Cooperative operation control method for electric vehicle rapid charging station equipped with battery energy storage
By combining distributed offline training and explicit policy functions, the problem of coordinated scheduling between electric vehicle fast charging stations and battery energy storage is solved, achieving efficient and scalable online decision-making and load smoothing, and overcoming the communication and privacy issues of centralized scheduling.
Patent Information
- Application Number
- CN202511246877.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-02
- Publication Date
- 2025-12-09
AI Technical Summary
Existing technologies struggle to effectively coordinate the scheduling of electric vehicle fast charging stations and battery energy storage, resulting in excessive communication and computing burdens, poor privacy, and weak scalability. This is particularly true in environments with multiple fast charging stations, where uncertainties and multi-stage uncertainties cannot be handled in real-time scheduling.
A distributed offline training method is adopted, which combines local value function training and back learning with explicit policy function and consensus algorithm to realize the online collaborative operation of fast charging station and battery energy storage. Dynamic programming is performed by using Bellman optimality principle and dual decomposition theory to reduce the dependence on centralized prediction and realize distributed decision-making.
It enables efficient collaborative operation of fast charging stations and battery energy storage without relying on predictive information, improves computing efficiency and scalability, ensures data privacy, reduces communication burden, and significantly improves the efficiency of online decision-making and load smoothing effect.
Smart Images
Figure CN121097784A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of electrical system operation optimization and control technology, specifically relating to a real-time control method for online collaborative operation of electric vehicle fast charging stations equipped with battery energy storage. Background Technology
[0002] The rise of electric vehicles (EVs) marks a key shift towards sustainable transportation, offering a viable solution to mitigating environmental problems and addressing climate change. The development of fast-charging technology for EVs is crucial for alleviating range anxiety and facilitating long-distance travel. Currently, with the advancement of ultra-fast charging technology, fast-charging power has reached over 300 kW. However, the high instantaneous power and intermittent nature of EV fast-charging loads can pose a significant threat to the stability and reliability of the power grid, leading to increased peak loads and transformer overloads. To address this issue, fast-charging stations are typically equipped with battery energy storage to smooth out fluctuating load power and reduce operating costs.
[0003] The stable and economical operation of the power grid requires the full coordination and utilization of fast charging stations and battery energy storage; however, achieving efficient online coordinated scheduling faces numerous challenges. On the one hand, the fast charging load of electric vehicles exhibits significant uncertainty. On the other hand, multiple fast charging stations with different operating characteristics interact with each other, further increasing the complexity of real-time scheduling decisions. Some researchers have proposed methods, including stochastic programming and distributed bar optimization, to handle multi-stage uncertainties. However, these studies primarily focus on determining day-ahead strategies, and their performance cannot be guaranteed in real-time scheduling. Other studies have extended these methods to online methods by combining rolling optimization, but rolling optimization requires high sensitivity to prediction accuracy.
[0004] In addition, most of the methods mentioned above are executed in a centralized manner, requiring a central controller to access information across the entire system and handle large-scale computations. This centralized approach suffers from problems such as excessive communication and computational burden, poor privacy, weak scalability, and the potential for single points of failure. The problem scale increases dramatically with the number of fast-charging stations. Some studies have proposed various distributed algorithms, such as the alternating direction multiplier method; however, these studies only focus on distributed optimization for offline deterministic problems and do not consider online sequential decision-making for uncertain problems. Summary of the Invention
[0005] The purpose of this invention is to overcome at least one deficiency of the prior art and provide a real-time control method for online collaborative operation of electric vehicle fast charging stations equipped with battery energy storage.
[0006] The technical solution adopted in this invention is:
[0007] The first aspect of the present invention provides:
[0008] A real-time solution method for online collaborative operation of fast charging stations for electric vehicles equipped with battery storage includes the following steps:
[0009] Distributed offline training: Starting from the end time period, each fast charging station and battery energy storage independently trains its local value function based on historical data, and passes the value function to the previous time period for reverse learning;
[0010] Online scheduling and control: Starting from the initial period, based on the trained value function, each fast charging station and battery energy storage autonomously decides the optimal operation action based solely on the local observation information of the current period (the fast charging demand of electric vehicles at the station and the current electricity purchase settlement rate), without needing to make any predictions for the exogenous components of subsequent periods, and then updates its own state to carry out the decision-making process for the next period.
[0011] Its basic mathematical model is as follows:
[0012]
[0013] in, This refers to the unit price of electricity purchased. The charging demand of electric vehicles at fast charging station n at time t; r n,t e represents the remaining charging demand of electric vehicles at fast charging station n at time t; n,t Let n be the energy level of the battery at time t. For tie line power; The difference between the electric vehicle's charging demand at time t and the actual charging power. and These represent the charging and discharging power of the battery's energy storage n at time t, respectively; ρ lf It is the penalty coefficient for load fluctuations; The marginal discharge cost of battery energy storage n; It is a predefined penalty factor for the remaining charging demand of fast charging station n;
[0014] The optimization objectives consist of four parts: energy cost, penalty cost for load fluctuation, battery energy storage degradation cost, and penalty cost for remaining fast charging demand of electric vehicles. At the same time, each fast charging station and battery energy storage needs to meet its own operational constraints and global power balance constraints.
[0015] In some instances, a spatiotemporal decoupling algorithm is used for real-time solutions. In the time dimension, based on the Bellman optimality principle, the original multi-time-period optimization problem is decomposed into T nested single-time-period dynamic programming problems:
[0016]
[0017] Among them, V t(·) represents the value function, indicating the opportunity cost of making decisions in the future time period. The value function after the decision;
[0018] In the spatial dimension, based on the dual decomposition theory, the dynamic programming problem is further decomposed into multiple single-agent dynamic programming problems, including the tie-line subproblem, the fast charging station subproblem, and the battery energy storage subproblem.
[0019]
[0020]
[0021] Where λ t As the dual variable of the global power balance constraint, it is considered as coordination information and incorporated into the state space of each agent. and These are the value functions / post-decision value functions for the fast charging station intelligent agent and the battery energy storage intelligent agent, respectively.
[0022] In some instances, λ t Discretize into K equal intervals, and define the lower bound, midpoint, and upper bound of each interval using... l t,k , λ t,k and This means that, for any given discrete interval, the fast charging station subproblem and the battery energy storage subproblem can be decomposed into:
[0023]
[0024] in Indicates from λ t,k To λ t+Δt,k′ The state transition probabilities are estimated from historical data.
[0025] In some instances, for the sub-problems after spatiotemporal decoupling, a distributed offline learning method is used, where each agent independently learns experiential knowledge from historical scenarios and embeds it into the value function V of the dynamic programming problem. t In (·), an analytical recursive formula is used for the offline training process to obtain the value function of each agent in a non-iterative and non-solution-based manner. For the fast charging station agent:
[0026]
[0027]
[0028] For battery energy storage intelligent agents:
[0029]
[0030] Where v is the derivative of V, i.e., the marginal value function; v post It is V post The derivative of the decision, i.e. the marginal value function after decision, is given in advance for all fast charging stations and battery energy storage agents. The marginal value function at the end of the cycle is given in advance. Then, the above analytical recursive formula is executed alternately in reverse order from t=T to t=1 to obtain the marginal value function and the marginal value function after decision for each cycle (t=1,…,T) covering all states.
[0031] In some instances, after the distributed offline training process is completed, a well-trained post-decision marginal value function embedded with empirical knowledge is used to assist each agent in making real-time decisions. Each agent utilizes a multi-threshold structured explicit policy function to incorporate the external information λ observed at each time interval. t The optimal action is directly mapped to the policy function.
[0032] The explicit policy function of the contactless agent is
[0033] The explicit policy function of the fast charging station agent is:
[0034]
[0035]
[0036] Explicit policy function for battery storage agent:
[0037]
[0038] In some instances, after obtaining the optimal policy for each agent, a consensus algorithm is used to achieve online collaborative operation of fast charging stations and battery energy storage within a distributed framework. This is achieved by first introducing a local consensus variable λ. n.t As a local estimate of global information, and λ n.t gradient ζ n,t Agents can share their local consensus variables with neighboring agents on a ring communication network. The distributed online scheduling steps based on the consensus algorithm are as follows:
[0039] S1: Initialization. In each time period t, each agent first initializes its estimated consensus variable to 0 and sets the iteration index h to 0.
[0040] S2: Information exchange. In each iteration, each agent communicates with its neighboring agents to obtain their latest estimates of the consensus variables.
[0041] S3: Dual variable update: Each agent updates the dual variable locally based on the received information;
[0042]
[0043] S4: Subproblem solving, each agent solves based on the updated λ. n It obtains its optimal response in parallel with the learned multi-threshold structure explicit policy function;
[0044] S5: Gradient update, each agent updates its local gradient based on the gradient estimates and correction terms of its neighboring agents. The correction term is given by the difference between the best responses of the two most recent iterations.
[0045]
[0046] S6: The termination condition is set as follows:
[0047]
[0048] Where σ is the preset precision, the iteration process terminates when all agents simultaneously meet the termination condition; then the real-time scheduling decision of each agent in the current time period is output, and the corresponding state variables are updated to prepare for the optimization of the next time period; otherwise, set h = h + 1, return to S2, and continue the above iteration process.
[0049] In some instances, the historical data includes electric vehicle charging demand and electricity prices.
[0050] In some instances, the self-state includes remaining charging demand and battery energy storage state of charge.
[0051] A second aspect of the present invention provides:
[0052] A computer-readable storage medium storing one or more programs, said one or more programs including instructions that, when executed by a computing device, cause the computing device to perform the method described in the first aspect of the present invention.
[0053] A third aspect of the present invention provides:
[0054] A computing device, comprising:
[0055] One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the one or more programs including steps for the method described in the first aspect of the invention.
[0056] The beneficial effects of this invention are:
[0057] This invention proposes a real-time solution method for online collaborative operation of electric vehicle fast charging stations equipped with battery storage. This method enables distributed coordination of fast charging stations and battery storage resources without relying on predictive information. The offline training process of the proposed dynamic programming algorithm is analytical and interpretable, significantly improving offline training efficiency. Furthermore, the explicit policy function proposed in this invention allows each agent to directly obtain online decisions without using commercial solvers, thereby greatly improving online computational efficiency. The distributed solution of this invention effectively addresses communication and privacy issues because private information (electric vehicle fast charging requirements) is stored only by local agents and does not need to be shared, eliminating the need for a central controller. Moreover, since all computations are performed on local agents, computational efficiency and scalability are significantly improved. Attached Figure Description
[0058] Figure 1 This is a flowchart of the method designed in this invention.
[0059] Figure 2 This is a schematic diagram of the fluctuation curve of net load power within a day.
[0060] Figure 3 This is a bar chart showing the average daily operating cost and delayed charging demand over a month. Detailed Implementation
[0061] The following description, in conjunction with the embodiments and accompanying drawings, provides further details.
[0062] Reference Figure 1 This invention establishes an optimization model for the online collaborative operation of electric vehicle fast charging stations and battery energy storage resources. The invention designs a real-time solution method for online collaborative operation combining dynamic programming and consensus algorithms. The method includes two stages: a distributed offline training stage, where, starting from the end time period, each fast charging station and battery energy storage independently trains its local value function based on historical data (electric vehicle charging demand, electricity price), and then passes the value function to the previous time period for back-learning; and an online scheduling stage, where, starting from the start time period, based on the trained value function, each fast charging station and battery energy storage decides its optimal operating action according to current local observation information, without needing to predict future charging load or other information, and then updates its own state (remaining charging demand, battery capacity, etc.) to proceed with the decision-making process for the next time period. Both the offline training and online decision-making processes are based on the proposed analytical explicit expressions without complex computation processes. Furthermore, both offline training and online decision-making processes are executed in a distributed manner, ensuring the data privacy of each fast charging station and the scalability of the algorithm.
[0063] The mathematical model for the online collaborative optimization problem of fast charging stations and battery energy storage resources in the power distribution network of this invention can be expressed as follows:
[0064]
[0065]
[0066] in, This refers to the unit price of electricity purchased. The charging demand of electric vehicles at fast charging station n at time t; r n,t e represents the remaining charging demand of electric vehicles at fast charging station n at time t; n,t Let n be the energy level of the battery at time t. For tie line power; The difference between the electric vehicle's charging demand at time t and the actual charging power. and These represent the charging and discharging power of the battery's energy storage n at time t, respectively; ρ lf It is the penalty coefficient for load fluctuations; The marginal discharge cost of battery energy storage n; It is a predefined penalty factor for the remaining charging demand of fast charging station n;
[0067] The optimization objectives consist of four parts: energy cost, penalty cost for load fluctuation, battery energy storage degradation cost, and penalty cost for remaining fast charging demand of electric vehicles. At the same time, each fast charging station and battery energy storage needs to meet its own operational constraints and global power balance constraints.
[0068] This invention first decouples the aforementioned optimization problem in the time dimension. Based on the Bellman optimality principle, the original multi-time-period optimization problem is decomposed into T nested single-time-period dynamic programming problems:
[0069]
[0070]
[0071] Among them, V t (·) represents the value function, indicating the opportunity cost of making decisions in the future time period. This invention introduces a post-decision value function, which allows the impact of future random information to be considered within the current time interval.
[0072] In the spatial dimension, this invention, based on dual decomposition theory, further decomposes the aforementioned dynamic programming problem involving multiple agents into multiple single-agent problems, including the tie-line sub-problem, the fast charging station sub-problem, and the battery energy storage sub-problem.
[0073]
[0074] Where λt As the dual variable of the global power balance constraint, it is considered as coordination information and incorporated into the state space of each agent. and These are the value functions / post-decision value functions for the fast charging station intelligent agent and the battery energy storage intelligent agent, respectively.
[0075] Furthermore, the present invention will λ t Discretize into K equal intervals, and define the lower bound, midpoint, and upper bound of each interval using... l t,k , λ t,k and This means that, for any given discrete interval, the fast charging station subproblem and the battery energy storage subproblem can be further decomposed into:
[0076]
[0077]
[0078] in Indicates from λ t,k To λ t+Δt,k′ The state transition probabilities are estimated from historical data. This allows for decoupling of the original problem in terms of time, space, and state intervals, significantly reducing the computational scale.
[0079] To address the sub-problems resulting from spatiotemporal decoupling, this invention proposes a distributed offline learning method. Each agent independently learns experiential knowledge from historical scenarios and embeds it into the value function V of the dynamic programming problem. t (·) In the offline training process, the value function is obtained in a non-iterative and non-solvable manner based on the proposed analytical recursive formula. For the fast charging station agent:
[0080]
[0081] For battery energy storage intelligent agents:
[0082]
[0083] Where v is the derivative of V, i.e., the marginal value function; v post It is V post The derivative of the decision, i.e. the marginal value function after decision, is given in advance for all fast charging stations and battery energy storage agents. The marginal value function at the end of the cycle is given in advance. Then, the above analytical recursive formula is executed alternately in reverse order from t=T to t=1 to obtain the marginal value function and the marginal value function after decision for each cycle (t=1,…,T) covering all states.
[0084] After the distributed offline training process is completed, this invention embeds a well-trained post-decision marginal value function based on empirical knowledge to assist each agent in making real-time decisions. This invention derives an explicit policy function for each agent, thereby obtaining the optimal action based on the current online state.
[0085] Explicit policy function of the contactless agent:
[0086]
[0087] The explicit policy function of the fast charging station agent is:
[0088]
[0089] Explicit policy function for battery storage agent:
[0090]
[0091] After obtaining the optimal strategy for each agent, this invention combines the proposed dynamic programming algorithm with a consensus algorithm to achieve online collaborative operation of fast charging stations and battery energy storage under a distributed framework. First, a local consensus variable λ is introduced. n.t As a local estimate of global information, and λ n.t gradient ζ n,t Agents can share their local consensus variables with neighboring agents on a ring communication network. The distributed online scheduling steps based on the consensus algorithm are as follows:
[0092] S1: Initialization. In each time period t, each agent first initializes its estimated consensus variable to 0 and sets the iteration index h to 0.
[0093] S2: Information exchange. In each iteration, each agent communicates with its neighboring agents to obtain their latest estimates of the consensus variables.
[0094] S3: Dual variable update: Each agent updates the dual variable locally based on the received information;
[0095]
[0096] S4: Subproblem solving, each agent solves based on the updated λ. n It obtains its optimal response in parallel with the learned multi-threshold structure explicit policy function;
[0097] S5: Gradient update, each agent updates its local gradient based on the gradient estimates and correction terms of its neighboring agents. The correction term is given by the difference between the best responses of the two most recent iterations.
[0098]
[0099] S6: The termination condition is set as follows:
[0100]
[0101] Where σ is the preset precision, the iteration process terminates when all agents simultaneously meet the termination condition; then the real-time scheduling decision of each agent in the current time period is output, and the corresponding state variables are updated to prepare for the optimization of the next time period; otherwise, set h = h + 1, return to S2, and continue the above iteration process.
[0102] The following are embodiments of the complete method implemented according to the content of this invention:
[0103] The effectiveness of this invention was verified on a power distribution system comprising three electric vehicle fast charging stations equipped with battery energy storage. Each fast charging station is equipped with six charging piles, each with a rated power of 350kW; and an upper limit is set for the extended charging power of each station for a single time period. With the upper bound of residual demand R n The rated power and capacity of the battery storage at each station are {100 / 3, 150 / 3, 250 / 3} kW and {100 / 12, 150 / 12, 250 / 12} kWh, respectively; the rated power and capacity of the battery storage at each station are {100, 150, 250} kW and {200, 300, 500} kWh, respectively; the charge / discharge efficiency η = 0.95; the marginal cost of degradation is 10 CNY / MWh; the penalty coefficients for load fluctuation and residual demand are set as ρ, respectively. lf = 1 / 15000 CNY / kW 2 and Using one year of fast-charging load data based on EVI-EnSite: the first 11 months were used for offline training, and the last month for validation; the electricity price adopted time-of-use pricing: 0.3818 CNY / kWh from 23:00 to 07:00, 1.3222 CNY / kWh from 10:00 to 15:00 and 18:00 to 21:00, and 0.8395 CNY / kWh for other times. The number of discrete intervals for the dual-price signal K = 23; the consistency convergence tolerance σ = e⁻²; the communication topology is a ring network, and the optimized scheduling time resolution is set to 5 minutes.
[0104] To verify the effectiveness of the proposed distributed online solution algorithm, this invention compares it with an ideal case that assumes centralized deterministic optimization is performed with accurate and complete future prediction information, which can serve as the best benchmark for comparison.
[0105] In this example, the effectiveness of the algorithm proposed in this invention is evaluated by four indicators: load fluctuation, operating cost, user charging satisfaction, and computation time, compared with the ideal situation.
[0106] In this example, the variance ratio is defined as the ratio of the load variance before and after optimization, used to evaluate load smoothing performance. A smaller variance ratio indicates better load smoothing. Figure 2 It is evident that the original fast-charging load of electric vehicles is highly volatile. By optimizing the fast-charging and battery energy storage charging and discharging behavior, the volatile load curve can be significantly alleviated, and the algorithm proposed in this invention produces a profile similar to the ideal situation.
[0107] In this example, cost reduction is mainly achieved by utilizing time-of-use electricity price arbitrage. Throughout the simulation month, the original system's operating cost was 5278.15 yuan. The average daily operating costs under ideal conditions and the method of this invention were 4538.77 yuan and 4540.67 yuan, respectively, demonstrating the economic advantage of the method of this invention. For the average daily delayed charging demand, the ideal condition and the method of this invention were 65.07 kWh and 78.76 kWh, respectively. Figure 3 ).
[0108] In terms of computational efficiency, for an optimization cycle of 1 month (31 days), with a time resolution of 5 minutes and a total of 8928 sequential decision processes, this invention can complete the solution within 5 minutes. In contrast, the ideal solution time for centralized optimization using a commercial solver is more than 24 hours. This invention greatly improves online computational efficiency.
[0109] The above is a further detailed description of the present invention and should not be considered as a limitation on the specific implementation of the present invention. For those skilled in the art, simple deductions or substitutions without departing from the concept of the present invention are all within the protection scope of the present invention.
Claims
1. A real-time solution method for online collaborative operation of fast charging stations for electric vehicles equipped with battery energy storage, comprising the following steps: Distributed offline training: Starting from the end time period, each fast charging station and battery energy storage independently trains its local value function based on historical data, and passes the value function to the previous time period for reverse learning; Online scheduling and control: Starting from the initial period, based on the trained value function, each fast charging station and battery energy storage autonomously decides the optimal operating action based solely on the local observation information of the current period, without needing to make any predictions for the exogenous components of subsequent periods, and then updates its own state to carry out the decision-making process for the next period. Its basic mathematical model is as follows: in, This refers to the unit price of electricity purchased. The charging demand of electric vehicles at fast charging station n at time t; r n,t e represents the remaining charging demand of electric vehicles at fast charging station n at time t; n,t Let n be the energy level of the battery at time t. For tie line power; The difference between the electric vehicle charging demand and the actual charging power at time t; and These represent the charging and discharging power of the battery's energy storage n at time t, respectively; ρ lf It is the penalty coefficient for load fluctuations; The marginal discharge cost of battery energy storage n; It is a predefined penalty factor for the remaining charging demand of fast charging station n; The optimization objectives consist of four parts: energy cost, penalty cost for load fluctuation, battery energy storage degradation cost, and penalty cost for remaining fast charging demand of electric vehicles. At the same time, each fast charging station and battery energy storage needs to meet its own operational constraints and global power balance constraints.
2. The real-time solution method for online collaborative operation of electric vehicle fast charging stations according to claim 1, characterized in that, A real-time solution is achieved using a spatiotemporal decoupling algorithm. In the time dimension, based on the Bellman optimality principle, the original multi-time-period optimization problem is decomposed into T nested single-time-period dynamic programming problems: Among them, V t (·) is the value function, representing the opportunity cost of making decisions in the future time period, V t post (·) represents the post-decision value function; In the spatial dimension, based on the dual decomposition theory, the dynamic programming problem is further decomposed into multiple single-agent dynamic programming problems, including the tie-line subproblem, the fast charging station subproblem, and the battery energy storage subproblem. Where λ t As the dual variable of the global power balance constraint, it is considered as coordination information and incorporated into the state space of each agent. and These are the value functions / post-decision value functions for the fast charging station intelligent agent and the battery energy storage intelligent agent, respectively.
3. The real-time solution method for online collaborative operation of electric vehicle fast charging stations according to claim 2, characterized in that, λ t Discretize into K equal intervals, and denote the lower bound, midpoint, and upper bound of each interval as λ. t,k , λ t,k and This means that, for any given discrete interval, the fast charging station subproblem and the battery energy storage subproblem can be decomposed into: in Indicates from λ t,k to λ t+Δt,k′ The state transition probabilities are estimated from historical data.
4. The real-time solution method for online collaborative operation of electric vehicle fast charging stations according to claim 3, characterized in that, For the subproblems resulting from spatiotemporal decoupling, a distributed offline learning method is used, where each agent independently learns experiential knowledge from historical scenarios and embeds it into the value function V of the dynamic programming problem. t In (·), an analytical recursive formula is used for the offline training process to obtain the value function of each agent in a non-iterative and non-solution-based manner. For the fast charging station agent: For battery energy storage intelligent agents: Where v is the derivative of V, i.e., the marginal value function; v post It is V post The derivative of the decision, i.e. the marginal value function after decision, is given in advance for all fast charging stations and battery energy storage agents. The marginal value function at the end of the cycle is given in advance. Then, the above analytical recursive formula is executed alternately in reverse order from t=T to t=1 to obtain the marginal value function and the marginal value function after decision for each cycle (t=1,…,T) covering all states.
5. The real-time solution method for online collaborative operation of electric vehicle fast charging stations according to claim 4, characterized in that, After the distributed offline training process is completed, a well-trained decision-making marginal value function embedded with empirical knowledge is used to assist each agent in making real-time decisions. Each agent utilizes a multi-threshold structure explicit policy function to incorporate the external information λ observed at each time period. t The optimal action is directly mapped to the policy function. The explicit policy function of the contactless agent is The explicit policy function of the fast charging station agent is: Explicit policy function for battery storage agent:
6. The real-time solution method for online collaborative operation of electric vehicle fast charging stations according to claim 5, characterized in that, After obtaining the optimal policy for each agent, a consensus algorithm is used to achieve online collaborative operation of fast charging stations and battery energy storage in a distributed framework. First, a local consensus variable λ is introduced. n.t As a local estimate of global information, and λ n.t gradient ζ n,t Agents can share their local consensus variables with neighboring agents on a ring communication network. The distributed online scheduling steps based on the consensus algorithm are as follows: S1: Initialization. In each time period t, each agent first initializes its estimated consensus variable to 0 and sets the iteration index h to 0. S2: Information exchange. In each iteration, each agent communicates with its neighboring agents to obtain their latest estimates of the consensus variables. S3: Dual variable update: Each agent updates the dual variable locally based on the received information; S4: Subproblem solving, each agent solves based on the updated λ. n It obtains its optimal response in parallel with the learned multi-threshold structure explicit policy function; S5: Gradient update, each agent updates its local gradient based on the gradient estimates and correction terms of its neighboring agents. The correction term is given by the difference between the best responses of the two most recent iterations. S6: The termination condition is set as follows: Where σ is the preset precision, the iteration process terminates when all agents simultaneously meet the termination condition; then the real-time scheduling decision of each agent in the current time period is output, and the corresponding state variables are updated to prepare for the optimization of the next time period; Otherwise, set h = h + 1, return to S2, and continue the above iteration process.
7. The real-time solution method for online collaborative operation of electric vehicle fast charging stations according to claim 1, characterized in that, The historical data includes electric vehicle charging demand and electricity prices.
8. The real-time solution method for online collaborative operation of electric vehicle fast charging stations according to claim 1, characterized in that, The self-state includes remaining charging demand and battery energy storage state of charge.
9. A computer-readable storage medium for storing one or more programs, characterized in that, The one or more programs include instructions that, when executed by a computing device, cause the computing device to perform the method of any one of claims 1 to 8.
10. A computing device, characterized in that, include: One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the one or more programs including steps for performing the method of any one of claims 1 to 8.