Multi-agent-oriented vehicle network interaction demand service method and system

Abstract

The invention discloses a multi-subject-oriented vehicle network interaction demand service method and system, and the method comprises the steps: constructing a multi-subject-oriented vehicle network interaction service framework, and building an optimization model which comprises a charging demand prediction objective function and a multi-subject constraint condition; based on the business framework, establishing a multi-source heterogeneous data processing module, collecting and processing multi-source heterogeneous data from the multiple subjects, and outputting a standardized data set; based on the processed multi-source heterogeneous data, an electric vehicle user portrait module is established, an improved BIRCH clustering algorithm is adopted to perform clustering analysis on the electric vehicle user data, and a user portrait is constructed; based on the established user portrait, establishing a user charging service demand prediction module, and predicting a user charging demand by adopting a TCN-DGAT fusion algorithm; the method can quickly respond to the demand change of the power grid and the user, and dynamically adjust the scheduling strategy, thereby remarkably improving the resource utilization efficiency and the response speed.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent power distribution network optimization operation, and particularly relates to a multi-agent-oriented vehicle-to-grid interaction demand service method and system. BACKGROUND

[0002] With the continuous improvement of the penetration rate of electric vehicles, the charging load and the vehicle energy storage resources have been deeply integrated into the power system regulation system, and the scale of the "vehicle-to-grid" two-way energy interaction is showing a rapid growth trend. The large-scale application of V2G technology enables electric vehicle clusters to simultaneously act as flexible loads and distributed energy storage resources of the power grid, participate in frequency regulation, peak regulation and other types of auxiliary services, and significantly improve the flexibility and renewable energy consumption capacity of the power system. The coordinated development of electric vehicles and power grids is gradually becoming an important part of the flexible regulation resource pool in the construction of new power systems, and the industry as a whole is showing the characteristics of scale, interaction and clean.

[0003] The prior art mainly relies on centralized scheduling architecture and one-way control mode. Typical implementation methods include aggregating electric vehicle charging and discharging resources through a virtual power plant (VPP), and relying on traditional SCADA systems for data collection and dispatching instruction issuance; at the user incentive level, most projects use a fixed price subsidy mechanism to guide users to participate in V2G response; and in terms of system coordination, limited power regulation is achieved through independent charging facility control systems and local grid monitoring.

[0004] However, the existing technology has a most obvious shortcoming: the centralized scheduling architecture cannot effectively adapt to the massive, dispersed and dynamically accessed electric vehicle resources, resulting in serious response delay and significant decline in resource adjustability. Actual operation data shows that the traditional monitoring system has significant delay in processing large-scale real-time electric vehicle states, resulting in a large mismatch between dispatching instructions and actual power grid demand. This rigid paradigm based on one-way control cannot cope with the multiple uncertainties brought about by vehicle mobility, user behavior heterogeneity and dynamic changes in power grid operation state, severely restricting the improvement of vehicle-to-grid interaction scale and resource utilization efficiency. SUMMARY

[0005] The purpose of the present application is to provide a multi-agent-oriented vehicle-to-grid interaction demand service method that realizes real-time response and dynamic optimization scheduling of electric vehicle clusters, effectively solving the response delay and resource mismatch problems under the centralized control mode. On the other hand, a multi-agent-oriented vehicle-to-grid interaction demand service system is provided.

[0006] Technical solution: The vehicle-to-grid interaction demand service method provided by the present application comprises:

[0007] A multi-agent oriented vehicle-grid interaction business framework is constructed, the business framework including three agents of electric vehicle owners, power grid companies and charging station operators, and the core being a charging demand prediction target function and multi-agent constraint conditions;

[0008] Based on the business framework, multi-source heterogeneous data from the multi-agent is collected and processed, and a standardized data set is output;

[0009] Based on the standardized data set, an improved hierarchical balanced iterative reduction clustering (BIRCH) algorithm is used to perform clustering analysis on electric vehicle user data, and a user portrait is constructed, the improvement including introducing a time decay factor and a streaming pruning mechanism;

[0010] Based on the established user portrait, a time convolution network-double graph attention (TCN-DGAT) fusion algorithm is used to predict user charging demand, the TCN-DGAT fusion algorithm including a TCN module and a DGAT module, wherein the TCN module is used for time feature extraction, and the DGAT module is used for dynamic spatial dependency modeling.

[0011] The present application realizes multi-agent benefit coordination and constraint planning by constructing a multi-agent (electric vehicle owners, power grid companies, charging station operators) oriented vehicle-grid interaction business framework and optimization model; the data is subjected to credibility and time sequence weighted fusion and abnormal cleaning through a multi-source heterogeneous data processing module, which significantly improves the data quality and reliability; the user portrait is constructed based on the improved BIRCH clustering algorithm (introducing a time decay factor and streaming pruning), which enhances the timeliness and dynamic adaptability of user behavior description; the TCN-DGAT fusion algorithm is used for charging demand prediction, which takes into account time feature extraction and dynamic spatial dependency modeling, thereby supporting real-time response and dynamic optimization scheduling of electric vehicle clusters, effectively overcoming the response delay and resource mismatch problem under the centralized control mode, and improving the system operation efficiency and resource allocation accuracy.

[0012] Preferably, the target function corresponding to the business framework is minimization of charging demand prediction error:

[0013]

[0014] In the formula, is the error of charging demand prediction, y k is the kth actual charging demand value, is the kth predicted charging demand value, and N is the sample number.

[0015] By setting the objective function as minimizing the charging demand prediction error, the technical solution can systematically optimize the output accuracy of the prediction model, significantly improving the estimation ability of future charging demand; This design forces the model to continuously learn and adjust parameters to approximate the real charging demand, thereby enhancing the reliability and practicality of the prediction results; The error minimization mechanism helps to reduce the risk of unreasonable resource allocation caused by prediction bias, providing more accurate data support for power grid dispatching and charging station operation, thereby improving the collaborative efficiency and response real-time of the entire vehicle-grid interaction system, effectively alleviating the imbalance between supply and demand, optimizing energy distribution and infrastructure utilization.

[0016] Preferably, the constraint conditions corresponding to the business framework include:

[0017] Power grid operation constraints: total power supply of the power grid Satisfies:

[0018]

[0019] The voltage V of each charging pile m,t Satisfies:

[0020] V min ≤V m,t ≤V max

[0021] In the formula, is the total power supply of the power grid, is the power of the mth charging pile, is the line loss, V m,t is the voltage of the charging pile m, V min and V max are the lower and upper limits of the voltage, respectively;

[0022] Charging pile constraints: actual charging power of the charging pile Satisfies:

[0023]

[0024] In the formula, is the actual charging power of the charging pile m, is the rated charging power of the charging pile m, is the maximum charging power acceptable by the vehicle v;

[0025] User behavior constraints: vehicle charging time satisfies:

[0026]

[0027] SOC δ after charging v Satisfies:

[0028]

[0029] wherein: is the charging time of the vehicle v, is the maximum charging time that the vehicle v can accept, δ v is the SOC of the vehicle v after charging, is the minimum SOC required by the vehicle v after charging.

[0030] By introducing three types of constraint conditions of power grid operation, charging pile capacity and user behavior, the technical scheme constructs a multi-dimensional collaborative optimization framework considering technical feasibility, equipment safety and user satisfaction: the power grid operation constraint ensures the balance of total power supply and the stability of each node voltage, effectively prevents system overload or voltage overrun, and guarantees the safe and reliable operation of the power grid; the charging pile constraint limits the actual power between the rated value and the vehicle acceptance capacity, which not only protects the charging equipment from over-limit loss, but also adapts to the individual charging needs of different vehicles; the user behavior constraint meets the charging time and minimum SOC requirements, fully respects the travel habits and expectations of the vehicle owner, and significantly improves the user experience and participation willingness. These constraint conditions work together to make the optimization scheduling strategy physically executable and resource-efficient, thereby systematically improving the stability, safety and user stickiness of the vehicle-grid interaction system.

[0031] Preferably, the multi-source heterogeneous data includes:

[0032] Electric vehicle owner side data: including structured data of vehicle real-time SOC, driving range, charging request time, and semi-structured data of user charging preference log, navigation destination information;

[0033] Power grid company side data: including structured data of real-time power supply, node voltage, line loss, and time series data of daily power fluctuation curve;

[0034] Charging station operator side data: including structured data of charging pile rated power, actual charging current, equipment fault code, and semi-structured data of charging pile operation state log;

[0035] External auxiliary data: including structured data and time series data.

[0036] By integrating multi-source heterogeneous data from electric vehicle owners, power grid companies, charging station operators and external environment, the data processing module builds a comprehensive and high-dimensional information perception system: the owner-side data (such as real-time SOC, driving range, charging preference) provides accurate user behavior insight, enabling the system to understand and predict individual needs; the grid-side data (such as real-time power, node voltage, fluctuation curve) reflects the running state and carrying capacity of the power grid in real time, providing a safety boundary for dynamic scheduling; the operator-side data (such as charging pile power, operating status, fault code) accurately depicts the actual availability and health of charging facilities, ensuring the physical feasibility of resource scheduling; external auxiliary data enhances the system's adaptability to the macro environment (such as traffic, weather). The deep integration and collaborative processing of multi-source data generates a highly standardized and credible data set, laying a solid data foundation for subsequent accurate prediction and optimal scheduling, greatly improving the panoramic perception and collaborative decision-making ability of the vehicle-grid interaction system in complex and variable environments.

[0037] Preferably, the processing of multi-source heterogeneous data from the multiple subjects includes assigning credibility weights and time sequence weights to the data, and weighted fusion:

[0038] Assign credibility weights ω to different data sources cre , whose expression is:

[0039]

[0040] In the formula: Err(D i ) is the historical observation error of data source D i ;

[0041] Assign time sequence weights ω to different time sequence data t , whose expression is:

[0042]

[0043] In the formula: Δt i = t-t i is the interval between the data collection time t i of data source D i and the current processing time t; λ f is the time decay factor;

[0044] Fuse the credibility weights and time sequence weights to calculate the comprehensive weight ω:

[0045]

[0046] In the formula:

[0047] Weighted fusion of multi-source observation values based on comprehensive weights:

[0048]

[0049] wherein: X f (t) is the fused target data value at time t, X i (t) is the observation value of the target data item at time t, ∈(t) is the fusion residual, satisfying ∈(t) ~ N(0, σ i )。 2

[0050] By assigning different data sources with credibility weights based on historical errors and time decay-based time sequence weights, and performing weighted fusion, the data processing mechanism realizes intelligent weighting and dynamic integration of multi-source heterogeneous data: the credibility weight gives greater influence to the data source with high historical accuracy, effectively improving the overall reliability and precision of the fused data; the time sequence weight gives priority to the more recent data with higher freshness, and weakens the interference of obsolete data through a time decay factor, significantly enhancing the real-time representation ability of the data set to the current system state; the combined weight of the two balances the accuracy and timeliness of the data, so that the final weighted fused data value can reflect both long-term stable rules and short-term dynamic changes. This process greatly reduces the negative impact of noise and outliers, generating high-quality and highly consistent standardized data, providing a solid and reliable data foundation for subsequent accurate prediction and optimization decision-making, thereby comprehensively improving the state perception accuracy and dynamic response capability of the vehicle-network interaction system.

[0051] Preferably, the processing of the multi-source heterogeneous data from the plurality of subjects further comprises outlier rejection:

[0052] Calculate the mean μ i and standard deviation σ i of each data source D i (1), X i (2),..., X i (n): i

[0053]

[0054] If an observation value X i (t) satisfies |X i (t) - μ i | > 3σ i , it is determined as an outlier Replace it with the current time fused value X f (t):

[0055]

[0056] ​​The modified standardized data set X'(t)={X1'(t), X2'(t), X3'(t), X4'(t)} is outputted.

[0057] By the abnormal value detection and correction mechanism based on the historical data mean and standard deviation, the data processing method significantly improves the integrity and quality of the data set: the reasonable fluctuation range of the data is dynamically set by using the statistical principle, which can effectively identify and eliminate abnormal observation values caused by equipment failure, transmission error or temporary interference; the weighted fusion value at the current time is used for replacement correction, which not only avoids the data missing problem caused by direct deletion, but also ensures the continuity and smoothness of the data sequence; this process greatly reduces the interference of abnormal data on the subsequent prediction and optimization model, prevents the decision risk of "garbage in, garbage out", and thus generates a highly reliable and consistent standardized data set, providing a stable and clean data basis for the vehicle-network interaction system, and ensuring the accuracy and robustness of the cluster response and scheduling strategy.

[0058] Preferably, the improvement in the improved hierarchical balanced iterative reducing clustering BIRCH algorithm comprises:

[0059] A time decay factor is introduced to optimize path selection, and the effective distance calculation formula is:

[0060] d e =d(x,c)×(1-λ·Δt)

[0061] In the formula, d e is the effective distance, d(x,c) is the original distance between the sample x and the cluster c, λ is the decay coefficient, which controls the influence strength of time on distance, and Δt is the time interval of cluster update;

[0062] If the cluster radius exceeds the threshold value r T after insertion, and the time sequence span in the cluster exceeds 24 hours, a splitting operation is forcedly performed.

[0063] The improvement in the improved hierarchical balanced iterative reducing clustering BIRCH algorithm further comprises a streaming pruning mechanism:

[0064] For a dormant cluster that has not been updated for more than 24 hours, if N<100, it is directly deleted; if N≥100, it is down-weighted as a historical cluster;

[0065] A time weight ω is introduced for each CF node:

[0066]

[0067] In the formula, N' is the number of samples within 24 hours, is the time of the last update, is the current time

[0068] Periodically delete low-activity branches and rebalance the CF tree structure after pruning to ensure that the tree height does not exceed the preset value.

[0069] By introducing a time decay factor to optimize path selection and using a streaming pruning mechanism, the improved BIRCH clustering algorithm significantly improves the timeliness, dynamic adaptability and computational efficiency of user portraits: the time decay factor weakens the influence of old history by weighting recent data, making the clustering results better reflect the current behavior patterns of users and enhancing the real-time accuracy of the portraits; forced splitting of clusters with large time spans and the use of a streaming pruning mechanism can automatically identify and delete low-activity dormant clusters or historical clusters, and preferentially retain high-frequency update clusters, effectively avoiding portrait deviations caused by outdated data and ensuring that the model continues to focus on the latest user group features; the introduction of time weights and periodic pruning to rebalance the CF tree structure not only reduces memory usage and computational complexity, but also ensures the algorithm's ability to quickly process streaming data, thereby supporting online dynamic updates and long-term evolution of user portraits and providing an efficient and robust clustering basis for accurately predicting user charging needs.

[0070] Preferably, the TCN module is used for time feature extraction, which includes:

[0071] A causal convolution is used to ensure that the output of time step t only depends on the current and past inputs, and its calculation formula is:

[0072]

[0073] In the formula, y t is the output of time step t; w k is the kth weight of the convolution kernel; x t-k is the value of the input sequence at time step t-k, and K is the size of the convolution kernel;

[0074] A dilated convolution is used to expand the receptive field, and its calculation formula is:

[0075]

[0076] In the formula, d is the dilation factor, i.e. the interval sampling step;

[0077] A residual connection is introduced, and its calculation formula is:

[0078] H(x)=F(x)+W s ·x

[0079] In the formula, H(x) is the output of the residual connection, F(x) is a function composed of one or more convolution layers, W s is a 1×1 convolution used to adjust the number of channels when the input and output dimensions do not match.

[0080] By adopting causal convolution, dilated convolution and residual connection, TCN module realizes efficient and robust feature extraction of charging demand time series: causal convolution strictly follows the time sequence, ensures that the model only predicts based on historical and current information, eliminates future data leakage, and guarantees the practicability and reliability of prediction in online deployment; dilated convolution greatly expands the receptive field of the model through an exponentially increasing dilation factor, greatly expanding the model's receptive field without significantly increasing the computational burden, enabling it to capture multi-scale temporal dependencies from short-term fluctuations to long-term trends; the introduction of residual connection effectively alleviates the gradient vanishing problem in deep networks, promotes the cross-layer flow of key information, and enhances the training stability and feature expression ability of the model. These mechanisms work together to enable TCN module to extract rich and accurate temporal features from complex and non-stationary charging demand data, laying a solid foundation for subsequent fusion of spatial information for accurate prediction.

[0081] Preferably, the DGAT module is used for dynamic spatial dependence modeling, comprising:

[0082] Dynamically generating an adjacency matrix:

[0083]

[0084] In the formula: is the feature of node i at time t-1; W a is a learnable weight matrix, and u is a similarity score vector. || is a vector concatenation operator symbol;

[0085] Calculate the attention coefficient:

[0086]

[0087] In the formula: W q and W k are the projection matrices of the query and the key, respectively, and is the attention parameter vector;

[0088] θ is an improved activation function, whose expression is as follows:

[0089]

[0090] In the formula: α is a fixed positive number, used to control the slope of the negative number region;

[0091] Normalize the attention weight:

[0092]

[0093] In the formula: N i (t) is the neighbor set of node i at t;

[0094] Aggregate neighbor information to update node representation:

[0095]

[0096] wherein: is the information of the current state, W v is the projection matrix of the value;

[0097] Introducing time encoding:

[0098] Φ(t) = [sin(w1t), cos(w1t), …, sin(w d t), cos(w d t)]

[0099] wherein: Φ(t) is a time encoding function, wherein w1, … w d are learnable frequency parameters used to generate time embeddings.

[0100] By dynamically generating an adjacency matrix, calculating attention coefficients, and introducing improved activation functions and time encoding, the DGAT module accurately and dynamically models spatial dependency relationships: the dynamic adjacency matrix can adaptively capture the changing spatiotemporal correlation according to the real-time features of the nodes, effectively reflecting the instantaneous interaction of charging demand in space; the coefficient calculation and normalization based on the attention mechanism enable the model to focus on the importance of different neighbor nodes differently, giving priority to spatial nodes with strong influence, and enhancing the flexibility of the model expression; the improved activation function adjusts the slope in the negative number region, improving the model's fitting ability for complex nonlinear relationships; the introduction of time encoding embeds time information such as periodicity and trend into spatial relationships, achieving deep integration of spatiotemporal features. These mechanisms collectively ensure that the model can accurately depict the dependence of charging demand on time and space in the power grid network, thereby significantly improving the overall prediction accuracy and generalization ability of cluster charging behavior.

[0101] In a second aspect, the vehicle-network interaction demand service system provided by the present application comprises:

[0102] A business framework modeling module is configured to build a multi-agent vehicle-network interaction business framework, which includes three main agents, namely electric vehicle owners, power grid companies, and charging station operators, and the core of the business framework is a charging demand prediction objective function and multi-agent constraint conditions.

[0103] A multi-source heterogeneous data processing module is configured to collect and process multi-source heterogeneous data from the multi-agent based on the business framework, and output a standardized data set.

[0104] A user portrait construction module is configured to perform clustering analysis on the electric vehicle user data based on the standardized data set by using an improved BIRCH algorithm based on hierarchy and balance iteration reduction, and to construct a user portrait, wherein the improvement includes introducing a time decay factor and a streaming pruning mechanism.

[0105] A charging demand prediction module is configured to predict the charging demand of users based on the established user portrait by using a TCN-DGAT fusion algorithm, wherein the TCN-DGAT fusion algorithm includes a TCN module and a DGAT module, the TCN module is configured to perform time feature extraction, and the DGAT module is configured to perform dynamic spatial dependency modeling.

[0106] Preferably, the business framework construction module is specifically configured to construct an objective function, wherein the objective function is a minimization of charging demand prediction error:

[0107]

[0108] In the formula, is the error of the charging demand prediction, y k is the kth actual charging demand value, is the kth predicted charging demand value, and N is the sample quantity.

[0109] Preferably, the business framework construction module is specifically configured to construct a constraint condition, wherein the constraint condition includes:

[0110] Power grid operation constraint: total power supply of the power grid satisfies:

[0111]

[0112] Each charging pile voltage V m,t satisfies:

[0113] V min ≤V m,t ≤V max

[0114] In the formula, is the total power supply of the power grid, is the power of the mth charging pile, is the line loss, V m,t is the voltage of the charging pile m, V min and V max are the lower and upper limits of the voltage, respectively;

[0115] Charging pile constraint: actual charging power of the charging pile satisfies:

[0116]

[0117] wherein, is the actual charging power of charging pile m, is the rated charging power of charging pile m, is the maximum charging power acceptable by vehicle v;

[0118] User behavior constraint: the charging time of vehicle v satisfies:

[0119]

[0120] Post-charging SOC δ v satisfies:

[0121]

[0122] wherein, is the charging time of vehicle v, is the maximum charging time acceptable by vehicle v, δ v is the post-charging SOC of vehicle v, is the minimum post-charging SOC required by vehicle v.

[0123] Preferably, the multi-source heterogeneous data processing module is specifically configured to indicate that the multi-source heterogeneous data includes:

[0124] Electric vehicle owner side data: including structured data of real-time SOC, driving range, and charging request time of vehicle, and semi-structured data of user charging preference log and navigation destination information;

[0125] Power grid company side data: including structured data of real-time power supply power, node voltage, and line loss, and time series data of daily power fluctuation curve;

[0126] Charging station operator side data: including structured data of charging pile rated power, actual charging current, and equipment fault code, and semi-structured data of charging pile operation state log;

[0127] External auxiliary data: including structured data and time series data.

[0128] Preferably, the multi-source heterogeneous data processing module is specifically configured to assign a credibility weight ω cre to different data sources, and the expression is:

[0129]

[0130] wherein: Err(D i ) is the historical observation error of data source D i ;

[0131] Assign time series weights ω to different time series data t Its expression is:

[0132]

[0133] In the formula: Δt i =tt i For data source D i Data collection time t i The interval from the current processing time t; λ f This is the time decay factor;

[0134] By combining the credibility weight and the time series weight, the comprehensive weight ω is calculated:

[0135]

[0136] In the formula:

[0137] Weighted fusion of multi-source observations based on comprehensive weights:

[0138]

[0139] In the formula: X f (t) represents the target data value fused at time t, X i (t) represents the data source D. i The observation of the target data item at time t, ∈(t) is the fusion residual, satisfying ∈(t)~N(0,σ 2 ).

[0140] Preferably, the multi-source heterogeneous data processing module is specifically used to calculate the D of each data source. i Historical data sequence X i (1),X i (2),...,X i The mean μ of (n) i With standard deviation σ i :

[0141]

[0142] If a certain observed value X i (t) satisfies |X i (t)-μ i |>3σ i It was determined to be an outlier. Use the current fusion value X f (t) substitution:

[0143]

[0144] The modified standardized data set X'(t)={X1'(t), X2'(t), X3'(t), X4'(t)} is output.

[0145] Preferably, the user portrait construction module is specifically configured to introduce a time decay factor to optimize path selection, and an effective distance calculation formula is:

[0146] d e =d(x,c)×(1-λ·Δt)

[0147] In the formula, d e is an effective distance, d(x,c) is an original distance of the sample x and the cluster c, λ is a decay coefficient, controls the influence strength of time on the distance, and Δt is a time interval of cluster updating;

[0148] If the cluster radius after insertion exceeds the threshold r T , and the time sequence span in the cluster exceeds 24 hours, a splitting operation is forced to be performed;

[0149] For a dormant cluster that is not updated for more than 24 hours, if N < 100, the dormant cluster is directly deleted; and if N ≥ 100, the dormant cluster is down-weighted as a historical cluster;

[0150] A time weight ω is introduced for each CF node:

[0151]

[0152] In the formula, N' is the number of samples within 24 hours, is the time of the last update, is the current time

[0153] Periodically, low-activity branches are deleted, and the CF tree structure is rebalanced after pruning to ensure that the tree height does not exceed a preset value.

[0154] Preferably, the charging demand prediction module is specifically configured to use causal convolution to ensure that the output of a time step t only depends on the current and past time inputs, and a calculation formula is:

[0155]

[0156] In the formula, y t is the output of the time step t; w k is the kth weight of the convolution kernel; x t-k is the value of the input sequence at the time step t-k, and K is the size of the convolution kernel.

[0157] A dilated convolution is used to expand the receptive field, and a calculation formula is:

[0158]

[0159] In the formula: d is the expansion factor, that is, the interval sampling step;

[0160] The residual connection is introduced, and its calculation formula is:

[0161] H(x) = F(x) + W s ·x

[0162] In the formula: H(x) is the output of the residual connection, F(x) is a function composed of one or more convolution layers, W s Is a 1x1 convolution used to adjust the number of channels in the case of input and output dimension mismatch.

[0163] Preferably, the charging demand prediction module is specifically used for dynamically generating an adjacency matrix.

[0164]

[0165] In the formula: Is the feature of node i at time t-1; W a Is a learnable weight matrix, and u is a similarity score vector.

[0166] || is a vector concatenation operator;

[0167] The attention coefficient is calculated:

[0168]

[0169] In the formula: W q And W k Are the projection matrices of the query and the key, respectively, and is an attention parameter vector;

[0170] Theta is an improved activation function, and its expression is as follows:

[0171]

[0172] In the formula: alpha is a fixed positive number, used to control the slope of the negative number area;

[0173] The attention weight is normalized:

[0174]

[0175] In the formula: N i (t) is the neighbor set of node i at t;

[0176] The neighbor information is aggregated to update the node representation:

[0177]

[0178] In the formula: Is the information of the current state, and W va projection matrix for values;

[0179] introducing time encoding:

[0180] Φ(t) = [sin(w1t), cos(w1t), …, sin(w d t), cos(w d t)]

[0181] where Φ(t) is a time encoding function, where w1, … w d are learnable frequency parameters used to generate time embeddings.

[0182] In a third aspect, the present application also provides a computer device, comprising a memory and a processor, and the memory has stored thereon a computer program which can be loaded and executed by the processor to implement the multi-agent oriented vehicle network interaction demand service method.

[0183] In a fourth aspect, the present application also provides a computer readable storage medium having stored thereon a computer program which, when executed by a processor, implements the multi-agent oriented vehicle network interaction demand service method.

[0184] Advantages: Compared with the prior art, the present application has the following remarkable advantages: 1. By constructing a multi-agent business framework, processing multi-source heterogeneous data and establishing a precise user portrait and demand prediction model, the system can quickly respond to the demand changes of the power grid and users, dynamically adjust the scheduling strategy, thereby significantly improving the resource utilization efficiency and response speed; 2. By using a multi-source data weighted fusion method that combines credibility and time sequence weight, an outlier elimination mechanism, and combining an improved TCN-DGAT fusion algorithm for spatiotemporal feature extraction, the prediction error is effectively reduced, providing a reliable data basis for precise scheduling of the power grid and charging stations; 3. By introducing a time decay factor and a streaming pruning mechanism to improve the BIRCH clustering algorithm, the user portrait can be dynamically updated to more accurately reflect the spatiotemporal variation of user charging behavior, thereby supporting more personalized services and more flexible scheduling strategies; 4. The optimization model takes into account multiple constraints such as power grid operation, charging pile capacity and user behavior, while pursuing prediction accuracy and scheduling economy, ensuring the power supply safety of the power grid, the safe operation of equipment and the core charging experience of users. BRIEF DESCRIPTION OF DRAWINGS

[0185] Figure 1 is a method flowchart of the present application;

[0186] Figure 2 is a system framework schematic diagram of the present application;

[0187] Figure 3 is a system flowchart of the present application;

[0188] Figure 4 The vehicle-network interaction multi-agent and its relationship diagram of the application;

[0189] Figure 5 The user charging demand result diagram output by the system of the application. DETAILED DESCRIPTION

[0190] The technical solutions of the application are further described below with reference to the drawings.

[0191] The embodiment of the application provides a multi-agent-oriented vehicle-network interaction demand service method, as shown in the figure, comprising the following steps: Figure 1

[0192] S1: a multi-agent-oriented vehicle-network interaction service framework is proposed, and multiple service agents are adapted;

[0193] S2: a multi-source heterogeneous data processing module is established, different sources and different types of data are processed, and user portrait establishment and prediction are facilitated;

[0194] S3: an EV user portrait module is established, different categories of EV owners are clustered through an improved BIRCH clustering algorithm, and an EV user portrait is established;

[0195] S4: a user charging service demand prediction module is established, and the charging demand of the user is predicted through a TCN-DGAT algorithm.

[0196] The multi-agent-oriented vehicle-network interaction service framework in step S1 comprises:

[0197] The service demand framework is oriented to multiple agents and considers cooperation of multiple departments. The agents include three major agents of EV owners, power grid companies and charging station operators.

[0198] EV owners play an important role in vehicle-network interaction, and are both consumers of energy and providers of distributed energy storage. The biggest demand of EV owners is charging demand, and the charging opportunities provided by charging stations should be matched with the charging demand of EV owners.

[0199] The power grid company is a manager and service integrator of the power system, needs to provide power to users, and integrates EV resources to supply power to the power grid when needed. When supplying power, the power grid company needs to prevent the impact of large-scale disordered charging of EVs on the power grid and ensure stable operation of the power grid.

[0200] ​The charging station operator, as an infrastructure service provider and resource aggregator, mainly undertakes the role of a bridge and intermediary, communicates EV owners and power grid companies, and realizes the bidirectional flow of energy and information between EVs and the power grid. The main demand of the charging station operator is the return on investment, that is, the economy, and the profit obtained by the operator is improved through intelligent operation and improvement of user stickiness.

[0201] The core of the multi-agent oriented vehicle network interaction service framework is EV charging demand prediction, which involves the objective function and the constraint conditions of multiple agents as follows:

[0202] S1.1 Objective function

[0203]

[0204] In the formula, is the error of charging demand prediction, y k is the kth actual charging demand value, is the kth predicted charging demand value, and N is the sample number.

[0205] S1.2 Constraint conditions

[0206] The constraint conditions include power grid operation constraints, charging pile constraints and user behavior constraints, which correspond to the power grid company, the charging station operator and the EV owner respectively. The specific constraint conditions are as follows:

[0207] Power grid operation constraints

[0208] The total power supply power of the power grid needs to be balanced with the sum of the power of all charging piles and the line loss:

[0209]

[0210] In the formula, is the total power supply power of the power grid, is the power of the mth charging pile, is the line loss.

[0211] In order to prevent voltage drop or sudden rise from damaging equipment, the voltage needs to be kept within a certain range:

[0212] V min ≤V m,t ≤V max

[0213] In the formula, V m,t is the voltage of the charging pile m, V min and V max are the lower and upper limits of the voltage respectively.

[0214] Charging pile constraints

[0215] The actual charging power of the charging pile needs to be less than the rated power of the charging pile and the maximum acceptable charging power of the vehicle at the same time:

[0216]

[0217] In the formula, is the actual charging power of the charging pile m, is the rated charging power of the charging pile m, is the maximum acceptable charging power of the vehicle v.

[0218] User behavior constraints

[0219] The owner of each vehicle has certain requirements for the charging time:

[0220]

[0221] In the formula: is the charging time of the vehicle v, is the maximum acceptable charging time of the vehicle v.

[0222] The owner also has certain requirements for the amount of electricity after the vehicle is charged:

[0223]

[0224] In the formula: δ v is the SOC of the vehicle v after charging, is the minimum required SOC of the vehicle v after charging.

[0225] The multi-source heterogeneous data processing module in step S2 includes:

[0226] S2.1 Multi-source Heterogeneous Data Source Definition

[0227] The sources of multi-source heterogeneous data in the vehicle network interaction scenario are clarified, covering three main bodies and external environment, and the specific classification is as follows:

[0228] EV owner side data D1: structured data (vehicle real-time SOC, driving range, charging request time), semi-structured data (user charging preference log, navigation destination information);

[0229] Power grid company side data D2: structured data (real-time power supply power, node voltage, line loss), time series data (intra-day power fluctuation curve);

[0230] Charging station operator side data D3: structured data (charging pile rated power, actual charging current, equipment fault code), semi-structured data (charging pile operation state log);

[0231] External auxiliary data D4: structured data, time series data.

[0232] S2.2 Data heterogeneity quantification and weight allocation

[0233] To quantify the data validity, data source credibility weight and time weight are introduced according to the credibility difference of different data sources:

[0234] Credibility weight ω cre : preset based on historical data error rate, reflecting the inherent reliability of data source

[0235]

[0236] In the formula: Err(D i ) is the historical observation error of data source D i .

[0237] Time weight ω t : considering the timeliness of data, time decay factor λ f is introduced.

[0238]

[0239] In the formula: Δt i = t-t i is the interval between data collection time t i of data source D i and the current processing time t; λ f takes the value range of 0.05-0.2.

[0240] Comprehensive weight ω, fusion credibility and time weight:

[0241]

[0242] In the formula: , to ensure the rationality of weight fusion.

[0243] S2.3 Multi-source data weighted fusion

[0244] For the same target data item (such as the effective power supply capacity of charging pile in a certain area at t time), multi-source observation values are fused:

[0245]

[0246] In the formula: X f (t) is the target data value fused at t time, X i (t) is the observation value of data source D i to the target data item at t time. ∈(t) is the fusion residual, satisfying ∈(t) ~ N(0, σ 2 ).

[0247] Based on the 3σ criterion to eliminate abnormal data before fusion, to avoid interference with subsequent modules:

[0248] 1) Calculate the historical data sequence X of each data source D_i i (1), X i (2),..., X i (n) mean μ i and standard deviation σ i :

[0249]

[0250] 2) If an observation value X i (t) satisfies |X i (t)-μ i |>3σ i , it is determined as an outlier Replace with the current time fusion value X f (t):

[0251]

[0252] 3) Output the corrected standardized data set X'(t) = {X1'(t), X2'(t), X3'(t), X4'(t)}.

[0253] The EV user portrait module in step S3 includes:

[0254] Before predicting the charging demand, it is necessary to first characterize the EV user portrait and clean and classify the user data. Clustering algorithm is one of the core tools for building user portrait, especially suitable for automatically discovering hidden features of user groups from massive data and grouping based on behavior or attribute similarity.

[0255] The present patent adopts the balanced iterative reduction and clustering using hierarchies (BIRCH) algorithm to cluster data.

[0256] S3.1 BIRCH algorithm steps

[0257] BIRCH algorithm is a high-efficiency hierarchical clustering algorithm specially designed for processing large-scale data sets. Its core idea is to compress data information by constructing a compact clustering feature tree, thereby greatly reducing the computational complexity while maintaining the quality of clustering.

[0258] The core of BIRCH is the Clustering Feature (CF), a summary of each sub-cluster containing three key statistics: the number of data points, the linear sum of all data, and the squared sum of all data points. With this information, BIRCH can compute the size, density, and distance of clusters without storing the original data.

[0259] The CF tree is a highly balanced tree structure composed of multiple layers of nodes (leaves and non-leaves). Each node contains multiple CF entries representing the summary information of its sub-clusters. The construction process maintains memory efficiency by dynamically adjusting branches and sub-clusters.

[0260] The traditional BIRCH algorithm starts by constructing the CF tree, inserting samples one by one from an empty tree. Starting from the root node, the nearest child node is selected based on the distance metric, and the process continues until the leaf node is located. Then, merging nodes is attempted by checking if there are any CFs in the leaf node that can be merged. After merging, the CF summary information of the parent nodes is updated from the leaf node upwards to ensure the balance of the tree.

[0261] The steps of the BIRCH algorithm are as follows:

[0262] Constructing the CF tree. First, data scanning and sub-cluster insertion are performed, reading data points one by one and inserting them into the CF tree. The algorithm determines whether a new point can be merged into an existing sub-cluster based on the distance threshold. If it can, the statistics of the corresponding CF are updated; otherwise, a new sub-cluster is created.

[0263] Dynamic adjustment of the CF tree. When the number of sub-clusters of a node exceeds the preset capacity, the node is split into two new nodes, and the sub-clusters are redistributed to maintain the balance of the tree. By adjusting the threshold parameter, the maximum radius (looseness) of the sub-clusters is controlled. If a new point causes the sub-cluster radius to exceed the threshold, splitting or creating a new sub-cluster is triggered.

[0264] Compressing the CF tree. If the initially constructed CF tree is too large, the data can be rescanned by increasing the threshold to merge similar sub-clusters, further reducing the size of the tree.

[0265] Global clustering. The traditional clustering algorithm is applied to the leaf nodes of the CF tree (i.e., the final set of sub-clusters) to convert the summary information into the final clustering result. This step addresses the problem of excessive subdivision of CF trees due to improper threshold settings.

[0266] Outlier handling and result optimization. Sub-clusters containing very few data points are marked as outliers, and then the outliers are redistributed to the nearest cluster or the cluster boundaries are adjusted through additional scanning to improve clustering quality.

[0267] S3.2 Improved BIRCH algorithm

[0268] Time-sensitive incremental insertion

[0269] The BIRCH algorithm is improved for path selection, and a time decay factor is introduced.

[0270] d e =d(x,c)×(1-λ·Δt)

[0271] In the formula: d e is the effective distance, d(x,c) is the original distance of sample x and cluster c. λ is the decay coefficient, which controls the influence strength of time on distance, and Δt is the time interval of cluster update.

[0272] If the cluster radius exceeds the threshold r T after insertion, and the time span in the cluster exceeds 24 hours, forced splitting is performed to prevent data from being too old.

[0273] Streaming pruning

[0274] For clusters that have not been updated for more than 24 hours (i.e. "dormant clusters"), if N < 100, they are directly deleted; if N ≥ 100, they are down-weighted to "history clusters" (only participate in merging, do not trigger splitting).

[0275] A time weight ω

[0276]

[0277] In the formula: N' is the number of samples within 24 hours. is the time of the last update, is the current time.

[0278] Every certain time, delete low-activity branches with ω < ω min , prune and rebalance the CF tree structure to ensure that the tree height does not exceed the preset value.

[0279] The user charging service demand prediction module in step S4 includes:

[0280] The user charging demand prediction module is the core module of the system, which is used to predict the charging demand of EV users. This module mainly consists of TCN and DGAT, and performs spatio-temporal prediction on user charging demand.

[0281] S4.1 Upper prediction model based on TCN

[0282] TCN is an improved network model based on CNN, which has a unique extended causal convolution structure and is more suitable for solving time series problems. The extended convolution allows the input of the previous layer to be used for expansion, extracting features from longer time interval data. The causal convolution ensures the causal relationship of the extracted features, that is, the output y t at time t terminates the input only before time t.

[0283] Causal Convolution

[0284] Causal convolution ensures that the output only depends on the current and past inputs, avoiding future information leakage. The formula is as follows:

[0285]

[0286] where y t is the output at time step t; w k is the kth weight of the convolution kernel; x t-k is the value of the input sequence at time step t-k, and K is the size of the convolution kernel.

[0287] Dilated Convolution

[0288] To increase the receptive field without increasing the computational load, TCN introduces dilated convolution, which expands the receptive field by skipping some positions through a dilation factor.

[0289]

[0290] where d is the dilation factor, i.e., the interval sampling step.

[0291] Residual Connection

[0292] To avoid gradient vanishing and improve the training efficiency of the model, residual connection is added in TCN. The output of each layer not only contains the result of convolution calculation, but also contains the weighted result of the input signal. This allows the network to be trained more easily and deeper.

[0293] H(x) = F(x) + W s ·x

[0294] where H(x) is the output of the residual connection, F(x) is a function composed of one or more convolution layers, W s is a 1x1 convolution used to adjust the number of channels when the input and output dimensions do not match.

[0295] S4.2 Lower-level prediction model based on DGAT

[0296] Dynamic Graph Attention Network (DGAT) is a graph neural network that combines dynamic graph structure and attention mechanism, aiming to process graph data that changes over time (such as social network evolution, traffic flow changes). Its core lies in dynamically capturing the changes of graph structure and adaptively adjusting the importance weights between nodes through attention mechanism. Unlike static GAT, DGAT models the evolution of node features and the dynamics of adjacent relationships in the time dimension.

[0297] The edge weights or connection states of dynamic graphs change over time, and the DGAT can generate an adjacency matrix by the following way:

[0298]

[0299] where: is the feature of node i at time t-1; W a is the learnable weight matrix, and u is the similarity score vector. || is the vector concatenation operator symbol.

[0300] The DGAT captures the dynamic dependencies between nodes at time step t through attention coefficients:

[0301]

[0302] where: W q and W k are the projection matrices for query and key, respectively, and is the attention parameter vector. To improve the activation function, its expression is as follows:

[0303]

[0304] where: a is a very small fixed positive number used to control the slope in the negative number area, usually set to 0.01.

[0305] The attention weights are normalized:

[0306]

[0307] where: N i (t) is the neighbor set of node i at t (defined by the dynamic adjacency matrix).

[0308] The node representation is updated in combination with the dynamic attention weight, and the neighbor information is aggregated:

[0309]

[0310] where: is the information of the current state, W v is the value projection matrix.

[0311] In order to capture the time dependency, time encoding is introduced:

[0312] Φ(t) = [sin(w1t), cos(w1t), …, sin(w d t), cos(w d t)]

[0313] where: Φ(t) is the time encoding function, where w1, … w d are learnable frequency parameters used to generate time embeddings.

[0314] Based on the similar inventive concept, the embodiment of the present application also provides a metrology data dynamic partitioning system corresponding to the metrology data dynamic partitioning method, comprising:

[0315] A business framework modeling module is configured to model a multi-agent vehicle-network interaction business framework, the business framework including three main agents of electric vehicle owners, power grid companies and charging station operators, and the core of the business framework is a charging demand prediction objective function and multi-agent constraint conditions;

[0316] A multi-source heterogeneous data processing module is configured to collect and process multi-source heterogeneous data from the multi-agent based on the business framework, and output a standardized data set;

[0317] A user portrait construction module is configured to perform clustering analysis on electric vehicle user data based on the standardized data set by using an improved hierarchical balanced iterative reduction clustering (BIRCH) algorithm, and construct a user portrait, the improvement including introducing a time decay factor and a streaming pruning mechanism;

[0318] A charging demand prediction module is configured to predict user charging demand based on the established user portrait by using a time convolution network-double graph attention (TCN-DGAT) fusion algorithm, the TCN-DGAT fusion algorithm including a TCN module and a DGAT module, wherein the TCN module is configured to extract time features, and the DGAT module is configured to model dynamic spatial dependency.

[0319] Further, the business framework modeling module is specifically configured to construct an objective function, the objective function being a minimization of charging demand prediction error:

[0320]

[0321] In the formula, is the error of the charging demand prediction, y k is the kth actual charging demand value, is the kth predicted charging demand value, and N is the sample quantity.

[0322] Further, the business framework modeling module is specifically configured to construct constraint conditions, the constraint conditions including:

[0323] Power grid operation constraint: total power supply of the power grid satisfies:

[0324]

[0325] Each charging pile voltage V m,t satisfies:

[0326] V min ≤ Vm,t ≤V max

[0327] In the formula, It is the total power supplied by the power grid. That is the power of the m-th charging station. It is line loss, V m,t It is the voltage of the charging pile m, V min and V max These are the lower and upper voltage limits, respectively.

[0328] Charging station constraints: Actual charging power of the charging station satisfy:

[0329]

[0330] In the formula, This is the actual charging power of charging pile m. This is the rated charging power of charging pile m. This is the maximum charging power that vehicle v can accept;

[0331] User behavior constraints: Vehicle charging time must meet the following requirements:

[0332]

[0333] SOCδ after charging v satisfy:

[0334]

[0335] In the formula: The charging time for vehicle v. δ is the maximum charging time that vehicle v can accept. v The state of charge (SOC) of the vehicle (v) after charging. The minimum SOC required for vehicle v after charging.

[0336] Furthermore, the multi-source heterogeneous data processing module is specifically used to describe that the multi-source heterogeneous data includes:

[0337] Electric vehicle owner-side data includes structured data such as real-time vehicle SOC, mileage, and charging request time, as well as semi-structured data such as user charging preference logs and navigation destination information.

[0338] Data from the power grid company side includes structured data on real-time power supply, node voltage, and line losses, as well as time-series data on intraday power fluctuation curves;

[0339] Data from the charging station operator side includes structured data such as the rated power of the charging pile, the actual charging current, and the equipment fault codes, as well as semi-structured data from the charging pile operation status log.

[0340] External auxiliary data: including structured data and time series data.

[0341] Furthermore, the multi-source heterogeneous data processing module is specifically used to assign confidence weights ω to different data sources. cre Its expression is:

[0342]

[0343] In the formula: Err(D i (D) is the data source. i Historical observation errors;

[0344] Assign time series weights ω to different time series data t Its expression is:

[0345]

[0346] In the formula: Δt i =tt i For data source D i Data collection time t i The interval from the current processing time t; λ f This is the time decay factor;

[0347] By combining the credibility weight and the time series weight, the comprehensive weight ω is calculated:

[0348]

[0349] In the formula:

[0350] Weighted fusion of multi-source observations based on comprehensive weights:

[0351]

[0352] In the formula: X f (t) represents the target data value fused at time t, X i (t) represents the data source D. i The observation of the target data item at time t, ∈(t) is the fusion residual, satisfying ∈(t)~N(0,σ 2 ).

[0353] Furthermore, the multi-source heterogeneous data processing module is specifically used to calculate D from each data source. i Historical data sequence X i (1),X i (2),...,Xi Mean of (n) i With standard deviation i :

[0354]

[0355]

[0356] If an observation value X i (t) satisfies |X i (t)-μ i |>3σ i , it is determined as an outlier Replace it with the fusion value X f (t) at the current time:

[0357]

[0358] Output the corrected standardized data set X'(t) = {X1'(t), X2'(t), X3'(t), X4'(t)}.

[0359] Further, the user portrait construction module is specifically configured to introduce a time decay factor to optimize path selection, and the effective distance calculation formula is:

[0360] d e = d(x,c) x (1-λ·Δt)

[0361] In the formula: d e is the effective distance, d(x,c) is the original distance between sample x and cluster c, λ is the decay coefficient, which controls the influence strength of time on distance, and Δt is the time interval of cluster update;

[0362] If the cluster radius exceeds the threshold r T after insertion, and the time sequence span in the cluster exceeds 24 hours, a splitting operation is forced to be performed;

[0363] For a dormant cluster that has not been updated for more than 24 hours, if N < 100, it is directly deleted; if N ≥ 100, it is down-weighted as a historical cluster;

[0364] A time weight ω is introduced for each CF node:

[0365]

[0366] In the formula: N' is the number of samples within 24 hours, is the time of the last update, is the current time

[0367] Periodically delete low-activity branches, and rebalance the CF tree structure after pruning to ensure that the tree height does not exceed the preset value.

[0368] Further, the charging demand prediction module is specifically configured to use causal convolution to ensure that the output of time step t only depends on the current and past time inputs, and the calculation formula is as follows:

[0369]

[0370] In the formula, y t is the output of time step t; w k is the kth weight of the convolution kernel; x t-k is the value of the input sequence at time step t-k, and K is the size of the convolution kernel;

[0371] The receptive field is expanded by using dilated convolution, and the calculation formula is as follows:

[0372]

[0373] In the formula, d is a dilated factor, that is, an interval sampling step;

[0374] The residual connection is introduced, and the calculation formula is as follows:

[0375] H(x) = F(x) + W s ·x

[0376] In the formula, H(x) is the output of the residual connection, F(x) is a function composed of one or more convolution layers, W s is a 1x1 convolution used to adjust the number of channels when the input and output dimensions do not match.

[0377] Further, the charging demand prediction module is specifically configured to dynamically generate an adjacency matrix:

[0378]

[0379] In the formula, is the feature of node i at time t-1; W a is a learnable weight matrix, and u is a similarity score vector. || is a vector concatenation operator symbol;

[0380] The attention coefficient is calculated as follows:

[0381]

[0382] In the formula, W q and W k are projection matrices of the query and the key, respectively, and is an attention parameter vector;

[0383] θ is an improved activation function, and its expression is as follows:

[0384]

[0385] where: a is a fixed positive number to control the slope of the negative region;

[0386] Normalizing attention weights:

[0387]

[0388] where: N i (t) is the neighbor set of node i at time t;

[0389] Aggregate neighbor information updates node representation:

[0390]

[0391] where: is the information of the current state, W v is the projection matrix of the value;

[0392] Introducing time encoding:

[0393] Φ(t) = [sin(w1t), cos(w1t),..., sin(w d t), cos(w d t)]

[0394] where: Φ(t) is the time encoding function, where w1,...w d are learnable frequency parameters to generate time embeddings.

[0395] Figure 2 is the overall architecture diagram of the vehicle network interaction demand service system. The core is the EV user portrait module and the charging demand prediction module. The EV user portrait module is driven by the BIRCH clustering algorithm, clusters the input EV user data information, and outputs the EV user portrait. The core of the user charging demand prediction module is the TCN and DGAT double-layer prediction model, which processes the input data and outputs the user charging demand prediction result. Figure 3 is the flow chart of the vehicle network interaction demand service system.

[0396] Figure 4 is the multi-agent involved in the vehicle network interaction demand service system, including EV owners, power grid companies and charging station operators. The owner and the power grid company: the owner obtains income by responding to the demand of the power grid, and the power grid company reduces the operating cost by scheduling EV resources. The owner and the operator of the charging station: the owner participates in the vehicle network interaction depending on the infrastructure of the operator, and the operator expands market influence through user scale. The power grid company and the operator: the power grid company relies on the operator to aggregate resources, and the operator obtains a share or dispatching fee by serving the demand of the power grid.

[0397] Figure 5The result graph generated after the charging demand prediction of the vehicle-network interaction demand service system shows that the peak of user charging demand concentrates at noon (about 12 o'clock) and in the evening (about 20 o'clock).

[0398] The application further discloses an electronic device.

[0399] Specifically, the electronic device can be a computer device such as a desktop computer, a notebook computer, a palm computer and a cloud server. The computer device can include but is not limited to a processor and a memory. The processor and the memory can be connected through a bus or other means. The processor can be a central processing unit (CPU). The processor can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, graphics processing units (GPU), embedded neural network processing units (NPU) or other dedicated deep learning coprocessors, discrete gate or transistor logic devices, discrete hardware components, etc. chips, or combinations of the above chips.

[0400] The memory, as a non-transitory computer readable storage medium, can be used to store non-transitory software programs, non-transitory computer executable programs and modules. The processor executes various functions and data processing of the processor by running the non-transitory software programs, instructions and modules stored in the memory. The memory can include a program storage area and a data storage area, wherein the program storage area can store application programs required by the control unit and at least one function; and the data storage area can store data created by the processor and the like. In addition, the memory can include a high-speed random access memory, and can also include a non-transitory memory. In some embodiments, the memory can optionally include a memory remotely arranged relative to the processor, and these remote memories can be connected to the processor through a network. Examples of the above network include but are not limited to the Internet, an intranet, a local area network, a mobile communication network and combinations thereof.

[0401] The application further discloses a computer readable storage medium.

[0402] Specifically, the computer readable storage medium is used to store a computer program, and the computer program is executed by the processor to realize the method in the above method embodiments.

[0403] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments of the method of the present application can be completed by instructing the relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the above-mentioned embodiments of the method. The storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), a random access memory (RAM), a flash memory, a hard disk drive (HDD) or a solid-state drive (SSD), etc. The storage medium can also include a combination of the above-mentioned types of memories.

Claims

1. A multi-agent oriented vehicle-network interaction demand service method, characterized in that, Comprise: Construct a multi-agent oriented vehicle-grid interaction business framework, which includes three major agents: electric vehicle owners, power grid companies and charging station operators; Based on the business framework, collect and process multi-source heterogeneous data from the multi-agent, and output standardized data sets; Based on the standardized data set, the improved BIRCH algorithm based on hierarchical balanced iterative reduction clustering is used for clustering analysis of electric vehicle user data to construct user portraits, and the improvement includes the introduction of time decay factor and streaming pruning mechanism; Based on the established user portrait, the TCN-DGAT fusion algorithm is used to predict the user's charging demand, and the TCN-DGAT fusion algorithm includes TCN module and DGAT module, wherein the TCN module is used for time feature extraction, and the DGAT module is used for dynamic spatial dependence modeling.

2. The vehicle-to-grid interaction demand service method according to claim 1, characterized by, The objective function corresponding to the business framework is to minimize the charging demand prediction error: wherein is the error of the charge demand prediction, y k is the kth actual charge demand value, is the kth predicted charge demand value, and N is the number of samples.

3. The vehicle-to-grid interaction demand service method of claim 1, wherein, The constraint conditions corresponding to the business framework include: Grid operation constraints: total power supplied by the grid P t gird satisfied: Each charging pile voltage V m,t Satisfies: V min ≤V m,t ≤V max where P t gird is the total power supplied by the grid, is the power of the mth charging station, P t loss is the line loss, V m,t is the voltage of the charging station m, V min and V max are the lower and upper voltage limits, respectively; Charging pile constraint: charging pile actual charging power Satisfies: wherein is the actual charging power of the charging station m, is the rated charging power of the charging station m, is the maximum charging power acceptable for the vehicle v; User behavior constraints: the vehicle charging time satisfies: Post-charge SOC delta v Satisfies: wherein: is the charging time of the vehicle v, is the maximum charging time acceptable for the vehicle v, δ v is the SOC of the vehicle v after charging, is the minimum SOC required for the vehicle v after charging.

4. The vehicle-to-grid interaction demand service method according to claim 1, characterized by, The multi-source heterogeneous data includes: Electric vehicle owner side data: including structured data of vehicle real-time SOC, driving mileage, charging request time, and semi-structured data of user charging preference log and navigation destination information; Power grid company side data: including structured data of real-time power supply power, node voltage and line loss, and time series data of daily power fluctuation curve; Charging station operator side data: including structured data of charging pile rated power, actual charging current and equipment fault code, and semi-structured data of charging pile operation state log; External auxiliary data: including structured data and time series data.

5. The vehicle-to-grid interaction demand service method of claim 1, wherein, The processing of multi-source heterogeneous data from the multi-agent includes credibility weight and time sequence weight allocation, weighted fusion: Assigning a credibility weight ω to different data sources cre The expression is: where: Err(D i ) is the historical observation error for data source D i . Assigning timing weights ω for different timing data t whose expression is where: Δt i = t - t i is the interval between the data collection time t i of the data source D i and the current processing time t; λ f is the time decay factor; Fusion of credibility weight and time sequence weight, calculation of comprehensive weight ω: In the formulae: Based on the comprehensive weight, the multi-source observation value is weighted and fused: where: X f (t) is the fused target data value at time t, X i (t) is the data source D i the observation of the target data item at time t, ∈(t) is the fusion residual, satisfying ∈(t) ~ N(0, σ 2 ).

6. The vehicle-to-grid interaction demand service method according to claim 1, characterized by, The processing of multi-source heterogeneous data from the multi-agent also includes outlier rejection: Computing the data sources D i the historical data sequence X i (1), X i (2),..., X i the mean μ of the data sources (1), X i and the standard deviation σ i : If an observation value X i (t) satisfies |X i (t)-μ i |>3σ i , it is determined as an abnormal value Replace it with the fusion value X f (t) at the current time Output the corrected standardized data set X'(t)={X1'(t),X2'(t),X3'(t),X4'(t)}.

7. The vehicle-to-grid interaction demand service method of claim 1, wherein, The improvement in the improved BIRCH algorithm based on hierarchical balanced iterative reduction clustering includes: Introduce time decay factor to optimize path selection, and the effective distance calculation formula is: d e = d(x, c) x (1 - l · At) where d e is the effective distance, d(x, c) is the original distance of sample x and cluster c, λ is the decay coefficient, which controls the strength of the influence of time on distance, and Δt is the time interval of cluster update. If the radius of the inserted cluster exceeds the threshold r T and the timing span within the cluster exceeds 24 hours, a split operation is forced. The improvement in the improved BIRCH algorithm based on hierarchical balanced iterative reduction clustering also includes streaming pruning mechanism: For dormant clusters that have not been updated for more than 24 hours, if N<100, delete directly; if N≥100, reduce the weight to historical cluster; Introduce a time weight ω for each CF node: where: N' is the number of samples over 24 hours, is the time of the last update, is the current time Periodically delete low activity branches, and rebalance the CF tree structure after pruning to ensure that the tree height does not exceed the preset value.

8. The vehicle-to-grid interaction demand service method of claim 1, wherein, The TCN module is used for time feature extraction, including: Using causal convolution to ensure that the output of time step t only depends on the input of current and past time, and its calculation formula is: where: y t is the output at time step t; w k is the kth weight of the convolution kernel; x t-k is the value of the input sequence at time step t-k, and K is the size of the convolution kernel. Using dilated convolution to expand the receptive field, and its calculation formula is: In the formula: d is the dilation factor, i.e. interval sampling step; The residual connection is introduced, and its calculation formula is: H(x) = F(x) + W s • x where H(x) is the output of the residual connection, F(x) is a function consisting of one or more convolutional layers, W s is a 1x1 convolution used to adjust the number of channels in case of input-output dimension mismatch.

9. The vehicle-to-grid interaction demand service method of claim 1, wherein, The DGAT module is used for dynamic spatial dependence modeling, which includes: The adjacency matrix is dynamically generated: ; where: is the feature of node i at time t - 1; is a learnable weight matrix, u is a similarity score vector. || is a vector concatenation operator symbol; The attention coefficient is calculated: ; where: and are the projection matrices for queries and keys, respectively, and are attention parameter vectors. To improve the activation function, its expression is as follows: ; where: is a fixed positive number used to control the slope of the negative region; The attention weight is normalized: ; where: is the set of neighbors of node i at time t; The neighbor information is aggregated to update the node representation: ; where: is information of the current state, is a projection matrix of the value; The time encoding is introduced: ; where: is a time encoding function, where is a learnable frequency parameter used to generate the time embedding. 10.A multi-agent oriented vehicle-network interaction demand service system, characterized in that, It includes: A business framework modeling module is configured to build a multi-agent vehicle network interaction business framework, which includes three main agents, namely, electric vehicle owners, power grid companies and charging station operators, and the core is a charging demand prediction objective function and multi-agent constraint conditions; A multi-source heterogeneous data processing module is configured to collect and process multi-source heterogeneous data from the multi-agent based on the business framework, and output a standardized data set; A user portrait construction module is configured to perform clustering analysis on electric vehicle user data based on the standardized data set by using an improved hierarchical balanced iterative reducing clustering (BIRCH) algorithm, and to construct a user portrait, the improvement including introducing a time decay factor and a streaming pruning mechanism; A charging demand prediction module is configured to predict user charging demand based on the established user portrait by using a time convolution network-double graph attention (TCN-DGAT) fusion algorithm, the TCN-DGAT fusion algorithm including a TCN module and a DGAT module, wherein the TCN module is used for time feature extraction, and the DGAT module is used for dynamic spatial dependence modeling.

Images