A battery swap station optimization scheduling method and system

Abstract

The application discloses a battery swap station optimization scheduling method and system, relates to the technical field of electric vehicle battery swap station operation optimization, and clearly defines the decision relationship between a power distribution network operator and a battery swap station operator through a two-stage stochastic programming model, provides a theoretical framework for the coordination and interaction between the two, can comprehensively consider the mixed queuing network service characteristics, demand uncertainty and service quality constraints, and realizes a computable robust optimal scheduling through a physical information neural network model, so that an effective battery swap station scheduling method which can effectively cope with demand uncertainty and collaboratively optimize economic benefits and service quality is provided.

Description

Technical Field

[0001] This invention relates to the field of electric vehicle battery swapping station operation optimization technology, specifically to a battery swapping station optimization scheduling method. Background Technology

[0002] With the rapid development of electric vehicles (EVs), battery swapping, as an efficient energy replenishment method, can provide flexible battery replacement services by decoupling the charging process from vehicle dwell time. Battery swapping stations (BSS) offer a rapid solution for EVs to replenish energy within minutes by separating vehicle and battery ownership and replacing charging with battery swapping.

[0003] Battery swapping stations, with their centralized management and charging of a large number of batteries, become flexible and controllable distributed energy storage units. They possess significant potential to participate in ancillary services markets such as grid peak shaving, frequency regulation, and backup power, bringing additional revenue to operators. In the actual operation of electric vehicle battery swapping stations, queuing issues mainly manifest in two aspects: first, the waiting time for electric vehicles after arrival and the battery swapping service process; and second, the cyclical process of batteries within the swapping station between "waiting to charge—charging—fully charged and ready for use." With the continuous growth in the number of electric vehicles, battery swapping stations face the dual pressures of limited service resources and significant demand fluctuations, leading to problems such as longer user waiting times, decreased service quality, and increased operating costs.

[0004] To address the aforementioned issues, traditional battery swapping station scheduling models often employ deterministic or simplified stochastic models. These methods track the state of charge (SoC) of each battery over time, calculating the charging and discharging processes to determine when a battery is fully charged or depleted. While these methods can accurately reflect the battery's operational trajectory at a microscopic level, they fail to accurately characterize the dynamic characteristics of vehicle arrival and internal battery cycling, and they do not fully consider the coupling issue between demand uncertainty and service quality requirements, leading to the following shortcomings:

[0005] 1) Lack of system-level queuing behavior modeling: It fails to explicitly characterize the random operational characteristics such as vehicle arrival, waiting, and service refusal, and cannot assess the overall service quality;

[0006] 2) Lack of user experience metrics: Key performance metrics such as average waiting time and demand fulfillment rate cannot be directly derived from a single-cell SoC. Summary of the Invention

[0007] The main objective of this invention is to propose an optimized scheduling method and system for battery swapping stations, aiming to solve the technical problem that existing battery swapping station scheduling models cannot accurately depict the dynamic characteristics of vehicle arrival and internal battery cycling, nor fully consider the coupling between demand uncertainty and service quality requirements.

[0008] To achieve the above objectives, this invention proposes an optimized scheduling method for battery swapping stations, comprising the following steps: S100, constructing a service model for battery swapping stations, establishing an operation model for battery swapping stations using a hybrid queuing network model, and calculating service quality indicators through the hybrid queuing network model; wherein, the service quality indicators include at least the average waiting time and the amount of unmet battery swapping demand, and the hybrid queuing network model is described by electric vehicle arrival rate, charging service rate, discharging service rate, and battery swapping service rate; S200, establishing a two-stage stochastic programming model, establishing a two-stage stochastic programming model using distribution network scheduling cost, battery swapping station operation cost, and service quality indicators as parameters; S300, establishing a robust optimization model, constructing a sub-Bruker fuzzy set based on the historical moment information of electric vehicle arrival rate, and transforming the two-stage stochastic programming model into a two-stage sub-Bruker optimization model; S400, solving the model, using a trained physical information neural network model to approximate the prediction of service quality indicators, and using the approximate prediction results as constraints to solve the two-stage sub-Bruker optimization model to output day-ahead hourly power dispatch instructions.

[0009] Preferably, in step S100, the hybrid queuing network model is composed of an open queuing network model and a closed queuing network model coupled together;

[0010] Open queuing network models include those that use Poisson processes to describe the arrival process of electric vehicles;

[0011] The closed queuing network model includes using an exponential distribution to describe charging service time and discharging service time, and using a constant value to describe battery swapping service time.

[0012] Preferably, in step S100, the average waiting time is calculated using Little's law, which is the product of the probability that the unmet battery swapping demand passes through the system at full load and the vehicle entry rate.

[0013] Preferably, step S200 specifically includes the following steps:

[0014] S201. Establish the objective function for power dispatch planning of the distribution network in the first stage;

[0015] S202. Establish line power flow constraints for the power system;

[0016] S203. Establish the objective function for the second phase of the battery swapping station participation in optimized scheduling and backup services;

[0017] S204. Based on the hybrid queuing network model, establish power regulation constraints and energy conversion constraints for battery swapping stations;

[0018] S205. Establish service quality constraints for battery swapping station operation based on service quality indicators.

[0019] Preferably, in step S300, the historical moment information includes the mean vector and covariance matrix of the historical data of electric vehicle arrival rate; the support set of the split-bar fuzzy set is the vector set of electric vehicle arrival rate.

[0020] Preferably, in step S300, the transformation process of the two-stage sub-Brubar optimization model includes: based on the sub-Brubar fuzzy set, transforming the expectation function for the continuous probability distribution in the two-stage stochastic programming model into the expectation function for finding the worst probability distribution within the sub-Brubar fuzzy set.

[0021] Preferably, in step S400, the training process of the physical information neural network model includes: training using a large number of data samples generated offline by the hybrid queuing network model; and introducing a constraint term reflecting the inherent physical relationship of service quality indicators into the loss function for training the physical information neural network model.

[0022] Preferably, the physical information neural network model uses a linear activation function, so that the trained physical information neural network model can be equivalently represented as several linear equality constraints, so as to be embedded in a two-stage sub-Bruker optimization model.

[0023] Preferably, in step S400, a nested column and constraint generation algorithm is used to solve the two-stage sub-Bruker optimization model; the column and constraint generation algorithm determines the worst-case probability distribution and updates the operational decision by alternately solving the main problem and sub-problems until convergence.

[0024] Preferably, in solving the subproblems of the column and constraint generation algorithm, the physical information neural network model is invoked to quickly obtain approximate values ​​of the service quality index corresponding to the current decision variable, replacing the direct solution of the hybrid queuing network model.

[0025] A battery swapping station optimized scheduling system includes:

[0026] The first model construction module uses a hybrid queuing network model to establish a battery swapping station operation model and calculates service quality indicators through the hybrid queuing network model. Among them, the service quality indicators include at least the average waiting time and the amount of unmet battery swapping demand. The hybrid queuing network model is described by electric vehicle arrival rate, charging service rate, discharging service rate and battery swapping service rate.

[0027] The second model building module establishes a two-stage stochastic programming model using distribution network scheduling costs, battery swapping station operating costs, and service quality indicators as parameters.

[0028] The optimization module constructs a sub-Brussels bar fuzzy set based on the historical moment information of electric vehicle arrival rate, and transforms the two-stage stochastic programming model into a two-stage sub-Brussels bar optimization model.

[0029] The calculation module uses a trained physical information neural network model to make approximate predictions of service quality indicators, and uses the results of the approximate predictions as constraints to solve a two-stage sub-Bruker optimization model to output day-ahead hourly power dispatch instructions.

[0030] The beneficial effects of this invention are as follows:

[0031] 1) This invention utilizes a hybrid queuing network model to establish a battery swapping station operation model. The hybrid queuing network model can accurately model the complex queuing behavior of electric vehicles and batteries in the actual operation of the battery swapping station through electric vehicle arrival rate, charging service rate, discharging service rate and battery swapping service rate. It can accurately depict the dynamic interaction between vehicle flow and battery flow at the overall level and reflect queuing and resource sharing characteristics.

[0032] 2) Simultaneously, this invention utilizes historical moment information of electric vehicle arrival rates to construct a sparse Brussels bar fuzzy set, thereby transforming the two-stage stochastic programming model into a two-stage sparse Brussels bar optimization model. This two-stage sparse Brussels bar optimization model can tightly integrate service quality indicators with scheduling optimization strategies under conditions of uncertain distribution or incomplete data, finding a more suitable optimal scheduling design for battery swapping stations under uncertain demand. Furthermore, this invention utilizes a pre-trained physical information neural network model to approximate the prediction of service quality indicators, solving the computational bottleneck of complex stochastic optimization models and greatly balancing model accuracy and computational efficiency.

[0033] Compared with existing technologies, the two-stage stochastic programming model of this invention clearly defines the decision-making relationship between distribution network operators and battery swapping station operators, providing a theoretical framework for their coordinated interaction. It can achieve computable robust optimal scheduling by comprehensively considering the service characteristics of hybrid queuing networks, demand uncertainty, and service quality constraints, and by using a physical information neural network model. Thus, it provides a battery swapping station scheduling method that can effectively cope with demand uncertainty and collaboratively optimize economic benefits and service quality. Attached Figure Description

[0034] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0035] Figure 1 This is a flowchart of Embodiment 1 of the present invention;

[0036] Figure 2 This is a schematic diagram of the framework for two-stage sub-Bruker optimization in this invention.

[0037] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0038] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0039] It should be noted that if the embodiments of the present invention involve directional indicators (such as up, down, left, right, front, back, etc.), the directional indicators are only used to explain the relative positional relationship and movement of the components in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indicators will also change accordingly.

[0040] Example 1

[0041] This embodiment provides a method for optimizing the scheduling of battery swapping stations.

[0042] In this embodiment, as Figure 1 and Figure 2 The method for optimizing the scheduling of battery swapping stations, as shown, includes the following steps:

[0043] S100, Construction of the service model for battery swapping stations: A hybrid queuing network model is used to establish an operation model for battery swapping stations, and service quality indicators are calculated through the hybrid queuing network model. Among them, the service quality indicators include at least the average waiting time and the amount of unmet battery swapping demand. The hybrid queuing network model is described by electric vehicle arrival rate, charging service rate, discharging service rate and battery swapping service rate.

[0044] S200. Establish a two-stage stochastic programming model, using distribution network dispatch cost, battery swapping station operating cost and service quality index as parameters.

[0045] S300. Establish a robust optimization model, construct a sub-Blu-bar fuzzy set based on the historical moment information of electric vehicle arrival rate, and transform the two-stage stochastic programming model into a two-stage sub-Blu-bar optimization model.

[0046] S400, Model Solving: The trained Physical Information Neural Network (PINN) model is used to make approximate predictions of service quality indicators. The approximate prediction results are used as constraints to solve the two-stage sub-Bruker optimization model, so as to output day-ahead hourly power dispatch instructions.

[0047] In this embodiment, a hybrid queuing network model is used to establish an operation model for the battery swapping station. This model accurately models the complex queuing behavior of electric vehicles and batteries in actual operation of the station by using electric vehicle arrival rate, charging service rate, discharging service rate, and battery swapping service rate. It can accurately depict the dynamic interaction between vehicle flow and battery flow at the overall level, reflecting both queuing and resource sharing characteristics. Simultaneously, this invention utilizes historical moment information of electric vehicle arrival rate to construct a sub-Bruker fuzzy set, thereby transforming the two-stage stochastic programming model into a two-stage sub-Bruker optimization model. This two-stage sub-Bruker optimization model can closely integrate service quality indicators with scheduling optimization strategies under uncertain distribution or incomplete data conditions, finding a more suitable optimal scheduling design for battery swapping stations under uncertain demand. Furthermore, this invention uses a trained physical information neural network model to approximate the prediction of service quality indicators, solving the computational bottleneck of complex stochastic optimization models and greatly balancing model accuracy and computational efficiency. Furthermore, the two-stage stochastic programming model of this invention clearly defines the decision-making relationship between distribution network operators and battery swapping station operators, providing a theoretical framework for their coordinated interaction. It can achieve computable robust optimal scheduling by comprehensively considering the service characteristics of hybrid queuing networks, demand uncertainty, and service quality constraints, and through a physical information neural network model. Thus, it provides a battery swapping station scheduling method that can effectively cope with demand uncertainty and collaboratively optimize economic benefits and service quality.

[0048] In this embodiment, the hybrid queuing network model is composed of an open queuing network model and a closed queuing network model coupled together. The open queuing network model includes a Poisson process to describe the electric vehicle arrival process. The closed queuing network model includes an exponential distribution to describe the charging service time and the discharging service time, and a constant value to describe the battery swapping service time.

[0049] Specifically, the arrival behavior of external electric vehicles (EVs) is random and independent, and the time interval between each charge / discharge is independent and constant. The time interval between each EV arrival or service completion is unaffected by previous events; that is, the time intervals exhibit an exponential distribution, conforming to a Poisson distribution. Therefore, in the open queuing network model, EVs arrive at the battery swapping station according to a Poisson process, with an arrival rate of... .

[0050] In this embodiment, if a fully charged battery is available at the station, the EV immediately accepts service; otherwise, it enters a waiting queue. The waiting queue has a limited capacity, denoted as . If the queue is full, subsequent arriving EVs will be denied service and this will be counted as demand loss.

[0051] In this embodiment, it is assumed that there are a total of The battery packs circulate within the system, and each battery has three states: Fully Charged Battery Queue (FB Queue), Depleted Battery Queue (DB Queue), and Charging. The battery swapping service consumes one Fully Charged Battery Queue and simultaneously generates one Depleted Battery Queue. Charging Piles are responsible for charging the Depleted Battery Queue.

[0052] In step S100, the average waiting time is calculated using Little's Law, which is the product of the probability that the unmet battery swapping demand passes through the system when it is at full load and the vehicle entry rate.

[0053] In this embodiment, the battery swapping service time The number of charging stations is constant, C, and the charging / discharging time follows an exponential distribution. The system state can be represented by a single tuple. To describe, where n represents the number of EVs in the system (including those currently being served and those waiting). ), b represents the number of fully charged batteries in stock ( ).

[0054] The system state changes as an EV arrives, service is completed, or battery charging is finished. For example, the arrival of an EV will cause the state to change from... Become A completed battery swap service will change the status from Become The changes in system state, including the arrival of the EV, battery charging, and service completion, depend only on the current state and are independent of previous history. Therefore, these transitions follow the Markov property, and the system state changes can be modeled using a state transition matrix. According to Markov process theory, a set of state transition equations can be established:

[0055] , Formula (1)

[0056] in:

[0057] It is the state transition rate matrix obtained from Markov's laws;

[0058] π represents the system in all its states. The steady-state probability distribution vector, , It is a unit column vector, and the probabilities of each state of the system are obtained by solving the state transition matrix; among them, the time period set of the battery swapping station operation problem is defined as T, and is indexed by time t.

[0059] Solving the above system of equations yields the steady-state probability. Based on steady-state probability The key service quality can be calculated. )index:

[0060] 1) Average waiting time (W) t According to Little's Law, waiting time can be calculated from the average number of EVs and the effective arrival rate in the system, expressed as:

[0061]

[0062] In the formula:

[0063] This indicates the average waiting time;

[0064] Represents the steady-state probability;

[0065] n represents the number of EVs in the system (including those currently being served and those waiting);

[0066] Indicates the capacity of the waiting queue;

[0067] This indicates the number of batteries circulating in the system within the battery swapping station.

[0068] Indicates the arrival rate of EVs;

[0069] Indicates the battery swapping service time;

[0070] , which represents the average number of trolleys in the system, and will be referred to as E[n] in the following text.

[0071] 2) Battery swapping demand loss rate ( When an EV arrives, if the waiting queue is full, the service is rejected. The loss rate equals the arrival rate multiplied by the probability that the system is in a full queue, expressed as:

[0072]

[0073] In addition, the average number of charging stations used for charging can be calculated. and the number of discharge piles used for discharge .

[0074]

[0075] Example 2

[0076] Based on Example 1, this example provides a specific solution for a two-stage stochastic programming model.

[0077] In this embodiment, step S200 specifically includes the following steps:

[0078] S201. Establish the objective function for power dispatch planning of the distribution network in the first stage;

[0079] S202. Establish line power flow constraints for the power system;

[0080] S203. Establish the objective function for the second phase of the battery swapping station participation in optimized scheduling and backup services;

[0081] S204. Based on the hybrid queuing network model, establish power regulation constraints and energy conversion constraints for battery swapping stations;

[0082] S205. Establish service quality constraints for battery swapping station operation based on service quality indicators.

[0083] Specifically, the objective function of power dispatch planning includes the purchase cost of electricity from the main grid, the settlement cost of power exchange with all battery swapping stations, the generation cost of all distributed generation (DG) generators in the distribution network, and the revenue from selling electricity to base loads.

[0084] The objective of the optimized scheduling in this invention is to minimize the power dispatch planning objective function, calculated as follows:

[0085]

[0086] In the formula:

[0087] This represents the electricity purchase cost of the main power grid during the dispatch period t;

[0088] This represents the settlement cost of the BSS during the scheduling period t;

[0089] This represents the cost of DG power generation during the scheduling period t;

[0090] This represents the revenue generated from selling electricity to ordinary loads during the dispatch period t.

[0091] in, The calculation formula is:

[0092]

[0093] in, This represents the active power output of the distributed generator located at node i during time period t, in kW. , , : These are the quadratic, primary, and zero-order cost coefficients for distributed generator i, respectively. These are inherent techno-economic parameters of the generator set, typically provided by the equipment supplier or obtained by fitting historical operating data, and are used to describe the quadratic function relationship between generation cost and output power. Δt: Represents the length of each scheduling period.

[0094] The calculation formula is:

[0095]

[0096] in This represents the wholesale electricity price purchased from the main grid during time period t, which is known market input data. This represents the active power purchased by the distribution network from the main grid during time period t, and is a decision variable in the model.

[0097] The calculation formula is:

[0098]

[0099] in, This represents the total number of battery swapping stations within the distribution network. This represents the time-of-use price for electricity in time period t. This is the retail electricity price published by the distribution network company to its internal users (including battery swapping stations), and it is the known input data. This represents the charging power of the battery swapping station bss during time period t. This represents the discharge power of the battery swapping station bss during time period t.

[0100] In this embodiment, the line power flow constraints of the power system specifically refer to the ability to describe the network state using linearized DistFlow (Distributed Power Flow) equations, including:

[0101] The active and reactive power of the nodes are balanced, and the total load composition of the nodes is given by formula (12):

[0102]

[0103] Node voltage drop:

[0104]

[0105] Node voltage and line power flow safety constraints:

[0106]

[0107] In the formula:

[0108] This represents the voltage of the j-th node at any time interval t;

[0109] This indicates the maximum voltage value of the node;

[0110] This indicates the minimum voltage value of the node;

[0111] Let represent the active power flowing into node j from the previous level node i at any time period t, and l represent the branch formed by ij;

[0112] This represents the maximum rated active power flowing into node j from its parent node i.

[0113] This represents the minimum rated active power flowing into node j from its parent node i.

[0114] Let represent the reactive power flowing into node j from the previous level node i at any time period t, and l represent the branch formed by ij;

[0115] This represents the maximum rated reactive power flowing into node j from its parent node i.

[0116] This represents the minimum rated reactive power flowing into node j from its parent node i.

[0117] This represents the output of renewable energy at node j during any time period t;

[0118] This represents the basic household active power load at node j during any time period t.

[0119] This represents the basic household reactive load at node j during any time period t.

[0120] V0 is a simplified reference voltage;

[0121] This represents the resistance of the line itself on branch l, which is connected from i to j;

[0122] This represents the reactance of the line itself on branch l, which is connected from i to j.

[0123] In this embodiment, the objective function for the second-stage participation of battery swapping stations in optimized scheduling and backup service is:

[0124]

[0125] In the formula:

[0126] The cost of battery charging is calculated using the following formula:

[0127] Formula (18)

[0128] The discharge benefit is expressed by the following formula:

[0129] Formula (19);

[0130] The formula for calculating battery degradation cost is as follows:

[0131] ;

[0132] The revenue from standby services is calculated using the following formula:

[0133] Formula (21);

[0134] The revenue from battery swapping services is calculated using the following formula:

[0135] Formula (22);

[0136] In formulas (14) to (19):

[0137] In the formula:

[0138] This represents the charging efficiency of the battery swapping station bss during a certain time period t.

[0139] This represents the time-of-use electricity price for period t;

[0140] Indicates charging time;

[0141] Represents various prices;

[0142] It is the slope that approximates the linear relationship between battery life and its slope.

[0143] This refers to the battery's rated capacity.

[0144] It refers to the cost of battery investment.

[0145] In this embodiment, the process of establishing power regulation constraints and energy conversion constraints for the battery swapping station using a hybrid queuing network model is as follows:

[0146] The functions include total charging and discharging power constraints of the battery swapping station, upper and lower limit constraints on the power of a single charging / discharging pile, average number of piles currently being used for charging and discharging at the battery swapping station, dynamic change constraints on the battery pack capacity of the battery swapping station, and energy and power constraints reserved for backup services.

[0147] The charging / discharging power constraint means that the total charging power / discharging power is the product of the power of a single charging pile and the number of working charging piles. The specific formula is as follows:

[0148]

[0149] Formulas (25) to (26) represent the upper and lower limits of the charging and discharging power of a single pile in a battery swapping station;

[0150] In formulas (23) to (26) (pile is used to indicate the variable of a single pile):

[0151] Indicates the total charging power;

[0152] This indicates the number of batteries that the battery swapping station will charge during time period t.

[0153] This represents the average charging power of a single charging pile at the battery swapping station during time period t.

[0154] This represents the total discharge power of the battery swapping station BSS during time period t.

[0155] This indicates the number of batteries discharged by the BSS at the battery swapping station during time period t.

[0156] This represents the average discharge power of a single discharge pile at the battery swapping station during time period t.

[0157] The correlation between the hybrid queuing network model and power refers to the relationship between the charging service rate / discharging service rate and the power of a single charging pile, and its specific expression formula is as follows:

[0158] Formula (27)

[0159] Formula (28)

[0160] In the formula:

[0161] Indicates the charging service rate;

[0162] Indicates the battery's charging power;

[0163] Indicates battery charging efficiency;

[0164] Indicates the charging service rate;

[0165] Indicates the battery's discharge power;

[0166] Indicates battery discharge efficiency;

[0167] Indicates the rated capacity of the battery;

[0168] Reserve and total energy constraint refers to the limitation that the reserve capacity provided by the BSS is limited by its total charging power and stored energy. The specific calculation formula is as follows:

[0169]

[0170] Formula (29)

[0171] Formula (30)

[0172] Formula (31)

[0173] Formula (32)

[0174] Formula (29): Upper and lower limits of reserve capacity constraints, the upward reserve capacity promised by the battery swapping station BSS in time period t. and downward reserve capacity It must be a non-negative value and cannot exceed its own set maximum available backup power. and .

[0175] Formula (30): Dynamic balance of total energy at the battery swapping station. This formula describes the dynamic change in the total state of charge of all batteries at the charging piles within the battery swapping station. Total energy at the current time period t. Equal to the total energy of the previous period Add the energy gained from charging during this period, and subtract the energy lost from discharging.

[0176] Formulas (31) and (32): Constraints on the impact of standby service on the state of energy. Formula (31) guarantees that even after an entire period of continuous upward standby is performed... Afterwards, the remaining total energy will not exceed the safe upper and lower limits. , Formula (32) guarantees that even after a continuous period of downward standby is performed... Afterwards, the total energy will not exceed the safe upper and lower limits. , .

[0177] The energy boundary conditions for the entire battery swapping station are calculated based on the connected batteries as follows:

[0178] Formula (33)

[0179] Formula (34)

[0180] Formula (33) defines the upper limit of energy: it is equal to the rated capacity of a single battery cell. Multiply by the number of charging piles currently in use, where C is the total number of piles; Formula (34) defines the lower limit of energy: it is determined by the maximum depth of discharge. Decide.

[0181] In this embodiment, the formula for establishing service quality constraints for battery swapping station operation based on service quality indicators is as follows:

[0182] Formula (35)

[0183] Formula (36).

[0184] Formula (35): Average waiting time constraint, the average user waiting time calculated by the hybrid queuing network model in time period t. It must be less than or equal to the service standard limit set by the operator. .

[0185] Formula (36): Demand loss rate constraint, the battery swapping demand loss rate calculated by the hybrid queuing network model in time period t. The percentage of vehicles refused service due to lack of a fully charged battery or a full queue must be lower than the acceptable limit set by the operator. .

[0186] Example 3

[0187] Based on any of the above embodiments, this embodiment provides a specific scheme for transforming a two-stage stochastic programming model into a two-stage bibliometric optimization model.

[0188] In this embodiment, the historical moment information includes the mean vector and covariance matrix of the historical data of electric vehicle arrival rate; the support set of the split-bar fuzzy set is the vector set of electric vehicle arrival rate.

[0189] In this embodiment, the transformation process of the two-stage sub-Brubar optimization model includes: based on the sub-Brubar fuzzy set, the expectation function for the continuous probability distribution in the two-stage stochastic programming model is transformed into the expectation function for finding the worst probability distribution within the sub-Brubar fuzzy set.

[0190] Specifically, for the uncertain demand variable, a fuzzy set of EV arrival rate is established, and the two-stage stochastic programming model is transformed into a two-stage fuzzy optimization model based on the fuzzy set. This is to address the EV battery swapping demand rate. The uncertainty is defined by a first moment (mean) ) and second moment (covariance) fuzzy set .

[0191] Formula (37)

[0192] in It is the support set of uncertain variables, denoted as .

[0193] This formula defines a set of probability distributions. Any probability distribution P belonging to this set must simultaneously satisfy three conditions:

[0194] 1. Its random variable The range of values ​​is limited to the support set Ξ (e.g., the arrival rate cannot be negative or exceed a certain physical limit).

[0195] 2. Its expected value (mean) must be equal to the mean vector calculated based on historical data. .

[0196] 3. Its covariance matrix must be equal to the covariance matrix calculated based on historical data. .

[0197] Based on this, the objective function under the two-stage sub-Bruker optimization model for battery swapping station operation can be obtained:

[0198]

[0199] Next, in order to solve the two-stage birus bar optimization model, constraints are applied to it. The specific process of constraint application is as follows:

[0200] 1) For uncertain demand variables, a sampling method is used to generate... A sample set consisting of discrete sample points;

[0201] 2) The objective function for obtaining the expectation of the continuous probability distribution in the two-stage sub-Bruker optimization model is approximated as a weighted sum of the N discrete sample points;

[0202] 3) Introduce weight variables As optimization variables, and constraining the sum of the weights of all samples to be 1, and requiring that the first and second moments of the weighted sample set be consistent with the preset mean and covariance matrices, the worst-case probability distribution is sought within the sample space. Therefore, the fuzzy set can be transformed into:

[0203] Formula (39)

[0204] This formula transforms the original description of the probability distribution P into a description of a set of sample weight vectors w. Here, we pre-generate a fixed set of discrete samples (scenes) ζs. Now, finding a worst-case "probability distribution" is equivalent to finding an optimal set of "sample weights". This set of weights must satisfy:

[0205] 1) The sum of all weights is 1 and non-negative (forming a valid discrete probability distribution).

[0206] 2) The mean of the sample weighted by this set of weights must be equal to the preset mean. ;

[0207] 3) The sample covariance weighted by this set of weights must be equal to the preset covariance. .

[0208] Therefore, the above objective function can be transformed into the following form:

[0209] Formula (40).

[0210] This is a standard form of two-stage Bruker optimization.

[0211] Phase 1 Decision: Decide on a decision x that is independent of uncertainty and must be made here and now, with a cost of [missing information]. .

[0212] The second-stage decision-making process involves selecting the worst-case weight distribution w after the first-stage decision is made to maximize subsequent costs. For each specific scenario ζs, an optimal response decision will then be made. To minimize the cost in this scenario .

[0213] Example 4

[0214] This embodiment provides a specific solution for the two-stage sub-Bruker optimization model, based on any of the above embodiments.

[0215] In this embodiment, to address the computational complexity issue caused by directly solving the hybrid queuing network model in the model, a physical information neural network is used to approximate the prediction of the key output indicators of the hybrid queuing network model. The key output indicators include the average number of charging piles used, the average number of discharging piles used, the average user waiting time, and the battery swapping demand loss rate. At the same time, in order to ensure the physical consistency of the approximate prediction results, the physical information neural network introduces constraints describing the inherent physical relationships between the indicators as part of the loss function during the training process.

[0216] In this embodiment, the solution steps for the two-stage sub-Bruker optimization model specifically include:

[0217] 1) Based on the historical data of random variables existing in the constraints of the optimization scheduling model, solve for the moment information of the uncertain variables and construct their fuzzy sets;

[0218] 2) Transform the standard two-stage sub-Brussels bar optimization model with uncertain vectors described by fuzzy sets into a sample-based two-stage sub-Brussels bar optimization model;

[0219] 3) The physical information neural network model is trained using a large number of data samples generated offline by the hybrid queuing network model, and the training results are converted into a form supported by the solver.

[0220] 4) A nested column and constraint generation algorithm is adopted, and a solver is used to solve the optimization scheduling model containing convex linear constraints and neural network constraints to obtain the power dispatching and battery swapping station operation instructions of the distribution network.

[0221] In this embodiment, the training process of the physical information neural network model is as follows:

[0222] In the loss function for training the physical information neural network model, a constraint term reflecting the inherent physical relationship of service quality indicators is introduced;

[0223] In solving the subproblems of the column and constraint generation algorithm, the physical information neural network model is called to quickly obtain approximate values ​​of the service quality index corresponding to the current decision variable, instead of directly solving the hybrid queuing network model.

[0224] In this embodiment, the sample data comes from a hybrid queuing network model. The hybrid queuing network model is used to calculate the charging and discharging power, the arrival rate, and the service quality index. The original training data can be obtained by randomly sampling multiple times. The training result is the mapping relationship between these parameters.

[0225] Specifically, the training data for the physical information neural network is generated by an offline computational hybrid queuing network model. The trained physical information neural network model is then integrated into C&CG as a constraint condition to replace the original complex matrix operations, thereby achieving rapid solution. In the construction of the physical information neural network model, a linear ReLU (Rectified Linear Unit) function is used as the activation function, thus equating the physical information neural network model with several linear equality constraints.

[0226] Specifically, a physical information neural network is established to approximate the key output results of the hybrid queuing network model.

[0227] In C&CG, solving the subproblems requires obtaining the average waiting time and battery swapping demand loss rate by solving the state transition matrix of the hybrid queuing network model based on the input charging / discharging strategy (i.e., decision variables). This embodiment uses PINN to create a surrogate model. The PINN training data is used to approximate the difficult-to-solve metrics, thus obtaining an approximate prediction close to the true value.

[0228] In this embodiment, a hybrid queuing network model is used as a simulation model to obtain large-scale data for PINN training. The PINN training process is as follows:

[0229] 1) Training: First, run the queuing network model offline to generate a large amount of "input-output" sample data. The input is the decision variables of the BSS, including charging power. Discharge power tram arrival rate The output is a QoS indicator, including latency. Loss rate and intermediate variables (including the number of charging stations used for charging) and the number of discharge piles used for discharge ).

[0230] 2) Physical Constraints: During the training of the physical information neural network model, in addition to minimizing the error between the predicted value and the true value (obtained from the offline queuing model), a physical constraint term is added to the loss function to eliminate approximate predictions that do not satisfy physical laws. The physical constraint term includes:

[0231] The product of service time (the sum of average waiting time and battery swapping service time) and EV actual service acceptance rate (the difference between EV arrival rate and battery swapping demand loss rate) should be equal to E[n].

[0232] The sum of the number of charging piles used for charging and the number of discharging piles used for discharging should be less than the total number of charging piles C.

[0233] The specific expressions for the physical constraints are as follows:

[0234] Formula (41)

[0235] 3) Application: The trained PINN is embedded into the solution of subproblems in C&CG, replacing the complex calculations of the original queuing model, thus transforming the subproblems into an easier-to-solve form and greatly improving the overall efficiency of the algorithm.

[0236] Specifically, the ReLU (Rectified Linear Unit) function is used as the activation function. Based on this linear activation function, PINN can be equivalent to several linear equality constraints. These constraints are then integrated into the solver's programming language using the Gurobi_ML (Gurobi Machine Learning) library, enabling the integration of prediction methods during the solution process. In the two-stage stochastic programming model, the Quality of Service (QoS) indicator, i.e., average latency... With average demand loss The average waiting time can be obtained from formulas (2) and (3). With average demand loss Quantitative relationships, in addition to average waiting time With average demand loss Since the variables need to be nonnegative, physical constraints need to be added in addition to the data-driven loss, and a weight hyperparameter needs to be introduced to balance the importance of the two objectives of "fitting the data" and "obeying physical laws". The specific physical information constraints are as follows:

[0237]

[0238] In this embodiment, the nested column and constraint generation algorithm specifically includes a nested column and constraint generation algorithm to decompose the original problem into a master problem (MP) and subproblems (SP).

[0239] The formula for calculating the main problem is:

[0240] MP= Formula (44);

[0241] This plane acts as the cutting plane. In each iteration of the C&CG algorithm, a new constraint (i.e., the cutting plane) is generated from the subproblem to continuously raise the bar. The value of is gradually brought closer to the true minimum cost of the second stage.

[0242] The formula for calculating the subproblems is:

[0243] Formula (45)

[0244] The subproblem is further decomposed into two types: SP-1 and SP-2.

[0245]

[0246] In the formula:

[0247] θ represents an auxiliary variable used to approximate the expected cost of the second stage;

[0248] This represents the value obtained from solving the decision variables in the first stage of the current round;

[0249] This indicates the cost of the second phase.

[0250] The probability weights represent the worst-case scenario.

[0251] This represents the total cost of the second phase;

[0252] The main problem and subproblems are solved alternately. Subproblems are decomposed into weighted solutions and minimum cost solutions with fixed weights. Subproblems are used to determine the worst-case demand distribution, while the main problem updates operational decisions, until the upper and lower bounds converge. The specific steps are as follows:

[0253] 1) Initialization: Set the number of iterations The lower realm Upper Realm Generate sample demand scenarios;

[0254] 2) Solving the main problem: Solving the current MP to obtain the first-stage decision. and auxiliary variables Update the Nether ;

[0255] 3) Solving subproblems: Substitute into SP-1; for each requirement scenario, solve SP-1 in parallel to obtain the optimal second-stage cost. Then solve for SP-2 to find the worst-case probability weights. Calculate the total cost of the second phase. Update the upper boundary In this step, the PINN model is used to quickly calculate QoS metrics.

[0256] 4) Convergence test: Check the difference between the upper and lower bounds. Is it less than a preset convergence threshold? If the condition is met, the algorithm converges and outputs the current optimal solution; otherwise, a new optimality cut is constructed using the solution from SP-2, as shown below:

[0257]

[0258] Add it back to the main problem and increase the number of iterations. And then solve the main problem again.

[0259] In this embodiment, a two-stage sub-Bruker optimization model containing convex linear constraints and PINN constraints is solved using a solver to obtain power dispatching instructions for the distribution network and operation instructions for the battery swapping station.

[0260] Specifically, the solution results of the two-stage distributed bar optimization model include the decision variables of the distribution network and the battery swapping station at each scheduling time t∈T. The solution results related to the distribution network include the results at each scheduling time. Any distributed generator Active power output and the power purchased from the main grid The solution results related to the distribution network are expressed as follows:

[0261] Formula (49);

[0262] The solution results related to the battery swapping station include each scheduling time. The charging power of any BSS at any previous battery swapping station Discharge power Upward reserve capacity and downward reserve capacity The solution result for the battery swapping station is expressed as follows:

[0263] Formula (50);

[0264] After solving the two-stage distributed bar optimization model, the power system dispatching agency can apply the above solution results to different dispatching times. Dispatch instructions are issued to distributed generators and battery swapping stations within the system, thereby enabling the economical, safe, and reliable operation of the entire distribution network. Based on these dispatch instructions, battery swapping stations execute corresponding charging / discharging plans and reserve power, allowing them to participate in the energy and backup services markets within the electricity market.

[0265] Example 5

[0266] This embodiment provides an optimized scheduling system for battery swapping stations, based on the above embodiments.

[0267] In this embodiment, a battery swapping station optimization scheduling system includes a first model building module, a second model building module, an optimization module, and a calculation module.

[0268] The first model construction module uses a hybrid queuing network model to establish a battery swapping station operation model and calculates service quality indicators through the hybrid queuing network model. Among them, the service quality indicators include at least the average waiting time and the amount of unmet battery swapping demand. The hybrid queuing network model describes these indicators through electric vehicle arrival rate, charging service rate, discharging service rate, and battery swapping service rate.

[0269] The second model building module establishes a two-stage stochastic programming model using distribution network scheduling costs, battery swapping station operating costs, and service quality indicators as parameters.

[0270] The optimization module constructs a sub-Brussels bar fuzzy set based on the historical moment information of electric vehicle arrival rate, and transforms the two-stage stochastic programming model into a two-stage sub-Brussels bar optimization model.

[0271] The calculation module uses a trained physical information neural network model to make approximate predictions of service quality indicators, and uses the results of the approximate predictions as constraints to solve a two-stage sub-Bruker optimization model to output day-ahead hourly power dispatch instructions.

[0272] Example 6

[0273] This embodiment provides some alternative solutions for the modeling method of hybrid queuing network models, based on any of the above embodiments.

[0274] In one possible embodiment, the open queuing network is described using an Erlang distribution or a phase-type distribution.

[0275] Specifically, the Erlang distribution and the more general phase distribution can be constructed by combining multiple consecutive, independent exponential distribution stages. This conforms to the memoryless and random nature of EV arrivals and the regularity of service times.

[0276] Specifically, a K-order Erlang distribution can accurately describe a service process as consisting of K independent, consecutive steps with the same service rate.

[0277] Applicable scenarios: When actual operational data shows that the battery charging process has very clear stages (for example, the first 80% of the constant current fast charging stage has a relatively fixed time, while the last 20% of the constant voltage charging stage has a large variation in time), using Erlang or phase distribution can more accurately characterize the real distribution of service time, thereby improving the simulation accuracy of the hybrid queuing network model.

[0278] In another possible embodiment, a time-varying Poisson process can also be used to describe the hybrid queuing network model.

[0279] Specifically, the standard Poisson process assumes that the average arrival rate λ of vehicles is constant within each scheduling period (e.g., 1 hour); the time-varying Poisson process allows the arrival rate λ(t) to become a function that varies with time within that period.

[0280] Applicable Scenarios: This solution is suitable for characterizing fine-grained demand fluctuations within scheduling periods. For example, during the evening peak hour of 18:00-19:00, actual vehicle arrivals may be highly concentrated in the first 20 minutes. Using a time-varying Poisson process can more accurately capture this "instantaneous peak" phenomenon, thereby making more accurate predictions of short-term congestion and service quality at battery swapping stations. It is suitable for operational scenarios with extremely high service level requirements.

[0281] Example 7

[0282] This embodiment provides some alternative solutions for the two-stage sub-Bruker optimization model based on any of the above embodiments.

[0283] In one possible embodiment, the progressive hedging (PH) algorithm is used to solve the two-stage sub-Bruker optimization model.

[0284] Specifically, PH first decomposes the original problem into multiple independent subproblems according to the uncertainty scenario and solves them in parallel. Then, by introducing a quadratic penalty term, it gradually "hedges" and penalizes the differences between the decision variables in the first stage of each scenario in multiple iterations, and finally makes it converge to a unified, non-anticipative solution.

[0285] Applicable Scenarios: When the number of uncertain scenarios (samples) is extremely large, the PH algorithm, due to its inherent parallel computing characteristics, may have a greater computational efficiency advantage than the serial master-sub-problem structure of C&CG. It provides an efficient solution path with a different architecture but consistent goal for solving the two-stage model proposed in this invention.

[0286] In another possible embodiment, stochastic dual dynamic programming (SDDP) is used to solve the two-stage bibliometric optimization model.

[0287] Specifically, SDDP is a method for solving multi-stage stochastic optimization problems by constructing a Benders decomposition of the future cost function (value function) from backwards. For solving two-stage SDDP optimization models, it can be viewed as a special implementation of Benders decomposition, where the algorithm continuously generates a lower bound linear approximation of the expected cost in the second stage through forward simulation and backward recursion.

[0288] Applicable Scenarios: Although SDDP is mainly used for multi-stage problems, its core idea, Bandel decomposition, is fully applicable to the two-stage model of this invention. In particular, if the model of this invention needs to be extended to multi-day, rolling optimization scheduling in the future, SDDP will become a very natural and efficient solution framework and an important alternative algorithm for future extended applications.

[0289] Example 8

[0290] This embodiment provides some optional solutions for approximate prediction of service quality indicators, based on any of the above embodiments.

[0291] In one possible implementation, Gaussian Process Regression (GPR) is used to approximate the prediction of service quality metrics.

[0292] Specifically, GPR is a non-parametric Bayesian regression method that can provide not only a predicted value, but also the confidence interval (a measure of uncertainty) of that predicted value.

[0293] Applicable Scenarios: In the initial stage of battery swapping station operation, when training data samples are limited, GPR is particularly suitable for initial calculations and simulations due to its excellent small-sample learning ability and resistance to overfitting. Furthermore, the prediction uncertainty information provided by GPR can be used for additional risk assessment of scheduling schemes, such as identifying operational intervals where model predictions are inaccurate or where potential risks are high.

[0294] In another possible implementation, gradient boosted trees (such as XGBoost, LightGBM) are used to make approximate predictions of service quality metrics.

[0295] Specifically, gradient boosting decision trees build a strong model by iteratively training a series of weak decision tree learners.

[0296] Applicable Scenarios: Gradient boosting decision trees typically exhibit extremely high prediction accuracy and training efficiency when processing tabular data. In scenarios where extreme prediction accuracy is desired and a certain level of model interpretability is required, gradient boosting decision trees are a strong competitor and alternative to PINN.

[0297] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, several equivalent substitutions or obvious modifications can be made without departing from the concept of the present invention, and all such modifications, achieving the same performance or purpose, should be considered within the scope of protection of the present invention.

Claims

1. A battery swap station optimization scheduling method, characterized in that, The method comprises the following steps: S100, a battery swap station service model is constructed, a battery swap station operation model is established by using a hybrid queuing network model, and a service quality index is calculated through the hybrid queuing network model; wherein the service quality index at least includes an average waiting time and an unmet battery swap demand, and the hybrid queuing network model is described by an electric vehicle arrival rate, a charging service rate, a discharging service rate and a battery swap service rate; S200, a two-stage stochastic programming model is established, and a two-stage stochastic programming model is established by taking a power distribution network scheduling cost, a battery swap station operation cost and the service quality index as parameters; S300, a robust optimization model is established, a distribution robust fuzzy set is constructed based on historical moment information of the electric vehicle arrival rate, and the two-stage stochastic programming model is converted into a two-stage distribution robust optimization model; S400, model solving, the service quality index is approximately predicted by using a physical information neural network model, and the result of the approximate prediction is taken as a constraint condition, the two-stage distribution robust optimization model is solved, and a day-ahead hour-level power scheduling instruction is output; the training process of the physical information neural network model comprises: a large amount of data samples generated offline by the hybrid queuing network model are used for training; in the loss function of training the physical information neural network model, a constraint term reflecting the internal physical relationship of the service quality index is introduced; the physical information neural network model adopts a linear activation function, so that the trained physical information neural network model can be equivalently represented as a plurality of linear equality constraint conditions, so as to be embedded into the two-stage distribution robust optimization model.

2. The method of claim 1, wherein, In step S100, the hybrid queuing network model is coupled by an open queuing network model and a closed queuing network model; The open queuing network model comprises an electric vehicle arrival process described by a Poisson process; The closed queuing network model comprises charging service time and discharging service time described by an exponential distribution, and battery swap service time described by a constant value.

3. The method of claim 1 or 2, wherein, In step S100, the average waiting time is calculated by Little's law, and the unmet battery swap demand is calculated by the product of the probability that the system is in a full load state and the vehicle entry rate.

4. The method of claim 1, wherein, Step S200 specifically comprises the following steps: S201, a power scheduling programming objective function of a first-stage power distribution network is established; S202, line flow constraints of a power system are established; S203, a battery swap station participating in optimization scheduling and standby service objective function of a second-stage is established; S204, based on the hybrid queuing network model, battery swap station power regulation constraints and energy conversion constraints are established; S205, based on the service quality index, battery swap station operation service quality constraints are established.

5. The method of claim 1, wherein, In step S300, the historical moment information includes a mean vector and a covariance matrix of the electric vehicle arrival rate historical data; and a support set of the distribution robust fuzzy set is a vector set of the electric vehicle arrival rate.

6. The method of Claim 1, wherein, In step S400, a nested column and constraint generation algorithm is used for solving the two-stage distribution robust optimization model; The column-and-constraint generation algorithm solves the master problem and the sub-problem alternately to determine the worst-case probability distribution and update the operation decision until convergence.

7. The method of claim 6, wherein, The sub-problem solving of the column-and-constraint generation algorithm comprises calling the physical information neural network model to quickly obtain an approximate value of the service quality index corresponding to the current decision variable, instead of directly solving the mixed queuing network model.

8. A battery swap station optimization scheduling system for implementing the method of any one of claims 1-7. Comprise: A first model construction module, which establishes a battery swap station operation model by using a mixed queuing network model, and calculates a service quality index through the mixed queuing network model; wherein the service quality index at least comprises an average waiting time and an unmet battery swap demand amount, and the mixed queuing network model is described by an electric vehicle arrival rate, a charging service rate, a discharging service rate and a battery swap service rate; A second model construction module, which establishes a two-stage stochastic programming model by taking a power distribution network scheduling cost, a battery swap station operation cost and the service quality index as parameters; An optimization module, which constructs a distribution robust fuzzy set based on historical moment information of the electric vehicle arrival rate, and converts the two-stage stochastic programming model into a two-stage distribution robust optimization model; A calculation module, which uses a physical information neural network model to approximately predict the service quality index, and uses a result of the approximate prediction as a constraint condition to solve the two-stage distribution robust optimization model, so as to output a day-ahead hourly power scheduling instruction.

Images