_t^{ch}\ge0,\qquad \Delta P_t^{dis}\ge0,
$$
and the instantaneous energy security constraint
$$
S_{n,t}^{\min}\le S_{n,t-1}+\left(\eta_n^{ch}P_{n,t}^{ch}-\frac{P_{n,t}^{dis}}{\eta_n^{dis}}\right)\hat{\Delta}t\le S_{n,t}^{\max}.
$$
I use \(\hat{\pi}_t\) and \(\hat{\Delta}t\) for the real-time signal value and the signal interval. The allocation model is a small linear program because the Lagrange multipliers have already converted future opportunity cost into a linear term. This is computationally attractive for an electric vehicle aggregator that receives a regulation signal every few seconds.
| Module | Decision variable | Main purpose |
|---|---|---|
| User response model | Response decision \(V>0\) | Screen electric vehicle users by bounded rationality |
| Minkowski aggregation | \(\Lambda_{n,t}\) | Build station-level virtual battery feasible region |
| Bidding model | \(P_t,R_t\) | Submit energy and frequency regulation bids |
| Real-time allocation | \(P_{n,t}^{ch},P_{n,t}^{dis}\) | Track regulation signal with future opportunity cost |
| Revenue allocation | \(\phi_n^{*}(v)\) | Distribute coalition value by multi-dimensional contribution |
Multi-dimensional Contribution and Improved Shapley Allocation
After the electric vehicle aggregator earns market revenue, it must distribute the coalition value among electric vehicle charging stations. A simple proportional allocation cannot reflect the heterogeneous service capability of different electric vehicle charging stations. The traditional Shapley value captures the average marginal economic contribution of each station, but it ignores non-economic dimensions such as regulation mileage, response speed, and actual regulation cost. I therefore build a four-dimensional contribution evaluation system and use it to improve the Shapley value.
I define four contribution indicators. The first is frequency regulation capacity contribution:
$$
x_{n1}=\sum_{t=1}^{T}R_{n,t}\Delta t.
$$
This indicator reflects the total regulation capacity that electric vehicle charging station \(n\) can provide over the scheduling horizon. The second is frequency regulation mileage contribution:
$$
x_{n2}=\sum_{j=2}^{N_{reg}}\left|P_{n,j}^{reg}-P_{n,j-1}^{reg}\right|.
$$
This indicator reflects how actively the electric vehicle charging station follows the regulation signal. The third is response rate contribution:
$$
x_{n3}=\frac{1}{N_{reg}}\sum_{j=1}^{N_{reg}}t_{n,j}^{delay},
$$
where \(t_{n,j}^{delay}\) is the response delay of station \(n\) under the \(j\)-th regulation instruction. A smaller delay means a stronger contribution, so this is a negative indicator. The fourth is frequency regulation cost contribution:
$$
x_{n4}=\sum_{j=1}^{N_{reg}}c_n^{cost}P_{n,j}^{dis}\Delta t_{reg}.
$$
This indicator reflects the implicit cost incurred by the electric vehicle charging station when it provides regulation service. A higher cost means that the station sacrifices more economic value to support the coalition, so it should receive appropriate recognition.
I normalize the contribution indicators to a comparable scale. For positive indicators, I use
$$
\hat{x}_{nr}=\frac{x_{nr}}{\sum_{n=1}^{N}x_{nr}}.
$$
For negative indicators, I use
$$
\hat{x}_{nr}=\frac{1/x_{nr}}{\sum_{n=1}^{N}1/x_{nr}}.
$$
The comprehensive contribution coefficient of electric vehicle charging station \(n\) is
$$
\varphi_n=\sum_{r=1}^{4}W_r\hat{x}_{nr},
$$
where \(W_r\) is the weight of indicator \(r\). I determine the weights through the covariance-analytic hierarchy process. The covariance matrix \(A=[a_{rl}]_{4\times4}\) captures the differences among electric vehicle charging stations on each indicator. I convert it into a relative covariance matrix \(B=[b_{rl}]_{4\times4}\) with
$$
b_{rl}=\frac{a_{rl}}{a_{rr}}.
$$
I then construct a positive reciprocal judgment matrix \(C=[c_{rl}]_{4\times4}\) such that \(c_{rl}>0\) and \(c_{rl}c_{lr}=1\). The geometric mean of row \(r\) is
$$
g_r=\left(\prod_{l=1}^{4}c_{rl}\right)^{1/4},
$$
and the normalized weight is
$$
W_r=\frac{g_r}{\sum_{l=1}^{4}g_l}.
$$
I check consistency through
$$
CI=\frac{\lambda_{\max}-4}{3},\qquad CR=\frac{CI}{RI},
$$
where \(RI=0.90\) for four indicators. A value \(CR<0.1\) indicates acceptable consistency. In my case, the resulting weight vector is
$$
\mathbf{W}=\left(0.3125,\ 0.1836,\ 0.3934,\ 0.1105\right)^{T}.
$$
The response rate indicator receives the highest weight, followed by regulation capacity. This is consistent with the market value of fast-response electric vehicle resources. I report the standardized contribution scores and the comprehensive contribution coefficients in Table 6.
| Electric vehicle charging station | Capacity contribution | Mileage contribution | Response rate contribution | Cost contribution | Comprehensive contribution \(\varphi_n\) |
|---|---|---|---|---|---|
| Station 1 | 0.1930 | 0.2441 | 0.2879 | 0.3508 | 0.2572 |
| Station 2 | 0.2861 | 0.2774 | 0.2303 | 0.0891 | 0.2408 |
| Station 3 | 0.2473 | 0.1561 | 0.2658 | 0.2403 | 0.2371 |
| Station 4 | 0.2736 | 0.3223 | 0.2159 | 0.3197 | 0.2650 |
The comprehensive contribution ranking is Station 4, Station 1, Station 2, and Station 3. Station 4 has strong capacity, mileage, and cost contributions, so it should receive a higher share than a purely economic marginal contribution would suggest. Station 3 has no single dominant dimension, so its comprehensive contribution is the lowest. This is exactly the kind of information that the traditional Shapley value misses when it treats each electric vehicle charging station as an identical economic player.
I define the coalition value as
$$
v(L)=(1-\kappa)\left(R(L)-G(L)\right),
$$
where \(R(L)\) is the market revenue of coalition \(L\), \(G(L)\) is the regulation cost, and \(\kappa\) is the aggregator operating cost coefficient. The traditional Shapley value of station \(n\) is
$$
\phi_n(v)=\sum_{L\subseteq N\setminus\{n\}}\frac{|L|!\left(N-|L|-1\right)!}{N!}\left[v(L\cup\{n\})-v(L)\right].
$$
The improved Shapley value is
$$
\phi_n^{*}(v)=\phi_n(v)+\xi v(N)\left(\varphi_n-\frac{1}{N}\right),
$$
where \(\xi\) is a revenue adjustment coefficient. Because
$$
\sum_{n=1}^{N}\left(\varphi_n-\frac{1}{N}\right)=0,
$$
the improved allocation remains collective-rational:
$$
\sum_{n=1}^{N}\phi_n^{*}(v)=v(N).
$$
I also require individual rationality, so that each electric vehicle charging station receives no less than its stand-alone value when the cooperative gain is positive. The correction term \(\xi v(N)(\varphi_n-1/N)\) rewards stations with high multi-dimensional contribution and reduces the share of stations with low contribution. This preserves the stability of the electric vehicle coalition while correcting the blind spot of the traditional Shapley value.
Simulation Results and Economic Interpretation
I test the mechanism on a representative day-ahead and real-time market data set with a two-second regulation signal and a thirty-minute bidding interval. The electric vehicle fleet includes private electric vehicles, ride-hailing electric vehicles, and commuter electric vehicles. I construct four heterogeneous electric vehicle charging stations. The composition is summarized in Table 7. The heterogeneity is intentional: it allows me to test whether the user response model and the improved Shapley value can distinguish electric vehicle stations with different service capabilities.
| Electric vehicle charging station | Private electric vehicle | Ride-hailing electric vehicle | Commuter electric vehicle | Total electric vehicle users |
|---|---|---|---|---|
| Station 1 | 88 | 125 | 0 | 213 |
| Station 2 | 62 | 100 | 169 | 331 |
| Station 3 | 0 | 168 | 165 | 333 |
| Station 4 | 181 | 0 | 0 | 181 |
I first compare the regulation potential with and without user response. Without user response, the aggregator assumes that every plugged-in electric vehicle is available. This is the theoretical upper bound. With user response, the aggregator uses only the electric vehicle users whose comprehensive utility is positive. I observe that the actual regulation potential is significantly smaller than the theoretical upper bound. The reduction is especially large during the evening peak, when many ride-hailing electric vehicles are connected but their response rate is low. In some evening hours, the potential boundary contracts by more than sixty percent. This result confirms that ignoring user response would lead to serious overbidding of electric vehicle frequency regulation capacity.
I also test the sensitivity of the electric vehicle response model to psychological parameters. When the risk preference coefficient increases, electric vehicle users become less averse to gains and losses, their decisions move closer to rational behavior, and the aggregated regulation potential expands. When the utility reference point increases, electric vehicle users require a higher comprehensive utility before they respond, so the response rate decreases and the regulation potential contracts. The direction of these effects is summarized in Table 8. The changes are smooth rather than abrupt, which indicates that the response model is stable enough for market operation.
| Sensitivity parameter | Change | Effect on electric vehicle response | Effect on aggregated regulation potential |
|---|---|---|---|
| Risk preference coefficient | Increase | Weaker loss aversion, more willingness | Expands |
| Risk preference coefficient | Decrease | Stronger loss aversion, less willingness | Contracts |
| Utility reference point | Increase | Higher utility hurdle for response | Contracts |
| Utility reference point | Decrease | Lower utility hurdle for response | Expands |
I then evaluate the economic performance of the co-optimized bidding and allocation strategy. I compare my model with a real-time cost minimization benchmark. The benchmark assumes proportional power allocation during bidding and minimizes only the current regulation cost during real-time allocation. It ignores the future opportunity cost of stored energy. My model embeds power allocation in the bidding problem and introduces the Lagrange multiplier into real-time allocation. The comparison is reported in Table 9. My model increases total revenue by 1.97 percent, reduces cost by 4.54 percent, and increases net revenue by 5.75 percent. The economic gain comes from two sources. First, the bidding model anticipates how power allocation will affect future state of charge. Second, the real-time allocation model avoids using cheap but energy-scarce electric vehicle stations too early.
| Strategy | Total revenue ($) | Cost ($) | Net revenue ($) | Change in net revenue |
|---|---|---|---|---|
| Co-optimized bidding and allocation | 2407.22 | 828.17 | 1579.05 | +5.75% |
| Real-time cost minimization | 2360.79 | 867.57 | 1493.21 | 0.00% |
I also examine computational performance. The bidding model is solved every thirty minutes, and the real-time allocation model is solved whenever a regulation signal arrives. When the number of electric vehicle charging stations is below sixty, the average computation time for one bidding interval is less than eighteen seconds. This is acceptable for rolling market clearing. When the number of electric vehicle charging stations increases further, the computation time grows. For large-scale deployment, I recommend a hierarchical aggregation strategy: group electric vehicle charging stations by geographic area or response characteristics, optimize each subgroup, and allocate power independently inside each subgroup. This decomposes a large electric vehicle optimization problem into several smaller problems and preserves computational tractability.
| Number of electric vehicle charging stations | Average computation time | Operational interpretation |
|---|---|---|
| Up to 60 | Below 18 s | Suitable for rolling bidding and real-time allocation |
| Around 100 | About 30 s | Acceptable with parallel or hierarchical decomposition |
| Large-scale cluster | Depends on decomposition | Use regional subgroup aggregation and parallel solving |
Finally, I analyze the revenue allocation results. The grand coalition net value after the aggregator operating cost is 1263.24 dollars. The sum of stand-alone net revenues of the four electric vehicle charging stations is 1194.57 dollars. Cooperation therefore creates a positive gain of 68.67 dollars, or 5.75 percent. This positive gain satisfies collective rationality and gives each electric vehicle charging station an incentive to join the coalition. I compare independent operation, traditional Shapley value, and improved Shapley value in Table 10.
| Electric vehicle charging station | Independent operation ($) | Traditional Shapley value ($) | Traditional share (%) | Improved Shapley value ($) | Improved share (%) |
|---|---|---|---|---|---|
| Station 1 | 319.29 | 322.15 | 25.50 | 331.19 | 26.22 |
| Station 2 | 281.94 | 297.00 | 23.51 | 285.40 | 22.59 |
| Station 3 | 146.66 | 172.31 | 13.64 | 155.97 | 12.35 |
| Station 4 | 446.68 | 471.78 | 37.35 | 490.69 | 38.84 |
The traditional Shapley value gives Station 3 a relatively large cooperative gain even though its comprehensive contribution is the lowest. This is the well-known weakness of a purely economic marginal contribution rule: a low-contribution electric vehicle charging station can receive a large cooperative gain if its presence changes the coalition value in some permutations. The improved Shapley value corrects this distortion. Station 4, which has the highest comprehensive contribution, increases its share from 37.35 percent to 38.84 percent. Station 3, which has the lowest comprehensive contribution, decreases its share from 13.64 percent to 12.35 percent. Station 1 also receives a higher share because its response rate and cost contributions are strong. The allocation result is therefore better aligned with the actual service capability of each electric vehicle charging station.
| Electric vehicle charging station | Traditional cooperative gain ($) | Improved cooperative gain ($) | Direction of correction |
|---|---|---|---|
| Station 1 | 2.86 | 11.90 | Rewarded for strong response and cost contribution |
| Station 2 | 15.06 | 3.46 | Reduced because its multi-dimensional contribution is moderate |
| Station 3 | 25.65 | 9.31 | Reduced because its comprehensive contribution is the lowest |
| Station 4 | 25.10 | 44.01 | Rewarded for leading capacity, mileage, and cost contribution |
I interpret these results from the perspective of long-term electric vehicle aggregator operation. If a low-contribution electric vehicle charging station receives the largest cooperative gain, high-contribution stations will gradually lose interest in providing high-quality regulation service. The electric vehicle coalition may still exist formally, but its service quality will erode. The improved Shapley value solves this incentive problem by linking revenue to multi-dimensional contribution. It does not abandon the economic marginal contribution logic of the Shapley value. Instead, it adds a contribution correction term that is zero-sum and collective-rational. This is important because the electric vehicle aggregator must simultaneously satisfy market rules, user fairness, and station participation.
Managerial and Market Implications
My results lead to several practical implications for electric vehicle aggregators. First, an electric vehicle aggregator should not treat user response as a fixed participation rate. Response willingness is a dynamic state that depends on incentive price, battery stress, initial state of charge, and parking duration. The aggregator should update the response rate before each bidding interval and adjust the electric vehicle regulation potential accordingly. Second, the aggregator should use a virtual battery boundary rather than a single capacity number. The Minkowski sum provides a power-energy feasible region that supports both bidding and real-time allocation. Third, the aggregator should price future opportunity cost in real-time allocation. A greedy allocation that minimizes current cost can exhaust a high-value electric vehicle station early and cause higher cost later. Fourth, the aggregator should allocate revenue by multi-dimensional contribution. The electric vehicle stations that provide fast response, high mileage, and low-cost regulation should be rewarded, because these attributes determine the quality of frequency regulation service.
From a market design perspective, my mechanism shows that electric vehicle frequency regulation is not only a power capacity problem. It is a user-behavior problem, a virtual battery aggregation problem, and a coalition incentive problem. The same electric vehicle cluster can have very different market value depending on whether user response is modeled, whether power allocation is co-optimized with bidding, and whether revenue is distributed fairly. The electric vehicle aggregator that ignores these dimensions may appear to have a large regulation potential on paper, but it will face delivery risk, deviation penalties, and station exit. The electric vehicle aggregator that integrates these dimensions can convert heterogeneous electric vehicle flexibility into a reliable and sustainable frequency regulation service.
I also note that the proposed mechanism is modular. The user response model can be recalibrated for different electric vehicle user populations. The Minkowski aggregation can be extended to include vehicle-to-grid discharging and charging efficiency. The bidding model can be adapted to market rules that separate upward and downward regulation. The contribution evaluation can include additional dimensions such as carbon reduction or voltage support if the electric vehicle aggregator participates in multiple markets. This modularity is valuable because electric vehicle markets are still evolving, and an aggregator needs a mechanism that can evolve with market rules without losing its physical and economic consistency.
Conclusion
I have presented an aggregated frequency regulation trading mechanism for heterogeneous electric vehicle resources. The mechanism couples user response, Minkowski-sum potential aggregation, stochastic bidding, real-time power allocation, multi-dimensional contribution evaluation, and improved Shapley revenue sharing. The user response model captures the fuzzy perception boundary and bounded rationality of electric vehicle users. The Minkowski sum preserves the power-energy coupling of individual electric vehicles and provides a station-level virtual battery region. The co-optimization model embeds power allocation in bidding and uses a Lagrange multiplier to represent future opportunity cost in real-time allocation. The improved Shapley value rewards electric vehicle charging stations according to capacity, mileage, response rate, and cost contribution, thereby aligning revenue with actual service capability.
My simulation results show that ignoring user response can overestimate electric vehicle regulation potential by more than sixty percent in some periods. The co-optimized strategy increases aggregator net revenue by 5.75 percent compared with a real-time cost minimization benchmark. The improved Shapley value corrects the over-reward of low-contribution electric vehicle charging stations and increases the share of high-contribution stations. These results support a simple conclusion: the sustainable operation of electric vehicle aggregated frequency regulation requires a mechanism that respects user heterogeneity, physical feasibility, market co-optimization, and fair coalition revenue allocation. I believe this mechanism provides a practical foundation for electric vehicle aggregators to participate in frequency regulation markets while maintaining user trust and station cooperation.
