Method and system for collaborative capacity evaluation of heterogeneous data of electric vehicle charging station

Abstract

The present application relates to the technical field of power system, especially to a heterogeneous data collaborative capacity evaluation method and system for electric vehicle charging stations, the method comprising: constructing a membership function model based on user charging behavior influencing factors and an original charging behavior database; constructing a charging behavior data generation model considering user behavior preferences based on the membership function model and according to charging behavior constraints; updating the data in the original charging behavior database through the charging behavior data generation model to obtain a sample charging behavior database and perform a data preprocessing operation; constructing and training an MF-LSTM charging load prediction model according to the membership function model and the preprocessed sample charging behavior database; and predicting the concentrated charging load of electric vehicles by using the trained MF-LSTM model, quantitatively analyzing the adjustable potential of the concentrated charging load of electric vehicles, and realizing accurate evaluation of the adjustable capacity of charging stations.

Description

Technical Field

[0001] This invention relates to the field of power system technology, and in particular to a method and system for assessing the capacity of heterogeneous data collaboration in electric vehicle charging stations. Background Technology

[0002] With the widespread adoption of electric vehicles, centralized charging stations have become a core adjustable resource on the demand side of the power grid. However, charging station data encompasses heterogeneous information from multiple sources, including charging pile operating status and power grid dispatch instructions, resulting in issues such as disorganized data formats and inconsistent time granularity, making unified modeling and collaborative analysis difficult. Furthermore, influenced by factors such as the uncertainty of user charging behavior and the widespread adoption of fast charging technology, charging load exhibits strong fluctuations, intermittency, and nonlinear characteristics, posing significant challenges to accurate capacity assessment and load regulation of charging stations.

[0003] Traditional assessment methods typically employ statistical analysis and shallow neural networks for prediction, which generally suffer from insufficient accuracy in capacity assessment and delayed control response. They also struggle to capture the long-term time-series dependence and complex nonlinear variation patterns of charging load, resulting in poor load forecasting performance. To address the shortcomings of traditional methods, current research introduces long short-term memory neural networks to optimize the time-series prediction capability of charging load, improving load forecasting accuracy to some extent. Simultaneously, preliminary modeling studies have been conducted on user willingness to participate in charging control, attempting to quantify the impact of user behavior on the adjustable potential of charging. However, existing technical solutions still face many unresolved key issues.

[0004] Existing long short-term memory neural network models rely solely on historical load data for prediction, resulting in insufficient generalization ability and scenario adaptability. Furthermore, user behavior modeling methods are relatively crude, often subjectively setting the proportion of user participation in regulation, leading to significant deviations between evaluation results and actual operating conditions. In addition, existing research largely focuses on theoretical algorithm simulation verification, failing to directly support real-time regulation of virtual power plants and dynamic grid dispatch decisions, severely restricting the efficient utilization of adjustable resources at charging stations.

[0005] The information disclosed in this background section is intended only to enhance the understanding of the overall background of the present invention and should not be construed as an admission or in any way implying that the information constitutes prior art known to those skilled in the art. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a method and system for collaborative capacity assessment of heterogeneous data of electric vehicle charging stations. Through heterogeneous data standardization, fuzzy modeling of user behavior, MF-LSTM model construction and dynamic capacity output, the adjustable capacity of charging stations can be accurately assessed.

[0007] To achieve the above objectives, the technical solution adopted by this invention is: a method for assessing the collaborative capacity of heterogeneous data in electric vehicle charging stations, comprising the following steps: Construct a membership function model based on factors influencing user charging behavior and a database of original charging behaviors; Based on the membership function model, a charging behavior data generation model that takes into account user behavior preferences is constructed according to the charging behavior constraints. The charging behavior data generation model is used to update the data in the original charging behavior database to obtain a sample charging behavior database and perform data preprocessing operations. Based on the membership function model and the preprocessed sample charging behavior database, an MF-LSTM charging load prediction model is constructed and trained. The trained MF-LSTM model is used to predict the centralized charging load of electric vehicles and to quantitatively analyze the adjustable potential of the centralized charging load of electric vehicles.

[0008] Furthermore, the factors influencing user charging behavior include internal and external factors. The internal factors include users' sensitivity to electricity prices and their level of anxiety about electricity consumption. The external factors include electricity price compensation and the initial state of charge of electric vehicles.

[0009] Furthermore, the construction of the membership function model based on factors influencing user charging behavior includes the following steps: The factors influencing user charging behavior are digitized and fuzzified to form values ​​in the range [0, 10]. Construct the output rules of the membership function model, and determine whether a user participates in power grid regulation based on the output rules.

[0010] Furthermore, the membership function model for electricity price compensation includes three levels: low, medium, and high. The membership functions for the initial state of charge and the degree of electricity anxiety both adopt the form of Gaussian functions. The membership function for the degree of electricity price sensitivity is divided into three levels: insensitive, relatively sensitive, and highly sensitive.

[0011] Furthermore, the step of constructing the output rules of the membership function model and determining whether a user participates in power grid regulation based on the output rules includes the following steps: Construct an output rule system comprising at least six decision rules, and then construct the output rule system into a rule matrix. Based on the determination rules, it is determined whether the user participates in power grid regulation, and based on the output rules, it is determined whether the user's participation in power grid regulation is output as an output result.

[0012] Furthermore, the charging behavior constraints include charging duration constraints, vehicle parking duration constraints, state of charge and charging power constraints, and electricity price impact coefficient constraints.

[0013] Furthermore, the process of updating the data in the original charging behavior database using the charging behavior data generation model includes the following steps: Load the original charging database and filter based on charging duration constraints; The membership function model is used to calculate the probability of each user participating in the demand response; For vehicles participating in power grid regulation, the vehicle parking time constraint, the state of charge and charging power constraint are applied. By constraining users' willingness to participate in regulation through the electricity price impact coefficient, sample charging behavior data is generated. Repeat the above steps until the data in the original charging database is updated to the sample charging behavior data.

[0014] Furthermore, the data preprocessing operations include anomaly processing, data sample set expansion, data standardization, and training set and test set partitioning.

[0015] Furthermore, the step of constructing and training the MF-LSTM charging load prediction model based on the membership function model and the preprocessed sample charging behavior database includes the following steps: Data fusion processing is performed on the membership function feature vector obtained from the membership function model and the charging load time series data in the sample charging behavior database to obtain input data; Configure the model parameters of the LSTM neural network and input the input data into the configured LSTM neural network model. The LSTM neural network includes a forget gate, an input gate, an output gate, and cell states. The MF-LSTM charging load prediction model is obtained by performing model training.

[0016] Furthermore, the evaluation metrics for model training include root mean square error (RMSE) and coefficient of determination. : ; ; In the formula, These are the model's predicted values. This represents the actual validation value, where n is the total number of data points; the smaller the RMSE, the better. The closer the value is to 1, the higher the model's prediction accuracy.

[0017] Furthermore, the average load change value is used. Average load offset percentage The adjustable time period and adjustable continuity quantitative analysis of the adjustable potential of the centralized charging load of electric vehicles, namely the average load variation value. The expression is: ; The average load offset percentage The expression is: ; In the formula, P is the load value without considering user preferences. The load value is taken into account user preferences, where T is the number of hours in the time period.

[0018] Furthermore, it also includes analyzing the impact of user behavior preferences on load adjustability potential under special scenarios, including: weekdays or holidays, summer or winter, residential areas or commercial areas.

[0019] This invention also provides a heterogeneous data collaborative capacity assessment system for electric vehicle charging stations, comprising: The function model building module is used to build a membership function model based on the factors influencing user charging behavior and a database of original charging behavior. The database generation and processing module is used to construct a charging behavior data generation model that takes into account user behavior preferences based on the membership function model and charging behavior constraints. The charging behavior data generation model is used to update the data in the original charging behavior database to obtain a sample charging behavior database and perform data preprocessing operations. The model building and training module constructs and trains the MF-LSTM charging load prediction model based on the membership function model and the preprocessed sample charging behavior database. The adjustable potential assessment module is used to predict the concentrated charging load of electric vehicles using the trained MF-LSTM model and to quantitatively analyze the adjustable potential of the concentrated charging load of electric vehicles.

[0020] Furthermore, the adjustable potential assessment module also includes a scenario analysis submodule, which is used to analyze the degree of influence of user behavior preferences on load adjustable potential under special scenarios, and output the load adjustable capacity, adjustable time period and adjustable sustainability index under each special scenario.

[0021] The beneficial effects of this invention are as follows: By introducing the membership function into the electric vehicle charging load prediction process, this invention integrates user behavior feature vectors with charging load time series data as model input. By combining the advantages of LSTM neural networks in processing time series data with the fuzzy reasoning ability of membership functions, the MF-LSTM prediction model can more accurately reflect the spatiotemporal variation of charging load and effectively improve the accuracy of charging load prediction. Attached Figure Description

[0022] 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 recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart illustrating the heterogeneous data collaborative capacity assessment method for electric vehicle charging stations in an embodiment of the present invention. Figure 2 This is a schematic diagram of the internal structure of the LSTM unit in an embodiment of the present invention; Figure 3 This is a schematic diagram of the MF-LSTM model structure in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of the heterogeneous data collaborative capacity assessment system for electric vehicle charging stations in an embodiment of the present invention. Detailed Implementation

[0024] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0025] It should be noted that when an element is referred to as being "fixed to" another element, it can be directly attached to the other element or there may be an intervening element. When an element is referred to as being "connected to" another element, it can be directly connected to the other element or there may be an intervening element. The terms "vertical," "horizontal," "left," "right," and similar expressions used herein are for illustrative purposes only and do not represent the only possible implementation.

[0026] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0027] like Figures 1 to 3 The method for assessing the capacity of heterogeneous data collaboration in electric vehicle charging stations, as shown, includes the following steps: Construct a membership function model based on factors influencing user charging behavior and a database of original charging behaviors; Based on the membership function model, a charging behavior data generation model that takes into account user behavior preferences is constructed according to the charging behavior constraints. The charging behavior data generation model is used to update the data in the original charging behavior database to obtain a sample charging behavior database and perform data preprocessing operations. Based on the membership function model and the preprocessed sample charging behavior database, an MF-LSTM charging load prediction model is constructed and trained. The trained MF-LSTM model is used to predict the centralized charging load of electric vehicles and to quantitatively analyze the adjustable potential of the centralized charging load of electric vehicles.

[0028] This invention introduces the membership function into the electric vehicle charging load prediction process, using the fusion of user behavior feature vectors and charging load time series data as model input. By combining the advantages of LSTM neural networks in processing time series data with the fuzzy reasoning ability of the membership function, the MF-LSTM prediction model can more accurately reflect the spatiotemporal variation of charging load and effectively improve the accuracy of charging load prediction.

[0029] Based on the above embodiments, the factors influencing user charging behavior include internal and external factors. The internal factors include the user's electricity price sensitivity and power anxiety, while the external factors include electricity price compensation and the initial state of charge of electric vehicle charging.

[0030] Based on the above embodiments, the construction of the membership function model based on factors influencing user charging behavior includes the following steps: The factors influencing user charging behavior are digitized and fuzzified to form values ​​in the range [0, 10]. Construct the output rules of the membership function model, and determine whether a user participates in power grid regulation based on the output rules.

[0031] Specifically, electric vehicle users, as the decision-makers regarding charging behavior, do not have a definite binary choice regarding whether to participate in grid regulation. Instead, their decision is a fuzzy decision influenced by a combination of internal and external factors. Because the user decision-making process is subjective and uncertain, it cannot be directly described by a precise mathematical model. Therefore, the membership function method in fuzzy inference systems is used to map the continuous values ​​of each influencing factor to fuzzy values ​​in the [0,10] interval. Output rules are then established to comprehensively determine whether users participate in grid regulation. Compared to the existing method of subjectively setting a fixed participation ratio, this step achieves objective quantification of user behavior preferences through the membership function, enabling subsequent load forecasting and capacity assessment to be based on data that more closely approximates the actual user decision-making characteristics.

[0032] This step selects four core influencing factors for modeling: electricity price compensation, initial state of charge, electricity price sensitivity, and energy anxiety. After fuzzification, each factor takes a value in the range [0,10]. A higher value indicates that the factor is more conducive to user participation in grid regulation. For electricity price compensation, this means a higher compensation amount; for initial state of charge, it means more sufficient remaining energy; for electricity price sensitivity, it means more sensitive users to changes in electricity prices; and for energy anxiety, it means a lower level of anxiety.

[0033] Compared with the existing technology that subjectively sets a fixed participation ratio, this step achieves objective quantification of user behavior preferences through membership functions, so that subsequent load forecasting and capacity assessment are based on data that is closer to the actual user decision-making characteristics.

[0034] Based on the above embodiments, the membership function model of the electricity price compensation includes three levels: low, medium, and high. The membership functions of the initial state of charge and the degree of electricity anxiety both adopt the form of Gaussian functions. The membership function of the degree of electricity price sensitivity is divided into three levels: insensitive, relatively sensitive, and highly sensitive.

[0035] Specifically, electricity price compensation is a direct external factor incentivizing users to participate in grid regulation, and the amount of compensation directly affects users' willingness to participate. Based on the amount of compensation, its membership degree is divided into three levels: low compensation, medium compensation, and high compensation, with threshold ranges of [0, 3.33], [3.33, 6.67], and [6.67, 10], respectively. A piecewise linear function is adopted, with parameters set to... , , , , .

[0036] Membership function of low electricity price compensation for: ; Membership function of electricity price compensation for: ; Membership function of high electricity price compensation for: ; in, The amount of electricity price compensation (unit: yuan / kWh) has been standardized to the range [0,10].

[0037] The design basis for the above piecewise linear membership function is that when the electricity price compensation is at an extremely low level ( When the compensation reaches a certain level, users' willingness to participate in regulation is close to zero. At this point, the membership degree for low compensation is 1, and the membership degree for medium-high compensation is 0. When the compensation enters the medium range ( to When the compensation level is between 1 and 2 (where 1 is the minimum), the membership degree is 1, indicating that the compensation level is attractive to most users; when the compensation exceeds 20%, the membership degree is 1. Afterwards, the high-compensation membership degree is 1, at which point the compensation amount is sufficient to incentivize the vast majority of users to participate in regulation. This three-stage division makes the transition region of the membership function ( to Between and to Maintain a linear gradient between steps to avoid unreasonable jumps in user behavior judgment.

[0038] The State of Charge (SOC) of an electric vehicle directly affects users' charging needs and regulatory flexibility. A lower SOC indicates a more urgent charging need and a lower willingness to participate in regulation; conversely, a higher SOC indicates greater flexibility in user participation in regulation. SOC is divided into four levels: low charge, relatively low charge, relatively high charge, and high charge, with corresponding threshold ranges of [0, 2.5], [2.5, 5], [5, 7.5], and [7.5, 10], respectively. A Gaussian function is used to construct the membership function, with σ set to 2.5 and the center value... Take values ​​of 0, 3.33, 6.67, and 10 respectively.

[0039] ; In the formula, SOC is the initial state of charge of the electric vehicle, which has been standardized to the range of [0,10]; i=1 corresponds to low charge, i=2 corresponds to relatively low charge, i=3 corresponds to relatively high charge, and i=4 corresponds to high charge.

[0040] The reason for choosing a Gaussian function instead of a piecewise linear function is that the State of Charge (SOC) has a smooth transition characteristic in its impact on user decisions; the change in user decisions between adjacent energy levels is gradual rather than abrupt. When the SOC is located in the boundary region between two levels (such as around SOC=2.5), the Gaussian function keeps both the low energy level and the lower low energy level membership values ​​at an intermediate level, more realistically reflecting the user's hesitation at this critical state. The choice allows for a moderate overlap between the different levels, ensuring both the distinguishability of the classification and avoiding overly sensitive boundary effects.

[0041] Battery anxiety refers to the degree of concern a user has about the remaining battery power of an electric vehicle, and it is a significant internal factor influencing user participation in control activities. The higher the anxiety level, the less willing the user is to participate in control activities that may affect the charging progress. Battery anxiety is categorized into four levels: no anxiety, low anxiety, medium anxiety, and high anxiety, with corresponding threshold ranges of [0, 2.5], [2.5, 5], [5, 7.5], and [7.5, 10], respectively. A Gaussian function is also used to construct the membership function, with σ set to 2.5 and the center value... Take values ​​of 0, 3.33, 6.67, and 10 respectively.

[0042] ; In the formula, The user's battery anxiety level has been standardized to the range [0,10]; i=1 corresponds to no anxiety, i=2 corresponds to low anxiety, i=3 corresponds to medium anxiety, and i=4 corresponds to high anxiety.

[0043] While both battery anxiety and initial state of charge (SOC) are related to remaining battery capacity, they reflect characteristics at different levels: SOC is an objective physical quantity reflecting the actual remaining battery capacity; battery anxiety is a subjective psychological quantity reflecting the user's subjective perception and level of concern about the remaining battery capacity. These two factors are modeled separately in a membership function model, enabling the model to distinguish between different decision-making patterns, such as objectively sufficient battery capacity but subjective anxiety, and objectively insufficient battery capacity but subjective lack of anxiety. At the same SOC level, the degree of battery anxiety may vary significantly among different users. For example, users with short commutes may not experience anxiety even with a low SOC, while users with long commutes may experience anxiety even with a high SOC.

[0044] Electricity price sensitivity reflects how well users react to changes in electricity prices. Higher sensitivity means users are more likely to participate in regulation due to incentives for electricity price compensation. Electricity price sensitivity is categorized into three levels: insensitive, moderately sensitive, and highly sensitive, with corresponding threshold ranges of [0, 3.33], [3.33, 6.67], and [6.67, 10], respectively. A piecewise linear function is used to construct the membership function, with parameters set to A2=0, B2=2.5, C2=5, D2=7.5, and E2=10.

[0045] Insensitive membership function : ; More sensitive membership function : ; Highly sensitive membership function : ; In the formula, The user's electricity price sensitivity has been standardized to the range of [0,10].

[0046] Based on the above embodiments, the step of constructing the output rules of the membership function model and determining whether a user participates in power grid regulation based on the output rules includes the following steps: Construct an output rule system comprising at least six decision rules, and then construct the output rule system into a rule matrix. Based on the determination rules, it is determined whether the user participates in power grid regulation, and based on the output rules, it is determined whether the user's participation in power grid regulation is output as an output result.

[0047] Specifically, the membership values ​​of the above four influencing factors are used as inputs to the fuzzy inference system. Output rules are formulated to comprehensively determine whether users participate in power grid regulation. The formulated output rules include 6 decision rules, and a 6-row, 7-column rule matrix R is constructed.

[0048] The expression for the regularity matrix R can be set as follows: ; Each row of the rule matrix corresponds to a fuzzy rule, and each column corresponds to a different combination of inputs. For example, the final output value is calculated through fuzzy inference to determine the user's willingness to participate in grid regulation. For example, the first four columns correspond to the level numbers of four input factors: electricity price compensation, initial state of charge, power anxiety, and electricity price sensitivity, respectively; the fifth and sixth columns are the rule weight parameters; and the seventh column is the output weight. After calculating the output value through fuzzy inference, whether the user participates in grid regulation is taken as the output result. If the output result is ≥0.5, the user is determined to participate in grid regulation (Res=1); if the output result is <0.5, the user is determined not to participate in grid regulation (Res=0). That is, the determination is made according to the following rules: ; The rule matrix contains typical decision-making patterns under different combinations of factors. For example, the first row of rules corresponds to conservative users with low compensation, low SOC, no anxiety, and low sensitivity, who will not participate in regulation even with low compensation; the fourth row of rules corresponds to proactive users with medium compensation, high SOC, low anxiety, and relatively high sensitivity, who are willing to participate in regulation at a medium compensation level.

[0049] The above steps construct a membership function model for four core influencing factors and six output rules, achieving fuzzy quantification of user charging behavior preferences. The resulting user behavior membership function feature vector will serve as one of the input features of the MF-LSTM model in subsequent steps, fusing it with charging load time series data to provide user behavior dimension input information for subsequent load forecasting.

[0050] Based on the above embodiments, the charging behavior constraints include charging duration constraints, vehicle parking duration constraints, state of charge and charging power constraints, and electricity price influence coefficient constraints.

[0051] Charging time constraints can be set based on the following considerations: Charging events with excessively short durations (such as less than 1 hour) are generally not worth regulating due to the urgent user charging needs and the limited benefits that can be obtained from regulation. Therefore, during the data generation process, charging data with a duration of more than 1 hour is first selected for subsequent processing, while invalid data with excessively short durations is removed.

[0052] Vehicle parking duration constraints can be set based on the following considerations: Electric vehicles can only participate in grid regulation during parking periods; therefore, parking duration is a key factor determining the adjustable potential. Based on user travel patterns, the probability distribution of parking duration is set for two time periods within a day: During the daytime (06:00-16:00): Users are mainly at work or out and about, and the stop duration is relatively short, following a Gaussian distribution with a mean of 4 hours and a standard deviation of 1 hour. .

[0053] ; Nighttime hours (00:00-06:00 and 16:00-24:00): Users are mainly resting at home, and the parking duration is relatively long, following a Gaussian distribution with a mean of 20 hours and a standard deviation of 4 hours. .

[0054] ; In the formula, The starting time for electric vehicle parking (unit: hour).

[0055] Based on the time interval between two consecutive charging records of the charging pile Determine the actual parking duration of the vehicle. : ; in, The theoretical parking duration is obtained by sampling from the Gaussian distribution mentioned above based on the time period. If a certain charging period is the last charging session for that charging pile during the day, then the period for which it participates in regulation is from the start of charging to 06:00 the next day. The above-mentioned time-segmented Gaussian distribution parameters are set based on the following: daytime users typically park for about 4 hours at their workplace, such as arriving in the morning and leaving after lunch; nighttime users typically park at home until the next morning, covering the nighttime period for about 20 hours, and the standard deviation reflects the degree of difference in the work and rest habits of different users.

[0056] State of charge and charging power constraints can be set based on the following considerations: to ensure the user's normal vehicle usage needs, the total amount of charging the user receives after participating in grid regulation and before leaving the charging station. It must be greater than or equal to its original charge. Meanwhile, charging power The charging power cannot exceed the maximum charging power of the charging station. .

[0057] ; Among them, the power constraint ensures that the charging process is carried out within the rated operating range of the charging pile and does not cause equipment overload; the power constraint ensures that the battery power is not lower than the power level under the original charging scheme when the user leaves, so as not to impair the user experience; these two constraints are the prerequisites for user participation in regulation. Only when the above constraints are met can the user who is judged to participate in regulation actually perform the regulation behavior.

[0058] The electricity price impact coefficient constraint can be set based on the following considerations: users' charging behavior is also affected by time-of-use pricing. During periods of low electricity prices, users are more willing to charge; during periods of high electricity prices, users tend to postpone charging. Therefore, an electricity price impact coefficient is introduced to correct users' willingness to participate in regulation: the impact coefficient for low-price areas is set at 1.2, and the impact coefficient for high-price areas is set at 0.8. This coefficient, correlated with the output of the membership function model, yields the corrected probability of participation in regulation. This reflects that in actual charging decisions, users consider the current base electricity price level in addition to the compensation amount; that is, during periods with lower base electricity prices, even with the same compensation amount, users' willingness to participate is higher.

[0059] Based on the above embodiments, the step of updating the data in the original charging behavior database using the charging behavior data generation model includes the following steps: Load the original charging database and filter based on charging duration constraints; The membership function model is used to calculate the probability of each user participating in the demand response; For vehicles participating in power grid regulation, the vehicle parking time constraint, the state of charge and charging power constraint are applied. By constraining users' willingness to participate in regulation through the electricity price impact coefficient, sample charging behavior data is generated. Repeat the above steps until the data in the original charging database is updated to the sample charging behavior data.

[0060] Based on the membership function model output combined with the physical constraints of charging behavior, sample charging behavior data that takes into account user behavioral preferences is generated using Monte Carlo simulation. This not only reflects the fuzzy characteristics of user decision-making but also satisfies the physical feasibility of charging behavior, providing a sufficient data foundation that conforms to actual physical laws for the subsequent training of the MF-LSTM model.

[0061] Based on the above embodiments, the data preprocessing operations include abnormal data processing, data sample set expansion, data standardization, and training set and test set partitioning.

[0062] Specifically, during the generation of sample charging behavior data, there may be issues such as outliers, missing values, and inconsistent units of measurement. Directly inputting these into the neural network can lead to instability in the training process and slow convergence. By employing four sub-steps—outlier data processing, data sample set expansion, data standardization, and training / test set partitioning—the raw data is transformed into a standardized format suitable for neural network training.

[0063] Anomaly processing addresses various outliers that may exist in the original charging data, such as abnormal charging power, abnormal charging duration, and empty data. This invention employs the following rules for anomaly cleaning: Data with charging power below 1kWh or above 60kWh was removed. The 1kWh lower limit was used to exclude invalid records caused by standby power consumption of charging piles, while the 60kWh upper limit corresponds to the maximum charging capacity of mainstream fast charging piles in one hour. Data with a charging duration of only one sampling point was also removed, as such data cannot provide effective time-series information. All empty data and missing values ​​were also removed. After the above cleaning, the quality and consistency of the data were ensured, avoiding interference from outliers in subsequent model training.

[0064] The charging power of electric vehicles typically exhibits a "rise-stabilize-fall" trend. However, due to sampling frequency limitations, some data points may be lost at the start and end of charging. To improve the data sample and enhance the model's fitting performance, linear interpolation is used to fill in and expand the high-power charging data at the start and end of charging. Specifically, for missing data between two adjacent sampling points, linear interpolation is performed based on the power values ​​and time intervals between the preceding and following sampling points to recover the power estimate at the intermediate moment. This process makes the charging power curve more complete at the start and end stages, helping the LSTM model learn the complete charging power change pattern.

[0065] Because different features have significantly different dimensions and value ranges, directly inputting them into a neural network can lead to unstable training and slow convergence. Therefore, it is necessary to standardize the data, converting it into standardized data with zero mean and unit variance.

[0066] Calculate the mean of the training data and standard deviation : ; ; Standardized formula: ; In the formula, The total number of vehicles. The original charging data for the i-th vehicle. This is the standardized data. After standardization, all features have a mean of 0 and a variance of 1, eliminating the impact of differences in units of measurement on model training. It is important to note that the standardization of the test set data was calculated using the training set. and Instead of recalculating, this ensures consistency in the data distribution between the training and test sets.

[0067] The training and test sets can be partitioned as follows: the preprocessed time series data is divided into training and test sets in an 8:2 ratio. The training set is used for model parameter learning, while the test set is used to verify the model's generalization ability and prediction accuracy. Since charging load exhibits significant weekly and daily periodicity, the data is partitioned chronologically to ensure the training set contains complete periodic features. The test set covers the time period following the training set, simulating real-world scenarios where historical data is used to predict future loads.

[0068] Based on the above embodiments, the step of constructing and training the MF-LSTM charging load prediction model according to the membership function model and the preprocessed sample charging behavior database includes the following steps: Data fusion processing is performed on the membership function feature vector obtained from the membership function model and the charging load time series data in the sample charging behavior database to obtain input data; Configure the model parameters of the LSTM neural network and input the input data into the configured LSTM neural network model. The LSTM neural network includes a forget gate, an input gate, an output gate, and cell states. The MF-LSTM charging load prediction model is obtained by performing model training.

[0069] This invention constructs a long short-term memory neural network model (MF-LSTM) that integrates membership function features. It combines user behavior membership function feature vectors with charging load time series data as input to the model, enabling the LSTM model to simultaneously consider the historical time series patterns of the load and user behavior preference features when predicting charging load, thus achieving accurate prediction of charging load.

[0070] Specifically, Long Short-Term Memory Neural Network (LSTM) is a special type of Recurrent Neural Network (RNN). By introducing three gating mechanisms—forget gate, input gate, and output gate—it solves the gradient vanishing and gradient explosion problems that exist in traditional RNNs, and can effectively process long sequence data.

[0071] The structure of an LSTM cell is as follows: Figure 2As shown, its internal structure includes cell states (CellState, and three gating units: ForgetGate: Determines which information to discard from the cell state.

[0072] Input Gate: Determines which new information is stored in the cell state.

[0073] Output gate: determines which information is output.

[0074] Based on the above embodiments, the LSTM neural network includes a forget gate, an input gate, an output gate, and cell states, which are calculated using the following formulas: ; ; ; ; ; ; in, Output for the forget gate. For the input gate state, To add a new candidate status, The cell state at time t. Output gate state The final hidden output at time t, , , , For matrix weights, , , , Here, σ is the bias term, sigmoid function, and tanh is hyperbolic tangent function; For the previous moment The hidden layer output; For the current moment The input feature vector; For the previous moment The cellular state.

[0075] like Figure 3 As shown, the structure of the MF-LSTM model mainly consists of three parts: the input layer, the hidden layer, and the output layer. Input Layer: The user behavior membership function feature vector (8-dimensional) is fused with the charging load time series data (1-dimensional) to form a 9-dimensional input vector. The 8-dimensional feature vector consists of the membership values ​​of four influencing factors: electricity price compensation (3-dimensional, 1-dimensional for each of the low / medium / high levels), initial state of charge (4-dimensional, 1-dimensional for each of the four levels), and power anxiety (4-dimensional, 1-dimensional for each of the four levels). After dimensionality reduction, these are combined with the electricity price sensitivity (3-dimensional, 1-dimensional for each of the three levels) to form the 8-dimensional vector. The 1-dimensional load data is the charging power value at the current moment; this allows the LSTM to simultaneously perceive load changes and user behavior characteristics at each input step, rather than treating user behavior as a static external variable.

[0076] Hidden Layers: Two LSTM hidden layers are used, each containing 100 neurons. Through multiple experiments, this structure achieves the optimal balance between prediction accuracy and computational complexity. The two-layer structure enables the model to extract features hierarchically. Specifically, the first LSTM layer primarily extracts short-term fluctuation patterns in charging load and low-level interactions of user behavior features. The second LSTM layer then builds upon this foundation to extract longer-term dependencies and higher-level feature combinations. The choice of 100 neurons has been validated through multiple experiments, achieving a balance between prediction accuracy and computational complexity. Too few neurons can lead to insufficient model capacity and underfitting, while too many neurons can increase training time and potentially cause overfitting.

[0077] Output layer: A fully connected layer is used to output the predicted charging load value at future time. The output of the second LSTM layer is mapped to the predicted charging load value at future time (1-dimensional).

[0078] The model is trained using the Adam optimizer with a learning rate of 0.001 and mean squared error (MSE) as the loss function. An early stopping strategy is employed during training: training is stopped if the validation set loss does not decrease for 10 consecutive epochs to prevent overfitting. The batch size can be set to 64.

[0079] Based on the above embodiments, the evaluation metrics for model training are the root mean square error (RMSE) and the coefficient of determination. : ; ; In the formula, These are the model's predicted values. This represents the actual validation value, where n is the total number of data points; the smaller the RMSE, the better. The closer the value is to 1, the higher the model's prediction accuracy.

[0080] Using the trained MF-LSTM model, we performed time-series predictions of charging load before and after considering user behavior preferences. By comparing and analyzing the two sets of prediction results, we quantitatively evaluated the adjustable potential of centralized charging load for electric vehicles and analyzed the potential differences under different scenarios.

[0081] Based on the above embodiments, the average load variation value is adopted. Average load offset percentage The adjustable time period and adjustable continuity quantitative analysis of the adjustable potential of the centralized charging load of electric vehicles, namely the average load variation value. The expression is: ; The average load offset percentage The expression is: ; In the formula, P is the load value without considering user preferences. The load value is taken into account user preferences, where T is the number of hours in the time period. This indicates that the load decreased after regulation. This indicates an increase in load after regulation.

[0082] At the same time, it is necessary to consider the adjustable time period and adjustable duration. By statistically analyzing the adjustable time period with a duration of more than 1 hour, the distribution of adjustable time and duration capability under different scenarios can be analyzed.

[0083] Based on the above embodiments, the method also includes analyzing the impact of user behavior preferences on load adjustability potential under special scenarios, including: weekdays or holidays, summer or winter, residential areas or commercial areas.

[0084] Different date scenarios: divided into weekdays and holidays. On weekdays, users have regular travel patterns, concentrated commuting times, and relatively stable parking durations, resulting in a clear bi-peak charging load characteristic in the morning and evening, with greater potential for adjustment. On holidays, users' travel is more random, the peak and valley characteristics of the charging load are weakened, and the potential for adjustment is relatively smaller.

[0085] Different seasonal scenarios: divided into summer and winter. In summer, the temperature is high, the energy consumption of electric vehicle air conditioning is high, the charging demand is high, the load peak-valley difference is large, and the adjustment potential is also large. In winter, the temperature is low, the battery performance decreases, the charging demand is also high, but users' anxiety about the battery level increases, and their willingness to participate in regulation is somewhat restricted, but the adjustment potential is slightly lower than in summer.

[0086] Different functional areas are divided into residential and commercial areas. Residential users park for longer periods at night, with adjustable time slots concentrated at night, resulting in large single-time adjustable capacity and long duration. Commercial users park for shorter periods during the day, with adjustable time slots scattered and small adjustable capacity.

[0087] By comparing and analyzing load data under different scenarios, the following can be output: the adjustable load capacity, adjustable time period distribution, and adjustable sustainability index under each scenario; and the ranking of the impact of various influencing factors (electricity price compensation, initial state of charge, electricity price sensitivity, and power anxiety) on the adjustable potential under different scenarios. For example, in the scenario of winter weekdays in residential areas, after considering user behavior preferences, the average load change from 06:00 to 20:00 during the day is -62.117kW, and the average load change from 20:00 to 06:00 the next day is 88.952kW, with an average load deviation percentage of 73.1%. This indicates that through reasonable incentive mechanisms and control strategies, the adjustable potential of electric vehicles can be fully tapped, providing effective demand-side response services to the power grid.

[0088] like Figure 4 As shown, the present invention also provides a heterogeneous data collaborative capacity assessment system for electric vehicle charging stations, comprising: The function model building module is used to build a membership function model based on the factors influencing user charging behavior and a database of original charging behavior. The database generation and processing module is used to construct a charging behavior data generation model that takes into account user behavior preferences based on the membership function model and charging behavior constraints. The charging behavior data generation model is used to update the data in the original charging behavior database to obtain a sample charging behavior database and perform data preprocessing operations. The model building and training module constructs and trains the MF-LSTM charging load prediction model based on the membership function model and the preprocessed sample charging behavior database. The adjustable potential assessment module is used to predict the concentrated charging load of electric vehicles using the trained MF-LSTM model and to quantitatively analyze the adjustable potential of the concentrated charging load of electric vehicles.

[0089] This invention introduces the membership function into the electric vehicle charging load prediction process, using the fusion of user behavior feature vectors and charging load time series data as model input. By combining the advantages of LSTM neural networks in processing time series data with the fuzzy reasoning ability of the membership function, the MF-LSTM prediction model can more accurately reflect the spatiotemporal variation of charging load and effectively improve the accuracy of charging load prediction.

[0090] Specifically, by organically combining the fuzzy reasoning capability of the membership function with the temporal modeling capability of the LSTM neural network—that is, the membership function is responsible for handling the fuzziness and uncertainty of user decisions, while the LSTM is responsible for capturing the long-term temporal dependence and nonlinear variation of charging load—the synergy of these two approaches enables the MF-LSTM model to more accurately reflect the spatiotemporal variation of charging load after being affected by user behavior compared to traditional single models such as BP, RNN, and GRU. Furthermore, by setting four physical constraints—charging duration limit, vehicle parking duration limit, state of charge limit, and charging power limit—the model assesses whether the decision-making level participates in regulation and the feasibility of the operation, making the load adjustment potential evaluated based on this database more consistent with the actual regulation needs of the power grid.

[0091] Based on the above embodiments, the adjustable potential assessment module further includes a scenario analysis submodule, which is used to analyze the degree of influence of user behavior preferences on the load adjustable potential under special scenarios, and output the load adjustable capacity, adjustable time period and adjustable sustainability index under each special scenario.

[0092] By analyzing the three dimensions of different dates, seasons, and functional areas, the system quantifies the differences in the adjustable potential of centralized charging load for electric vehicles, clarifies the degree of influence of each influencing factor on the adjustable potential in different scenarios, and outputs the adjustable load capacity, adjustable time period, and adjustable sustainability indicators. This provides a quantitative basis for power grid dispatching departments to formulate differentiated demand response strategies in different scenarios.

[0093] Those skilled in the art should understand that this invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to this invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for assessing the collaborative capacity of heterogeneous data in electric vehicle charging stations, characterized in that, Includes the following steps: Construct a membership function model based on factors influencing user charging behavior and a database of original charging behaviors; Based on the membership function model, a charging behavior data generation model that takes into account user behavior preferences is constructed according to the charging behavior constraints. The charging behavior data generation model is used to update the data in the original charging behavior database to obtain a sample charging behavior database and perform data preprocessing operations. Based on the membership function model and the preprocessed sample charging behavior database, an MF-LSTM charging load prediction model is constructed and trained. The trained MF-LSTM model is used to predict the centralized charging load of electric vehicles and to quantitatively analyze the adjustable potential of the centralized charging load of electric vehicles.

2. The method for assessing the capacity of heterogeneous data collaboration in electric vehicle charging stations according to claim 1, characterized in that, The factors influencing user charging behavior include internal and external factors. The internal factors include users' sensitivity to electricity prices and their level of anxiety about electricity consumption. The external factors include electricity price compensation and the initial state of charge of electric vehicles.

3. The method for assessing the capacity of heterogeneous data collaboration in electric vehicle charging stations according to claim 2, characterized in that, The construction of the membership function model based on factors influencing user charging behavior includes the following steps: The factors influencing user charging behavior are digitized and fuzzed to form values ​​in the range [0, 10]. Construct the output rules of the membership function model, and determine whether a user participates in power grid regulation based on the output rules.

4. The method for assessing the capacity of heterogeneous data collaboration in electric vehicle charging stations according to claim 2, characterized in that, The membership function model for electricity price compensation includes three levels: low, medium, and high. The membership functions for the initial state of charge and the degree of electricity anxiety both adopt the form of Gaussian functions. The membership function for the degree of electricity price sensitivity is divided into three levels: insensitive, relatively sensitive, and highly sensitive.

5. The method for assessing the capacity of heterogeneous data collaboration in electric vehicle charging stations according to claim 3, characterized in that, The process of constructing the output rules of the membership function model and determining whether a user participates in power grid regulation based on the output rules includes the following steps: Construct an output rule system comprising at least six decision rules, and then construct the output rule system into a rule matrix. Based on the determination rules, it is determined whether the user participates in power grid regulation, and based on the output rules, it is determined whether the user's participation in power grid regulation is output as an output result.

6. The method for assessing the capacity of heterogeneous data collaboration in electric vehicle charging stations according to claim 1, characterized in that, The charging behavior constraints include charging duration constraints, vehicle parking duration constraints, state of charge and charging power constraints, and electricity price impact coefficient constraints.

7. The method for assessing the capacity of heterogeneous data collaboration in electric vehicle charging stations according to claim 6, characterized in that, The process of updating the data in the original charging behavior database using the charging behavior data generation model includes the following steps: Load the original charging database and filter based on charging duration constraints; The membership function model is used to calculate the probability of each user participating in the demand response; For vehicles participating in power grid regulation, the vehicle parking time constraint, the state of charge and charging power constraint are applied. By constraining users' willingness to participate in regulation through the electricity price impact coefficient, sample charging behavior data is generated. Repeat the above steps until the data in the original charging database is updated to the sample charging behavior data.

8. The method for assessing the capacity of heterogeneous data collaboration in electric vehicle charging stations according to claim 1, characterized in that, The data preprocessing operations include anomaly processing, data sample set expansion, data standardization, and training and test set partitioning.

9. The method for assessing the capacity of heterogeneous data collaboration in electric vehicle charging stations according to claim 1, characterized in that, The step of constructing and training the MF-LSTM charging load prediction model based on the membership function model and the preprocessed sample charging behavior database includes the following steps: Data fusion processing is performed on the membership function feature vector obtained from the membership function model and the charging load time series data in the sample charging behavior database to obtain input data; Configure the model parameters of the LSTM neural network and input the input data into the configured LSTM neural network model. The LSTM neural network includes a forget gate, an input gate, an output gate, and cell states. The MF-LSTM charging load prediction model is obtained by performing model training.

10. The method for assessing the capacity of heterogeneous data collaboration in electric vehicle charging stations according to claim 9, characterized in that, The evaluation metrics for model training are the root mean square error (RMSE) and the coefficient of determination. : ; ； In the formula, These are the model's predicted values. This represents the actual validation value, where n is the total number of data points; the smaller the RMSE, the better. The closer the value is to 1, the higher the model's prediction accuracy.

11. The method for assessing the capacity of heterogeneous data collaboration in electric vehicle charging stations according to claim 1, characterized in that, Using average load variation Average load offset percentage The adjustable time period and adjustable continuity quantitative analysis of the adjustable potential of the centralized charging load of electric vehicles, namely the average load variation value. The expression is: ； The average load offset percentage The expression is: ； In the formula, P is the load value without considering user preferences. The load value is taken into account user preferences, where T is the number of hours in the time period.

12. The method for assessing the capacity of heterogeneous data collaboration in electric vehicle charging stations according to claim 1, characterized in that, It also includes analyzing the impact of user behavior preferences on load adjustability potential under special scenarios, such as weekdays or holidays, summer or winter, and residential or commercial areas.

13. A heterogeneous data collaborative capacity assessment system for electric vehicle charging stations, characterized in that, include: The function model building module is used to build a membership function model based on the factors influencing user charging behavior and a database of original charging behavior. The database generation and processing module is used to construct a charging behavior data generation model that takes into account user behavior preferences based on the membership function model and charging behavior constraints. The charging behavior data generation model is used to update the data in the original charging behavior database to obtain a sample charging behavior database and perform data preprocessing operations. The model building and training module constructs and trains the MF-LSTM charging load prediction model based on the membership function model and the preprocessed sample charging behavior database. The adjustable potential assessment module is used to predict the concentrated charging load of electric vehicles using the trained MF-LSTM model and to quantitatively analyze the adjustable potential of the concentrated charging load of electric vehicles.

14. The heterogeneous data collaborative capacity assessment system for electric vehicle charging stations according to claim 13, characterized in that, The adjustable potential assessment module also includes a scenario analysis submodule, which is used to analyze the impact of user behavior preferences on the load adjustable potential under special scenarios, and output the load adjustable capacity, adjustable time period and adjustable sustainability index under each special scenario.

Images