Electric vehicle charging station load prediction and schedulable resource aggregation method and system based on CNN-TSMIX model

Abstract

The invention discloses an electric vehicle charging station load prediction and schedulable resource aggregation method and system based on a CNN-TSMIX model, and relates to the field of electric vehicle charging station energy management. Load data of a plurality of charging stations are collected, and data preprocessing is carried out; training a global model by using the preprocessed data; training a local model suitable for each charging station by finely adjusting global model parameters; predicting the power of each time point under each charging station and the number of various automobiles at each time point by using the established local model in a rolling updating mode; and establishing a consumer psychological model, evaluating the flexible resource adjustment potential of each charging station by using the obtained predicted value, and aggregating the flexible resource adjustment potential into a virtual power plant according to a rule. According to the invention, high-precision prediction of the load of the charging station is realized. Moreover, the energy use efficiency of the charging station is optimized, and a virtual power plant is formed to provide services for a power grid on the premise that the charging experience of an electric vehicle user is not affected.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electric vehicle charging station energy management, more particularly to an electric vehicle charging station load prediction and dispatchable resource aggregation method and system based on a CNN-TSMixer model. BACKGROUND

[0002] At present, under the background of the country vigorously promoting the development of new energy technology, electric vehicles, as an important carrier of green travel, are popularizing in every corner of society at an unprecedented speed. According to statistics, the number of new energy electric vehicles continues to rise, not only reflecting the public's positive response to environmental protection concepts, but also indicating a profound change in the future transportation energy structure. However, the widespread popularity of electric vehicles poses new challenges to traditional power distribution networks, especially the randomness and concentration of their charging demand, which poses a potential threat to the stable operation and energy management of the power grid.

[0003] Accurate electric vehicle load prediction is crucial for predicting energy management systems. It can help power grid operators understand the distribution and scale of future electric vehicle charging demand in advance, so as to reasonably plan the capacity of the power grid, optimize the allocation of power resources, and avoid power grid overload or power shortages due to sudden increases in electric vehicle charging demand. By accurately predicting electric vehicle load, the power system can be more intelligent and efficient in scheduling and management, ensuring power supply and demand balance, improving the economic and reliable operation of the power grid, and providing more stable and convenient charging services for electric vehicle users, promoting the popularity of new energy vehicles and the promotion of green travel.

[0004] Different types of vehicles (such as buses, private cars, and taxis) have significant differences in their use of charging stations, resulting in large fluctuations in charging station load at different times. At the same time, the imbalance of power grid load also brings operational risks. There is currently little research on charging stations participating in power market dispatch.

[0005] Therefore, how to provide a method that can accurately predict charging station load power and different types of vehicles at the same time to reasonably evaluate the adjustable capacity of flexible resources of charging stations is a problem that needs to be solved by those skilled in the art. SUMMARY

[0006] Therefore, the present application provides an electric vehicle charging station load prediction and dispatchable resource aggregation method and system based on a CNN-TSMixer model to solve the problems in the background art.

[0007] To achieve the above purpose, the present application adopts the following technical solutions:

[0008] A load prediction and schedulable resource aggregation method for electric vehicle charging stations based on a CNN-TSMixer model, comprising:

[0009] Step one: collect load data of multiple charging stations and perform data preprocessing;

[0010] Step two: train a global model suitable for multiple charging stations using the data processed in step one, and save the global model parameters with the best prediction performance;

[0011] Step three: use the global model parameters of step two to train local models suitable for each charging station by adjusting the global model parameters obtained in step two;

[0012] Step four: use the local model established in step three to predict the power at each time point and the number of vehicles at each time point under each charging station in a rolling update manner;

[0013] Step five: establish a consumer psychology model to evaluate the flexibility resource adjustment potential of each charging station using the prediction values obtained in step four;

[0014] Step six: according to the voltage regulation and peak regulation service requirements of the power grid, aggregate the flexibility resources into voltage regulation resource clusters and peak regulation resource clusters.

[0015] Optionally, the data preprocessing includes abnormal value processing and data normalization operation.

[0016] Optionally, step two uses the CNN-TSMixer algorithm to train a global model suitable for all time series, continuously adjusts the model parameters, and saves the model parameters with the best prediction performance for constructing local models;

[0017] The CNN-TSMixer algorithm is as follows: TSMixer expands the function of a univariate linear model by adding a cross-variable feedforward layer, thereby forming a linear model capable of handling multivariate prediction; the architecture of TSMixer is relatively simple, based on a multilayer perceptron, and consists of two steps: mixer layer and time projection;

[0018] Mixer layer: the core part of time mixing and feature mixing; time mixing is achieved through a fully connected layer, a Re LU activation function and a dropout layer; the rows of the input data represent time and the columns represent features; after transposition, the data is processed to share features in the time domain; before leaving the time mixing unit, the matrix is transposed again and sent to the feature mixing unit;

[0019] The feature mixing unit is composed of two MLPs, applied to the feature domain, and does not need to be transposed because the features are already on the horizontal axis; a normalization layer and a residual connection are included in the mixer to facilitate the model to learn a deeper data representation while maintaining computational efficiency;

[0020] Temporal projection: responsible for generating predictions, the matrix is processed by a fully connected layer after transposing again, and the final output matrix places the features on the horizontal axis and the time steps on the vertical axis.

[0021] Optionally, the step two further comprises, using a CNN network to extract spatial features from a large amount of historical load data, and using the CNN to capture spatial features and TSMixer to process time sequence dependence; using a convolution kernel, performing convolution and pooling operations on the load feature data set to extract numerous features in the load data, and inputting the extracted features into the TSMixer model along with the load feature data set to train the global and local models.

[0022] Optionally, the global model training process is as follows:

[0023] In order to speed up the convergence speed of the model, the load data is normalized;

[0024] The normalized load data is input into the established global model for training;

[0025] The training early stopping algorithm is used to test the trained model through the validation set, and the parameters of the model with the best prediction performance are saved.

[0026] Optionally, the local model training process is as follows:

[0027] In order to speed up the convergence speed of the model, the load data of the charging station is normalized;

[0028] The normalized load data is input into the established local model, and the local model loads the parameters trained by the global model, and then the training is performed;

[0029] The training early stopping algorithm is used to test the trained model through the validation set, and the parameters of the model with the best prediction performance are saved.

[0030] Optionally, the consumer psychology model comprises:

[0031] Differentiate user types and build a basic response model;

[0032] Users respond to price changes by adjusting their electricity consumption periods, and introduce a power demand price elasticity coefficient to measure the impact of price changes on electricity consumption;

[0033] The influencing factors of the cross-charging station scheduling are taken into account, when the mutual support between the charging stations is considered, the additional power consumption is generated when the electric vehicles travel across the stations, and the users will respond only when the electricity price incentive exceeds the cost; the greater the distance between the two stations, the smaller the slope of the linear interval of the user psychological response model;

[0034] The travel time and the queuing waiting time are considered in the cross-station scheduling, and these time costs need to be converted into the user's psychological expected price, which affects the user's response willingness;

[0035] The user psychology price response model considering the cross-station scheduling is constructed as the consumer psychology model by combining the influence of the price change on the power consumption, the influencing factors of the cross-charging station scheduling, and the travel time and the queuing waiting time, and the user's response behavior to the price change is described;

[0036] The flexible resource adjustment potential of each charging station is obtained by substituting the future period power and vehicle quantity prediction values of each charging station obtained in step four into the consumer psychology model.

[0037] Optionally, the constraint is specifically:

[0038] When the cross-charging station scheduling is considered, the travel distance of the electric vehicle satisfies the constraint:

[0039]

[0040] In the formula, L max is the maximum distance of the electric vehicle from the i charging station to the n charging station;

[0041] The queuing waiting time of the electric vehicle satisfies the constraint:

[0042]

[0043] In the formula, is the maximum waiting time of the electric vehicle in the n charging station.

[0044] Optionally, the user psychology price response model considering the cross-charging station scheduling is as follows:

[0045]

[0046] In the formula, ρ pv is the participation of the electric vehicle in the demand response incentive; λ EV,max is the upper limit of the load transfer rate of the electric vehicle; a EV is the upper limit of the load transfer rate of the electric vehicle to the peak-valley electricity price difference dead zone; p EV is the power consumption cost of the cross-charging station scheduling of the electric vehicle; K EV is the slope of the linear interval; is the function of converting the time into the economic cost, travel time for an electric vehicle to travel from the i-th charging station to the n-th charging station; queue time at the n-th charging station, VOT is the psychological value / expense corresponding to the unit time (such as per hour), and the unit is yuan / hour.

[0047] A load prediction and schedulable resource aggregation system for electric vehicle charging stations based on a CNN-TSMixer model, comprising:

[0048] a charging behavior data collection module for collecting load data of the charging stations;

[0049] a data preprocessing module for eliminating outliers in the collected data and completing part of the missing values in the data collection;

[0050] a global model training module for training a global model to provide initial training parameters for the training of a local model;

[0051] a local model training module for training a prediction model more suitable for each charging station by fine-tuning the global model parameters;

[0052] a load prediction module for completing the actual prediction task;

[0053] a flexibility resource evaluation module for evaluating the flexibility resource adjustment potential of each charging station;

[0054] a flexibility resource aggregation module for dividing the flexibility resources into a voltage regulation resource cluster and a peak shaving resource cluster.

[0055] According to the above technical solution, compared with the prior art, the present disclosure provides a load prediction and schedulable resource aggregation method and system for electric vehicle charging stations based on a CNN-TSMixer model, which collects load data of numerous charging stations and performs data preprocessing; trains a global model using the data processed in step one; trains a local model suitable for each charging station by fine-tuning the global model parameters; uses the established local model to predict the power at each time point and the number of vehicles at each time point under each charging station in a rolling update manner; establishes a consumer psychology model, uses the obtained prediction value to evaluate the flexibility resource adjustment potential of each charging station and aggregates resources. The present disclosure uses the CNN-TSMixer algorithm to achieve high-precision prediction of the load of the charging station. At the same time, the present disclosure also involves the evaluation of schedulable resources, aiming to optimize the energy use efficiency of the charging station, reduce the operating cost, and improve the charging experience of electric vehicle users. By combining advanced deep learning algorithms with intelligent energy management technology, the present disclosure provides an innovative solution for the intelligentization and sustainable development of electric vehicle charging stations. BRIEF DESCRIPTION OF DRAWINGS

[0056] In order to make the technical solutions in the embodiments of the present application or the prior art clearer, the accompanying drawings needed in the embodiments or prior art description will be briefly introduced. Obviously, the accompanying drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without any creative effort on the basis of the provided drawings.

[0057] Figure 1 The method flowchart provided by the present application.

[0058] Figure 2 The TSMixer structure diagram provided by the present application

[0059] Figure 3 The convolution feature extraction process diagram provided by the present application;

[0060] Figure 4a 、 Figure 4b The taxi user and private car user psychological electricity price response model display diagrams provided by the present application in the charging station respectively;

[0061] Figure 5a 、 Figure 5b The taxi user and private car user psychological electricity price response model display diagrams provided by the present application in the charging station considering cross-charging station scheduling respectively;

[0062] Figure 6 The data flow diagram of the method provided by the present application;

[0063] Figure 7 The system module relationship diagram provided by the present application;

[0064] Figure 8 The total load of each period and the number of various types of vehicles in the charging station A provided by the present application;

[0065] Figure 9 The total load of each period and the number of various types of vehicles in the charging station B provided by the present application;

[0066] Figure 10 The total load of each period and the number of various types of vehicles in the charging station C provided by the present application;

[0067] Figure 11 The arrangement diagram of the power stations A, B and C provided by the present application;

[0068] Figure 12 The schedulable resource evaluation diagram provided by the present application;

[0069] Figure 13 The schedulable resource aggregation diagram provided by the present application;

[0070] Figure 14 A schematic diagram of the device provided by the embodiment of the application is shown in the following figure; DETAILED DESCRIPTION

[0071] The technical solutions in the embodiments of the application will be clearly and completely described in connection with the drawings in the embodiments of the application. Obviously, the described embodiments are only a part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the application.

[0072] The embodiment of the application discloses a method for predicting and aggregating load of an electric vehicle charging station based on a CNN-TSMixer model, as shown in the following figure, which comprises the following steps: Figure 1

[0073] Step one: collect load data of multiple charging stations and perform data preprocessing;

[0074] Step two: train a global model suitable for multiple charging stations using the data processed in step one, and save the global model parameters with the best prediction performance;

[0075] Step three: use the global model parameters in step two to train local models suitable for each charging station by adjusting the global model parameters obtained in step two;

[0076] Step four: use the local model established in step three to predict the power at each time point and the number of vehicles at each time point under each charging station in a rolling update manner;

[0077] Step five: establish a consumer psychology model to evaluate the flexibility resource adjustment potential of each charging station using the prediction values obtained in step four;

[0078] Step six: according to the voltage regulation and peak regulation service requirements of the power grid, aggregate the flexibility resources into voltage regulation resource clusters and peak regulation resource clusters.

[0079] Step one, collect load data of multiple charging stations and perform data preprocessing; preprocessing includes but is not limited to operations such as abnormal value processing and data normalization.

[0080] Step two, using the data processed in step one, train a global model suitable for all time series using the CNN-TSMixer algorithm, constantly adjust the model parameters, and save the model parameters with the best prediction performance for constructing local models.

[0081] The detailed introduction of the CNN-TSMixer algorithm is as follows:

[0082] ​In the field of forecasting, utilizing cross-variable information helps to improve the accuracy of predictions. For example, factors such as weather, whether it is a holiday, and so on can affect the travel of electric vehicles, and in turn, affect the load of electric vehicle charging stations. Therefore, it is meaningful to build a prediction model that can utilize numerous features. This has led to the creation of TSMixer. TSMixer extends the functionality of univariate linear models by adding a cross-variable feedforward layer, forming a linear model capable of handling multivariate predictions. The architecture of TSMixer is relatively simple and is mainly based on multilayer perceptron (MLP), as shown in Figure 2

[0083] Mixing layer: This is the core part of time mixing and feature mixing. Time mixing is achieved through a fully connected layer, a ReLU activation function, and a dropout layer. The rows of the input data represent time, and the columns represent features. After transposition, the data is processed to share features across the time domain. Before leaving the time mixing unit, the matrix is transposed again and sent to the feature mixing unit.

[0084] The feature mixing unit consists of two MLPs (Multilayer Perceptron) applied to the feature domain and does not require transposition as the features are already on the horizontal axis. The mixer includes a normalization layer and a residual connection to facilitate the model in learning deeper data representations while maintaining computational efficiency.

[0085] Time projection: This step is responsible for generating predictions. The matrix is processed through a fully connected layer after transposition, and the final output matrix places features on the horizontal axis and time steps on the vertical axis.

[0086] In summary, TSMixer enhances the predictive ability of linear models through effective time and feature mixing, aiming to utilize various covariates and static features for multivariate predictions.

[0087] Convolutional Neural Network (CNN) is a deep learning model widely used in image processing and computer vision fields. The design of CNN is inspired by the structure of the biological visual cortex, which simulates the way the human brain processes visual information. CNN can automatically learn feature representations without manual feature extraction. Through multiple convolution and pooling operations, CNN can automatically learn complex temporal and spatial features in the load data, avoiding the limitations of manual feature extraction. In particular, CNN is sensitive to local dependencies in input data, which enables it to effectively capture potential patterns in load changes, especially in seasonal fluctuations, holiday effects, and daily changes.

[0088] In the task of load forecasting, such as Figure 3 ​The CNN network is used for spatial feature extraction on a large amount of historical load data, and the CNN is used to capture spatial features and the TSMixer is used to process time sequence dependence, so that the accuracy and response speed of load prediction are further improved.

[0089] The load feature data set is convolved and pooled by using a convolution kernel, and a plurality of features in the load data are fully extracted, and the extracted features are input into the TSMixer model along with the load feature data set to train the global and local models.

[0090] The input data of the global-local model constructed by the application includes but is not limited to charging station historical load, temperature, humidity, rainfall, air pressure, whether it is a working day, and year, month, day, hour, minute and second information.

[0091] The global model training process is as follows:

[0092] In order to accelerate the convergence speed of the model, the load data set is normalized;

[0093] The above data is input into the established global model for training;

[0094] The training early stopping algorithm is used to test the training model through the validation set, and the parameters of the model with the best prediction performance are saved.

[0095] Step three, based on the global model prediction parameters of step two, for each charging station, a local prediction model more suitable for the charging station is trained. The established model and the input data form are consistent with step two, and are not repeated here. The difference between the local model training and the global model of step two is that the initial parameters of the model are initialized by the model itself, and the initial training parameters of the local model are determined by the global model parameters.

[0096] The local model training process is as follows:

[0097] In order to accelerate the convergence speed of the model, the charging station load data is normalized;

[0098] The above data is input into the established local model, and the local model loads the parameters trained by the global model, and then is trained;

[0099] The training early stopping algorithm is used to test the training model through the validation set, and the parameters of the model with the best prediction performance are saved.

[0100] Step four, using the local model obtained in step three, the power at each time point of each charging station and the number of each type of car at each time point are predicted in a rolling update manner;

[0101] Step five, since the consumers driving different types of vehicles have great differences in charging demand and habits, it is necessary to establish a psychological model for consumers driving different vehicles respectively. In actual engineering, since the buses will be charged by the bus company, the bus company will automatically choose to charge at the valley time (night) of the power grid in order to maximize the profit, so when evaluating the adjustable potential of the charging station, the buses are not considered.

[0102] The following will introduce the consumer psychology model established for taxi users and private car users in detail.

[0103] The user's response behavior to the electricity price is reflected in adjusting the power consumption period according to the price signal, changing the power consumption mode in different time periods, etc. The electricity demand price elasticity coefficient is the relative change of electricity consumption caused by the relative change of electricity price, that is

[0104]

[0105] In the formula: Δq and Δρ are the increments of electricity consumption and electricity price respectively; q and ρ are the electricity consumption and electricity price before the change of electricity price respectively.

[0106] However, the user response model established based on the electricity demand price elasticity matrix cannot reflect the psychology of consumers. The stimulation of the sales price to the user has a minimum just noticeable difference, when the sales price is less than the minimum just noticeable difference, the user basically has no response; and when the sales price is greater than an upper limit value, the user will no longer have more transferable load, and the response ability tends to be saturated.

[0107] Figure 4a 、 Figure 4b Figures are the psychological price response models of taxi users and private car users in the charging station respectively;

[0108] For a charging station, the dispatchable participation rate of electric vehicle load is:

[0109]

[0110] In the formula: ρ ev is the electric vehicle demand response incentive; λ EV,max is the upper limit of the electric vehicle load transfer rate; a EV is the upper limit of the peak-valley electricity price difference dead zone of the electric vehicle load transfer rate; K EV is the slope of its linear zone, since the private car is more sensitive to load change than the taxi, the K EV of the private car is greater than that of the taxi.

[0111] The above psychological price response model considers that the electric vehicle is at the charging station at this time, and does not consider the mutual support phenomenon of adjacent charging stations. When considering mutual support between charging stations, the dead zone interval will also change. This is mainly because the distance between two charging stations causes additional power consumption when the electric vehicle travels from one place to another. From a purely economic point of view, the user needs to be given an incentive greater than the cost of the additional power consumption for the user to participate in the response. In addition, due to the different distances of different charging stations, the slope of the linear interval of the user psychological price response model between the two charging stations will also change. The greater the distance between the two charging stations, the smaller the slope of the linear region of the user psychological price response model.

[0112] When considering cross-charging station scheduling, consumers need to consider not only the cost of additional power consumption, but also the problem of additional time consumption. The additional time consumption includes the travel time and the waiting time of the car arriving at another charging station.

[0113] The additional time consumption time is:

[0114]

[0115] In the formula, is the driving time of the electric vehicle from the i-th charging station to the n-th charging station; is the queuing time at the n-th charging station. The driving time is as follows:

[0116]

[0117] where Ω q is the set of shortest path segments through which the electric vehicle travels to another charging station. L n，q is the distance from the q-th road segment to the n-th charging station. According to the functional relationship between speed and road traffic volume, the speed v q of the q-th road segment can be expressed as:

[0118]

[0119] In the formula, v AVE is the average speed. v q，min is the minimum speed of the electric vehicle. ρ q (t) represents the traffic density of the q-th road segment at time t, ρ q，min and ρ q，max are the minimum and maximum traffic densities of the q-th road segment, respectively. α and β are congestion influence coefficients. N q (t) represents the number of vehicles traveling on the q-th road segment at time t. It is particularly noted that the number of vehicles traveling on the q-th road segment at time t can be obtained from the relevant traffic department or some commercial electronic map company (Gaode Map).

[0120] Waiting time is:

[0121]

[0122] where K q is the queuing coefficient. Q n represents the number of queuing electric vehicles in the charging station. The number of queuing includes the number of original electric vehicles in the charging station and the number of electric vehicles participating in cross-charging station support from other charging stations.

[0123] To calculate the specific additional consumption time, the following method can be used:

[0124] First, calculate the driving time required from the current vehicle node to the destination charging station Considering the total time spent on different routes, the energy is different, and the waiting time after superposition to the charging station is different, so a certain number of driving route solutions is reserved, called K, and finally the best route is selected.

[0125]

[0126] T ij , e ij are the time and energy cost required from road node i to road node j, represent whether the kth route from the starting point to the end point selects the road from road node i to road node j, 0 is not selected, and 1 is selected. V' is a set of all road nodes. e ij can be calculated by the distance L n,q(i→j) from node i to road node j. ω is the coefficient for converting time into cost, so that it can be added with energy cost, so that the comprehensive cost is minimized.

[0127]

[0128] b1 = |K|, b n = |K|, b i = 0 (10)

[0129] The number of vehicles flowing out of the starting point, the number of vehicles flowing into the end point, and the net flow of intermediate nodes is 0 (vehicles must enter from one end and exit from the other end of a road). The continuity of vehicle travel is ensured.

[0130]

[0131] Limit each route to a maximum of one pass, i.e. vehicles cannot repeatedly travel on the road between node i and node j. R is V' excluding the starting node v1 where the vehicle departs and the node (end point) v na set of other road nodes besides the charging stations.

[0132]

[0133] The vehicle must have a certain amount of electricity left when it arrives at the charging station. E1 is the amount of electricity the vehicle has when it departs, E min is the minimum amount of electricity allowed for the vehicle's battery.

[0134] Then the distance traveled by the kth route from the start to the end is

[0135]

[0136] The time taken by the kth route from the start to the end is

[0137]

[0138] The energy cost of the kth route from the start to the end is

[0139]

[0140] According to the following rules, the electric vehicle is preliminarily judged whether it is willing to accept the cross-station scheduling instruction. As long as one of the k routes meets the following three requirements, it is considered that the electric vehicle is preliminarily willing to accept the cross-station scheduling.

[0141] A. When considering cross-charging station scheduling, the electric vehicle driving distance meets the constraint:

[0142]

[0143] In the formula, L max is the maximum driving distance that the electric vehicle is willing to accept from the i charging station to the n charging station.

[0144] B. The incentive I EV of the electric vehicle exceeds the dead zone a EV in the user's psychological electricity price response model and the additional energy cost p EV consumed by cross-station driving.

[0145] I EV > a EV + p EV (17)

[0146] C. The cross-station driving time is less than the user's upper limit of time consumption T maxQ .

[0147]

[0148] According to the preliminary willingness to accept cross-station scheduling of the vehicle and the original number of electric vehicles of the charging station, the waiting time is updated And again, whether to participate in cross-site scheduling, s EV,n = 1 represents willingness to participate.

[0149]

[0150] In the formula is the acceptable maximum waiting time of the electric vehicle in the charging station.

[0151] For the vehicle willing to participate in cross-site scheduling, the response quantity is calculated according to the user psychology price response model:

[0152]

[0153] In the formula: p pv is the electric vehicle participating in demand response incentive; λ EV,max is the upper limit of the electric vehicle load transfer rate; a EV is the upper limit of the electric vehicle load transfer rate to the peak-valley electricity price difference dead zone; p EV is the electric quantity fee consumed by the automobile cross-charging station scheduling; K EV is the slope of the linear region; is the function of converting time into economic cost, is the driving time of the electric vehicle from the i th charging station to the n th charging station; is the queuing time at the n th charging station, and VOT is the psychological value / fee corresponding to the unit time (such as per hour), and the unit is yuan / hour.

[0154] Figure 5a , Figure 5b are respectively the taxi user and private car user psychology price response model display diagram considering cross-regional scheduling in the charging station;

[0155] The predicted value obtained in step four is combined with the consumer psychology model and constraints established by the application to determine the number of various types of vehicles willing to participate in scheduling in each charging station. The number of vehicles willing to participate in scheduling is multiplied by the average charging power of the vehicle, and the flexibility resource adjustment potential of each charging station can be evaluated.

[0156] Aggregation method:

[0157] The charging station is divided into an aggregation body for voltage regulation and an aggregation body for peak regulation. It is assumed that based on the predicted data, the nodes with voltage close to the allowable limit of the distribution network, 0.93 p.u. and 1.07 p.u., are set as nodes to be regulated. The voltage regulation aggregation of the adjustable resources of the charging station is carried out according to the following principles:

[0158] It is assumed that the node to be regulated is i, and the sensitivity coefficient based on the adjustable resource of the charging station is The influence of the unit active power injection of node j on the voltage amplitude of node i (when i=j, the influence of the power injection of the node on the voltage amplitude of the node) is as follows.

[0159]

[0160] In the formula, is the power of the charging station of node j participating in the demand response scheduling. is the cost of the power grid scheduling the demand response of the charging station.

[0161]

[0162]

[0163]

[0164] In the formula, ΔV i，min is the difference between the voltage of the node i to be regulated and the safety boundary. EV is a set composed of charging station nodes. is the scheduling amount of the charging station of node j, is the cross-station mutual assistance amount of node g to node j. is the maximum adjustable amount of the charging station of node j, is the maximum cross-station mutual assistance capability of the charging station of node g to the charging station of node j. is the maximum adjustable capability of the charging station of node g.

[0165] Reserve the voltage regulation resource, and the remaining adjustable resource of the charging station is called a peak regulation resource, and the resource is aggregated again according to the spatial position, response capacity, acceptable scheduling period of the charging station, acceptable advance notice time of the charging station, and historical response reputation index of the charging station.

[0166] Figure 6 is a data flow diagram of the method of the application;

[0167] A load prediction and schedulable resource aggregation system of an electric vehicle charging station based on a CNN-TSMixer model, as shown in Figure 7 , comprising:

[0168] A charging behavior data acquisition module is configured to acquire load data of the charging station.

[0169] A data preprocessing module is configured to eliminate abnormal values in the acquired data and complete part of missing values in the data acquisition.

[0170] A global model training module is configured to train a global model to provide initial training parameters for the training of a local model.

[0171] a local model training module configured to train a prediction model more applicable to each charging station by fine-tuning global model parameters;

[0172] a load prediction module configured to complete an actual prediction task;

[0173] a flexibility resource evaluation module configured to evaluate flexibility resource adjustment potential of each charging station.

[0174] a flexibility resource aggregation module configured to divide flexibility resources into a voltage regulation resource cluster and a peak shaving resource cluster.

[0175] Specifically, the method of the present application is verified.

[0176] Example 1

[0177] The simulation data set of this example uses data of three tram charging stations in a city to simulate and illustrate, wherein the data sampling interval is 1 hour.

[0178] First, according to step one, the data from the data acquisition module is preprocessed. Then, according to step two, one year of charging behavior data is used for training, and half a year of data is used for cross-validation. In all model training stages, the data set is preprocessed according to step one, and the local model training is not repeated. When training the global model, one year of data from the three charging stations is combined in sequence (non-overlapping), and the combination form is as follows: the first 365*24 rows of data are for charging station A, then 365*24 rows of data are for B, and so on. The combined data is input into the proposed CNN-TSMixer model, which automatically extracts features and expands the original data set features. The expanded data set is input into the TSMixer model to train the global model, and after the training is completed, the model parameters with the best performance are saved.

[0179] According to step three, the global model obtained in step two is used to construct a local model applicable to each charging station. The input data for training the local model is the data of the charging station.

[0180] According to step four, the trained local model is used to predict the total load of the charging station and the number of public cars, private cars and taxis in the next hour in a rolling update manner.

[0181] Figure 8 , Figure 9 , Figure 10 The total load of each time period in charging stations A, B and C is shown respectively. In order to intuitively show the accuracy of the prediction algorithm proposed in this paper on the whole day scale, the given graph is the comprehensive result obtained after completing the prediction of all time periods of the day, rather than the single hourly prediction curve.

[0182] According to step five, the user response willingness of each charging station in each time period can be calculated by combining the user psychology electricity price response model and the incentives given to the participants, and the schedulable amount of each charging station in each time period can be evaluated by using the predicted value obtained in step four. The arrangement and distance of charging stations A, B and C are shown in Figure 11 .

[0183] For the sake of brevity, only the schedulable amount of charging station A is evaluated and displayed here, and the schedulable amount of charging stations B and C can be evaluated in the same way as that of A.

[0184] Figure 12 To consider the waiting for charging and transferring to other charging stations for charging, and the situation of other charging stations coming here for reverse power transmission, and to consider the display effect of the above two situations at the same time.

[0185] Figure 13 The display diagram of the clustering result according to the voltage out-of-limit nodes is shown, charging station A is located at node 7, charging station B is located at node 20, and charging station C is located at node 30, the voltage out-of-limit nodes are node 17 and node 18, and the voltage regulation cluster 1, the voltage regulation cluster 2 and the peak regulation cluster are formed respectively.

[0186] Example two

[0187] As shown in Figure 14 , an apparatus for implementing a CNN-TSMixer model-based electric vehicle charging station load prediction and schedulable resource aggregation system includes a prediction device host 12, which includes the following apparatus components:

[0188] One or more processors or processing units 16, system memory 28, and bus 18 connecting different system components.

[0189] Bus 18 represents one or more of several types of bus structures, including a memory bus or memory controller, a peripheral bus, a graphics acceleration port, a processor or local bus using any of a variety of bus architectures. For example, these architectures include, but are not limited to, industry standard architecture (ISA) bus, microchannel architecture (MAC) bus, enhanced ISA bus, video electronics standards association (VESA) local bus, and peripheral component interconnect (PCI) bus.

[0190] The prediction device host 12 typically includes a variety of computer system readable media. These media can be any available media that can be accessed by the prediction device host 12, including both volatile and nonvolatile media, removable and non-removable media.

[0191] System memory 28 can include computer system readable media in the form of volatile memory, such as random access memory (RAM) 30 and / or cache memory 32, and / or non-volatile memory, such as read only memory (ROM) 33. Predictive device host 12 can also include other removable / non-removable, volatile / non-volatile computer system storage media. By way of example only, a hard disk drive 34 can be used for reading from and writing to non-removable, non-volatile magnetic media. While Figure 11 A magnetic hard disk drive, not shown, can be used for reading from, and writing to, a non-removable, non-volatile magnetic media such as a "hard drive". Although not specifically shown, alternate embodiments include optical and / or semiconductor one or more non-removable, non-volatile memory systems (also generally considered to be a storage device). As the term is used herein, a "storage device" is used in its normal sense and can include a disk (magnetic), optical, a semiconductor, or other storage medium. If implemented using a semiconductor-based or other integrated circuit chip, executable code can be provided by fabrication as configuration logic during integrated circuit manufacturing. These elements work to communicate persistent information as data signals among other components set forth herein.

[0192] Device 12 can also communicate with one or more external devices 14 such as a keyboard, a pointing device, a display 24, etc.; one or more devices that enable a user to interact with device 12; and / or one or more devices that enable device 12 to communicate with one or more other computing devices. Such communication can occur via Input / Output (I / O) interface(s) 22. Still yet, device 12 can communicate with one or more networks, such as one or more local area networks (LAN), wide area networks (WAN), telecommunication networks, and / or the Internet, etc., through network interface card(s) 20. As Figure 11 illustrated, network interface card(s) 20 communicate with the other components of predictive device host 12 via bus 18. It should be appreciated that although not shown, other hardware and / or software components could be used in conjunction with device 12. These include, but are not limited to, microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data archival storage systems, etc.

[0193] Processing unit(s) 16 can execute instructions, for example, for the operation of the predictive device host 12 and / or one or more applications 36. These application programs (or simply "applications") can include one or more software programs designed and constructed for implementing embodiments of the present application. Processing unit(s) 16 can also communicate with one or more input and / or output devices, such as display 24 and keyboard 38, via user interface adapter 24 and I / O interface 22. In this regard, processing unit(s) 16 can represent a

[0194] The various embodiments described in this specification are implemented in a progressive manner, each embodiment focusing on the differences from other embodiments, and the same or similar parts between embodiments can be mutually referred to. For the apparatus disclosed by the embodiments, since it corresponds to the method disclosed by the embodiments, the description is relatively simple, and the relevant parts can be referred to the description of the method part.

[0195] The above description of disclosed embodiments enables one of ordinary skill in the art to make or use the application. Various modifications to these embodiments will be readily apparent to those of ordinary skill in the art, and the generic principles defined herein can be applied to other embodiments without departing from the spirit or scope of the application. Therefore, the application is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for load forecasting and schedulable resource aggregation of electric vehicle charging stations based on the CNN-TSMixer model, characterized in that, include: Step 1: Collect load data from multiple charging stations and perform data preprocessing; Step 2: Use the data processed in Step 1 to train a global model applicable to multiple charging stations, and save the parameters of the global model with the best prediction performance; Step 3: Using the global model parameters obtained in Step 2, train a local model suitable for each charging station by adjusting the global model parameters obtained in Step 2. Step 4: Using the local model established in Step 3, predict the power at each charging station and the number of various types of vehicles at each time point using a rolling update method. Step 5: Establish a consumer psychology model and use the predicted values ​​obtained in Step 4 to assess the flexibility and resource adjustment potential of each charging station. Step Six: In accordance with the requirements of power grid voltage regulation and peak shaving services, aggregate flexibility resources into voltage regulation resource clusters and peak shaving resource clusters.

2. The method for load prediction and schedulable resource aggregation of electric vehicle charging stations based on the CNN-TSMixer model according to claim 1, characterized in that, Step two uses the CNN-TS Mixer algorithm to train a global model applicable to all time series, continuously adjusts the model parameters, and saves the model parameters with the best prediction performance to construct local models. The CNN-TSMixer algorithm is as follows: TSMixer extends the functionality of a univariate linear model by adding a cross-variable feedforward layer, thereby forming a linear model capable of handling multivariate predictions; the architecture of TSMixer is relatively simple, based on a multilayer perceptron, and consists of two steps: a mixer layer and a temporal projection. Hybridization layer: The core of temporal and feature mixing; temporal mixing is achieved through fully connected layers, ReLU activation function, and dropout layer. The rows of the input data represent time, and the columns represent features. After transposition, the data is processed to share features in the temporal domain. Before leaving the temporal mixing unit, the matrix is ​​transposed again and sent to the feature mixing unit. The feature mixing unit consists of two MLPs applied to the feature domain. No transposition is required because the features are already on the horizontal axis. The mixer includes a normalization layer and residual connections to facilitate the model learning deeper data representations while maintaining computational efficiency. Temporal projection: responsible for generating predictions. The matrix is ​​transposed again and processed by a fully connected layer. The final output matrix places the features on the horizontal axis and the time step on the vertical axis.

3. The method for load prediction and schedulable resource aggregation of electric vehicle charging stations based on the CNN-TSMixer model according to claim 1, characterized in that, Step two also includes using a CNN network to extract spatial features from a large amount of historical load data, while using CNN to capture spatial features and TSMixer to handle temporal dependencies; using convolution kernels to perform convolution and pooling operations on the load feature dataset to extract numerous features from the load data, and inputting the extracted features along with the load feature dataset into the TSMixer model to train global and local models.

4. The method for load prediction and schedulable resource aggregation of electric vehicle charging stations based on the CNN-TSMixer model according to claim 1, characterized in that, The overall process of global model training is as follows: To accelerate model convergence, the overloaded data is normalized. The normalized load data is input into the established global model for training; The training early stopping algorithm is used to test the trained model through a validation set, while saving the parameters of the model with the best prediction performance.

5. The method for load prediction and schedulable resource aggregation of electric vehicle charging stations based on the CNN-TSMixer model according to claim 1, characterized in that, The overall process of local model training is as follows: To accelerate model convergence, the charging station load data is normalized. The normalized load data is input into the established local model, and the local model loads the parameters trained by the global model, and then training is performed. The training early stopping algorithm is used to test the trained model through a validation set, while saving the parameters of the model with the best prediction performance.

6. The method for load prediction and schedulable resource aggregation of electric vehicle charging stations based on the CNN-TSMixer model according to claim 1, characterized in that, Establishing a consumer psychology model specifically includes: Differentiate user types and build basic response models; Users respond to electricity price changes by adjusting their electricity consumption periods, and the electricity demand price elasticity coefficient is introduced to measure the impact of electricity price changes on electricity consumption. Factors influencing cross-charging station scheduling include the fact that, when considering mutual support between charging stations, electric vehicles will consume additional electricity when traveling between stations, and users will only respond if the electricity price incentive exceeds this cost; the greater the distance between the two stations, the smaller the linear interval slope of the user's psychological response model. Cross-station dispatching needs to consider travel time and queuing time. These time costs need to be converted into users' expected prices, which affect their willingness to respond. A user cross-station route selection model should be constructed to reduce travel time and travel costs. Combining the impact of electricity price changes on electricity consumption, the influencing factors of cross-charging station scheduling, as well as travel time and queuing time, a user psychology electricity price response model considering cross-station scheduling is constructed as a consumer psychology model to describe users' response behavior to electricity price changes. Using the predicted power and number of vehicles for each charging station in the future obtained in step four, and substituting them into the consumer psychology model, the flexibility resource adjustment potential of each charging station is obtained.

7. The method for load prediction and schedulable resource aggregation of electric vehicle charging stations based on the CNN-TSMixer model according to claim 6, characterized in that, The constraints are specifically as follows: When considering cross-charging station scheduling, the electric vehicle's travel distance meets the following constraints: In the formula L max The maximum distance for an electric vehicle from charging station i to charging station n; Electric vehicle charging queue waiting time meets the following constraints: In the formula, T max Q represents the maximum waiting time for an electric vehicle at charging station n.

8. A method for load prediction and schedulable resource aggregation of electric vehicle charging stations based on the CNN-TSMixer model according to claim 6, characterized in that, The user psychology electricity price response model considering electric vehicle cross-charging station scheduling is shown below: In the formula: ρ pv Incentives for electric vehicles to participate in demand response; λ EV,max This represents the upper limit of the load transfer rate for electric vehicles. a EV The upper limit of the dead zone for peak-valley electricity price difference based on the load transfer rate of electric vehicles; p EV The additional electricity cost incurred for vehicle dispatching across charging stations; K EV The slope of its linear region; To convert time into a function of economic cost, Let be the travel time of an electric vehicle from the i-th charging station to the n-th charging station; VOT represents the queuing time at the nth charging station, where VOT is the psychological value / cost per unit time (e.g., per hour), expressed in yuan per hour.

9. The method for load prediction and schedulable resource aggregation of electric vehicle charging stations based on the CNN-TSMixer model according to claim 1, characterized in that, In accordance with the requirements of power grid voltage regulation and peak shaving services, flexibility resources are aggregated into voltage regulation resource clusters and peak shaving resource clusters: Select voltage regulation resource clusters using the following formula: In the formula, The power of charging station j participating in demand response scheduling; The cost of charging stations participating in demand response for power grid dispatching; In the formula, ΔV i，min It is the difference between the voltage at the node i to be regulated and the safety boundary; S EV A set consisting of charging station nodes; Let J be the scheduling quantity for charging station j. For cross-site mutual aid between node g and node j; This represents the maximum adjustable value for charging station j itself. The maximum cross-station mutual assistance capacity of charging station g to charging station j; The maximum adjustable capacity of the charging station at node g; The remaining adjustable resources of charging stations are called peak-shaving resources, which are then aggregated again according to spatial location, response capacity, the time period during which the charging station can accept scheduling, the advance notice time that the charging station can accept, and the historical response reputation indicators of the charging station.

10. A load forecasting and schedulable resource aggregation system for electric vehicle charging stations based on a CNN-TSMixer model, employing the load forecasting and schedulable resource aggregation method for electric vehicle charging stations based on a CNN-TSMixer model as described in any one of claims 1-9, characterized in that... include: The charging behavior data acquisition module is used to collect load data of the charging station; The data preprocessing module is used to remove outliers from the collected data and to fill in missing values ​​in the collected data. The global model training module is used to train the global model and provides initial training parameters for the training of local models. The local model training module is used to train a prediction model that is more suitable for each charging station by fine-tuning the global model parameters. The load forecasting module is used to complete the actual forecasting tasks; The flexibility resource assessment module evaluates the flexibility resource adjustment potential of each charging station. The flexibility resource aggregation module is used to divide flexibility resources into voltage regulation resource clusters and peak shaving resource clusters.

Images