Electric vehicle charging load analysis system based on deep learning and real-time traffic simulation

Abstract

The invention provides an electric vehicle charging load analysis system based on deep learning and real-time traffic simulation. The system comprises an offline simulation deduction module, a deep neural network training module and an online simulation deduction module. An energy consumption model is constructed by collecting an electric vehicle starting toll station record, ETC door frame vehicle passing information and service area charging station data in an expressway, and offline simulation deduction is achieved. The deep neural network adopts a self-attention mechanism model, and combines a multi-head attention mechanism and time sequence coding to predict the EV charging load. The on-line simulation deduction module collects vehicle data in real time and carries out prediction based on real-time traffic simulation software. And performing weighted analysis and dynamic adjustment on simulation and deep learning prediction results through a feedback mechanism so as to improve the prediction precision. Through the distributed calculation and real-time data processing technology, mass data can be efficiently processed, accurate charging demand prediction is provided, and the method adapts to dynamic changes in different scenes.

Description

Technical Field

[0001] The present invention relates to the field of electric vehicle load analysis, and in particular to an electric vehicle charging load analysis system based on deep learning and real-time traffic simulation. Background Art

[0002] As a kind of power load, the charging load of electric vehicles (EVs) will inevitably have a certain impact on the stability of power systems in some regions as the number of EVs increases and large-scale EVs are connected to the power grid. The time series method is the most common EV load analysis method. It takes some observations from the load data in chronological order, finds the relevant feature vectors, and assumes that this feature can continue to persist, so as to complete the analysis of future loads. The regression analysis prediction method does not require a large amount of historical load data support. It is a statistical method with a simple calculation framework, which can construct the correlation between the influencing factors and load values ​​of load prediction. The charging data of large-scale charging stations is difficult to obtain directly, so in recent years, the origin-destination (OD) method is mainly used to establish travel demand and spatiotemporal models in combination with transportation networks. Considering dynamic traffic information, the OD matrix analysis method and Dijkstra path search algorithm are often used to plan the driving path of electric vehicles, simulate their dynamic driving process and charging behavior, and analyze the probability density function of EV charging load in combination with travel chains.

[0003] However, the prediction accuracy of traditional analysis methods is increasingly insufficient, because the overall historical state set by traditional analysis methods is always stable, ignoring the continuous development of society. The continuous increase in the number of EVs in the future may make it more difficult to predict the load; the charging data of large-scale charging stations is difficult to obtain directly, which limits the effectiveness of the analysis method. The data of charging stations may be restricted by commercial sensitivity or privacy issues, making it impossible for researchers to directly use actual charging data for analysis; it is crucial to accurately predict EV charging demand and real-time monitoring of traffic flow in highway systems; existing methods usually lack comprehensive systems and methods for EV charging load prediction and road network traffic simulation. Therefore, it is particularly important to analyze EV charging load. Summary of the invention

[0004] The purpose of the present invention is to solve the problem of how to establish an accurate energy consumption model through historical data and traffic flow model, and to analyze charging load prediction and traffic flow simulation in real time based on real-time traffic simulation data and ETC gantry vehicle passing record data.

[0005] To achieve the above objectives, the present invention provides an electric vehicle charging load analysis system based on deep learning and real-time traffic simulation, which is as follows:

[0006] An electric vehicle charging load analysis system based on deep learning and real-time traffic simulation, comprising: an offline simulation deduction module, a deep neural network training module, an online simulation deduction module, and a data analysis module;

[0007] The operations performed by the offline simulation module include:

[0008] Collect the electric vehicle (EV) starting toll station record data, ETC gantry vehicle passing information, and service area charging station data on provincial highways; based on the collected data, build a highway network topology model and a revised energy consumption model;

[0009] In order to ensure the data quality and stability of the offline simulation module, data verification, exception handling and real-time monitoring measures are adopted; among them, data verification performs multiple verifications on the ETC gantry vehicle passing information, EV starting toll station records and service area data to ensure the accuracy and integrity of the data; multiple verifications include: verifying whether the timestamp conforms to the standard time format (ISO 8601); verifying whether the vehicle speed data exceeds a reasonable upper limit, and verifying whether the toll amount is a negative value; comparing the ETC gantry vehicle passing information and toll station records to ensure that the time and vehicle ID information of each vehicle passing the toll station and the ETC gantry are consistent. If there is any inconsistency, it is marked as potential erroneous data;

[0010] Abnormality handling: Introduce abnormality detection algorithms to monitor the collected data in real time, discover and handle abnormal values ​​and missing values; use regression analysis technology to fill in missing data to ensure data continuity and consistency;

[0011] Real-time monitoring: Establish a data monitoring system to monitor any abnormal situations in the data collection and processing process in real time, and issue warnings and handle them in a timely manner (mark abnormal data and filter abnormal data) to ensure the stability and reliability of data flow;

[0012] 1. A macro traffic flow model (Lighthill-Whitham-Richards, LWR) is established to reflect the real-time status of the energy consumption model; that is, when the macro traffic flow model shows that the vehicle is in a congested state, the low speed or frequent parking of the vehicle will increase the energy consumption, while when the vehicle is in a high-speed driving state, the energy consumption will be relatively stable; the expression of the macro traffic flow model is:

[0013]

[0014] In the formula, represents the vehicle density on road section x at time t, Indicates traffic flow;

[0015] 2. Establish an energy consumption model, the expression is as follows:

[0016] P 消耗 =F 阻力总 ×v

[0017] Where P 消耗 is the energy consumption (power) per unit time (one hour), F 阻力总 is the total resistance of the vehicle, v is the speed of the vehicle; where F 阻力总 is the resultant force of all resistances, including rolling resistance F 滚动 , air resistance F 空气 ; Rolling resistance F 滚动 The expression is as follows:

[0018] F 滚动 =C r ×m×g

[0019] In the formula, C r is the rolling resistance coefficient, m is the vehicle mass, and g is the acceleration due to gravity;

[0020] Air resistance F 空气 The expression is as follows:

[0021]

[0022] Where ρ is the air density, A is the effective cross-sectional area of ​​the vehicle, and C d is the aerodynamic drag coefficient, v is the speed of the vehicle;

[0023] 3. Further optimize the energy consumption model through electric vehicle (EV) dynamic equations;

[0024] Specifically, the electric vehicle (EV) dynamics equation is expressed as follows:

[0025] F 净 =F 驱动 -F 阻力总

[0026] In the formula, F 净 is the net force acting on the EV, causing the EV to accelerate; F 驱动 is the driving force provided by the motor, which can be expressed as the motor torque T 电机 The traction force is divided by the wheel radius r and is expressed as follows:

[0027]

[0028] In summary, the basic driving equation of electric vehicles (EVs) can be obtained:

[0029]

[0030] In the formula, m×a represents the resistance, and a is the acceleration of the vehicle;

[0031] 4. In order to consider the impact of the energy recovery system of electric vehicles (EVs) on the basic driving equation, an additional coefficient is introduced to represent the efficiency of the energy recovery system;

[0032] Energy recovery system efficiency: Electric vehicles (EVs) are equipped with a braking energy recovery system to convert the kinetic energy generated during braking into electrical energy and store it in the battery. However, the recovery system is not 100% efficient, so the efficiency of energy recovery η needs to be considered. 回收 ; If the efficiency of braking energy recovery is η 回收 , then the actual recovered energy is the braking energy multiplied by η 回收 ;

[0033] Therefore, the revised energy consumption model is expressed as:

[0034] P 消耗 =F 阻力总 ×v+P 加速 +P 制动能量回收

[0035] Where P 加速 is the energy required for vehicle acceleration, P 制动能量回收 is the energy recovered by braking energy;

[0036] The offline simulation module simulates the driving trajectory of electric vehicles (EVs) according to the road network topology model and the revised energy consumption model, including the departure point, destination, route sections and stop time, and obtains the energy consumption and corresponding charging requirements of each electric vehicle (EV) during driving;

[0037] Finally, the offline simulation module generates charging load demand data for each service area based on the simulated driving trajectory, the energy consumption of electric vehicles (EVs) during driving, and the corresponding charging demand. This data is used to reflect the amount of charging services that the service area should provide in different time periods. This data is used as training data for the deep learning model and can also be used for charging load prediction in the subsequent online simulation module.

[0038] Furthermore, the operations performed by the deep neural network training module include: data preparation and preprocessing and designing a neural network model. The specific process is as follows:

[0039] Combined with the revised energy consumption model, the Simulation of Urban MObility (SUMO) software is used to simulate the driving process of electric vehicles under different road sections, speeds and traffic flow conditions. By inputting the speed, acceleration, road conditions and slope information of the simulated EV, the energy consumption forecast of each vehicle is obtained. Then, by integrating the data of all vehicles, the charging load demand of each service area is calculated.

[0040] Specifically, data preparation and preprocessing require that 80% of the energy consumption data generated by the modified energy consumption model be used for the training set and 20% for the test set; the training set is used to train the model, and by adjusting the model parameters, it can learn the relationship between the input data and the target output, and the test set is used to test the generalization ability of the model, that is, to test the performance of the model on data that has never been seen;

[0041] In the data preparation stage, deviation standardization (Min-Max normalization) is performed, that is, all numerical features are normalized to the interval [0,1] or [-1,1] to ensure that features of different dimensions can participate in model training together, so as to be input into the neural network model for training; the expression is as follows:

[0042]

[0043] Where X is any value in the energy consumption data, X norm is the normalized data value, X min is the minimum value in the energy consumption data, X max is the maximum value in the energy consumption data;

[0044] Furthermore, a deep neural network model is designed through the attention mechanism to predict charging load and identify electric vehicle (EV) types. Multi-Head Attention and Temporal Fusion Transformer (TFT) are used to improve the performance of the model, as follows:

[0045] The expression of the multi-head attention mechanism is as follows:

[0046]

[0047] Where Q, K, and V are query, key, and value matrices, respectively. is the scaling factor;

[0048] The outputs of all attention heads are concatenated and the final multi-head attention output is obtained through linear transformation. The expression is as follows:

[0049] MultiHead(Q,K,V)=Concat(head1,head2,...,head h )W O

[0050] Where W O is the output linear transformation weight matrix;

[0051] The goal of the temporal fusion transformer (TFT) is to transform energy consumption data into a format suitable for model processing and capture the changing trend in the time dimension. The specific operations include: temporal encoding of energy consumption data, adding timestamp, date, and hour information to enhance the model's perception of time characteristics;

[0052] Temporal Fusion Transformer (TFT) includes: input feature processing layer, separation feature encoding layer, long-term and short-term attention layer, and dynamic convolution layer; the details are as follows:

[0053] 1. Input feature processing layer: the time series encoded data will be used as input and enter the input feature processing stage;

[0054] Specifically, the input multidimensional data features (vehicle speed, vehicle density, traffic flow, and energy consumption data) are preprocessed to meet the requirements of the model;

[0055] Preprocessing includes: standardization, normalization, and denoising preprocessing operations on multidimensional data features;

[0056] 2. Separate feature encoding layer: split the processed input features and encode them independently, so that the model can better learn the specific information of different features;

[0057] Specifically, the processed input features are separated using an embedding layer to capture the pattern of each feature;

[0058] Furthermore, the function of the embedding layer is to map the features of the separation feature encoding layer into dense vectors that can capture the semantic relationship between the features. The specific steps are as follows:

[0059] Define the basic parameters of the embedding layer, including: vocabulary size and embedding dimension:

[0060] Vocabulary Size: The total number of categories, which is the same as the number of electric vehicles (EVs) collected;

[0061] Embedding Dimension: The dimension of the vector space to which each category is mapped. The embedding dimension is selected to be the same as the mapping vector of each category.

[0062] Initialize the embedding matrix;

[0063] The core of the embedding layer is an embedding matrix, which is randomly initialized at the beginning of training and optimized by gradient descent during training. The dimension expression of the embedding matrix is ​​as follows:

[0064] Embedding Matrix Size=Vocabulary Size×Embedding Dimension

[0065] Where, Embedding Matrix Size represents the size of the embedding matrix, Vocabulary Size represents the vocabulary size, and Embedding Dimension represents the embedding dimension;

[0066] Map each electric vehicle (EV) category to the corresponding vector in the embedding matrix to achieve category indexing; that is, each category is assigned a unique index, and the embedding vector is searched for corresponding to different electric vehicle (EV) types using the numbers 0, 1, 2, 3, ..., m (m is a natural number), and so on. The corresponding embedding vector is searched in the embedding matrix through the index, and the embedding vector corresponding to the number can be represented as the corresponding row of the embedding matrix;

[0067] Train the embedding layer;

[0068] During the model training process, the embedding matrix is ​​optimized through the back-propagation algorithm, and the weights of the embedding layer (embedding matrix) are updated according to the gradient of the loss function to better capture the potential relationship of the category features: Forward propagation, in the forward propagation stage, the category index is passed through the embedding layer to find the corresponding embedding vector, and these embedding vectors are used as the input of the model;

[0069] Back-propagation algorithm, that is, in the back-propagation phase, the gradient of the loss function is calculated and the weights in the embedding matrix are updated;

[0070] Application embedding layer;

[0071] After training is completed, the embedding layer can be used to convert new category features into their corresponding embedding vectors, so that the model can process new data; in the prediction stage, the new category features are searched for corresponding embedding vectors through the embedding layer, and these embedding vectors are provided as input to the model for prediction;

[0072] In order to improve the generalization ability of the model, after the initial embedding layer processing, further feature extraction can be performed through multiple embedding layers. Each layer will be processed by a nonlinear activation function to capture deeper feature associations.

[0073] 3. Long-term and short-term attention layer; that is, the feature sequence after separation feature encoding enters the long-term and short-term attention layer to focus on which time points are most important for current energy consumption or traffic prediction;

[0074] Specifically, for the encoded independent features, the attention mechanism is used to capture the dependencies of different time steps. The long-term and short-term attention mechanisms are combined to focus on the important information in the input sequence, better handle long-term dependencies, and enhance the understanding of the association between sequences.

[0075] The long and short-term attention expressions are as follows:

[0076]

[0077] Where P is the position encoding matrix;

[0078] 4. Dynamic convolution layer: further extract local features from independent features and dynamically adjust the convolution kernel according to the context to improve the response ability to time series changes;

[0079] Specifically, dynamic convolution is used to extract the temporal features from the independent features. The size and weight of the convolution kernel will be dynamically adjusted according to different moments to adapt to the characteristics of different time periods.

[0080] The expression is as follows:

[0081] Fused Features=Concat(Temporal Features,Static Features)W F

[0082] Where W F is the linear transformation matrix of feature fusion;

[0083] 5. Output the final prediction result through the fully connected layer; the output of the embedding layer is connected to the fully connected layer (FullyConnected Layer), and the final model output is generated through nonlinear transformation; for example, energy consumption prediction value or traffic flow prediction value;

[0084] 6. Model optimization: adjust the parameters of the embedding layer and subsequent network layers by calculating the error and back-propagating the error, and gradually optimize the prediction results;

[0085] In practice, the ETC gantry vehicle passing records and charging pile data may fail or be incomplete due to communication failures and environmental factors affecting some detectors. In this case, the adaptive spatiotemporal interpolation and generative adversarial network (GANs) fusion method is used to ensure the accuracy and stability of the prediction. In response to possible data failure, the online simulation module calculates the ETC gantry vehicle passing information and the data of the failed charging piles based on the existing OD traffic through simulation and prediction, thereby filling the gaps in data missing, so that the prediction deviation is not too large or divergent, and ensuring the integrity and reliability of the prediction model. The online simulation module can effectively reduce the prediction deviation caused by data failure and improve the stability and accuracy of the system in practical applications.

[0086] The online simulation module takes into account vehicle type, driving trajectory, and road congestion to achieve accurate charging demand prediction. It uses real-time data and prediction results to conduct highly refined real-time traffic simulation to simulate the actual driving process of EVs on the road network. By dynamically adjusting the traffic flow and charging demand model, it can accurately predict and reflect the charging demand of each service area, and realize real-time monitoring and optimization of road network traffic flow.

[0087] In order to cope with the growing amount of data and frequently updated environmental conditions, the online simulation module adopts the following measures:

[0088] Distributed computing architecture: Introducing distributed computing architecture and using Hadoop big data processing technology to ensure that the system can efficiently process massive amounts of data;

[0089] Real-time data processing: Apache Flink technology is used to achieve rapid processing and response of real-time data, ensuring the real-time and stability of the system;

[0090] Elastic scalability: The system is designed with good elastic scalability, and can dynamically adjust computing resources according to actual needs to ensure efficient operation during peak hours;

[0091] Fault-tolerance mechanism: Introduce a fault-tolerance mechanism to ensure that the system can still run stably in the event of partial node failure or data loss, thereby improving system reliability;

[0092] Finally, the prediction results provided by the online simulation module provide important decision-making support for traffic management and energy planning departments, which can be used to optimize the supply strategy of charging facilities, improve the utilization efficiency and service level of charging infrastructure, and promote the development of intelligent transportation systems towards smart cities.

[0093] The data analysis module performs the following tasks:

[0094] Perform weighted analysis on the prediction results of the online simulation module and the prediction data of the deep neural network training module, and introduce a feedback mechanism to evaluate the prediction result deviation and obtain a comprehensive prediction value;

[0095] Specific methods include: prediction result weighting method and feedback mechanism;

[0096] The expression of the prediction result weighting method is as follows:

[0097] Weighted analysis = ω 仿真 ×Prediction value 仿真 +ω 深度学习 ×Prediction value 深度学习

[0098] In the formula, ω 仿真 is the weight of the online simulation module; ω 深度学习 is the weight of the deep neural network training module; the predicted value 仿真 It is the predicted value obtained by the online simulation module; predicted value 深度学习 is the predicted value obtained by the deep neural network training module;

[0099] In the present invention, a dynamic weight adjustment mechanism based on prediction error is used to evaluate the deviation of historical prediction results through actual detection values, and the confidence (weight) of the weight estimation algorithm and feedback mechanism is estimated according to the deviation value; the specific steps are as follows:

[0100] 1. Save the prediction results of the online simulation module and the deep neural network training module as historical values;

[0101] 2. After a certain time interval (one week), collect the actual measurement values ​​of the charging load in the actual service area;

[0102] 3. Compare the EV starting toll station record data, ETC gantry vehicle passing information, and service area charging station data collected by the deep neural network training module with the actual measured values ​​of EV starting toll station record data, ETC gantry vehicle passing information, and service area charging station data collected by the online simulation deduction module to calculate the prediction deviation;

[0103] The purpose of calculating the prediction deviation is to evaluate the difference between the predicted values ​​of the online simulation module and the deep neural network training module and the actual measured values, and then adjust the weights of the two to optimize the future prediction accuracy;

[0104] The prediction deviation is calculated by the mean square error (MSE). First, the charging load measurement value of the actual service area is collected, which is recorded as A(t); the comprehensive prediction value of the online simulation deduction module is recorded as S(t); and the comprehensive prediction value of the deep neural network training module is recorded as D(t).

[0105] Calculation error: The error of the online simulation module and the deep neural network training module is calculated by the mean square error (MSE), and the formula is as follows:

[0106]

[0107] Where P(t) is the predicted value, i.e., the online simulation module S(t) or the deep neural network training module D(t), and A(t) is the actual value, t is the time, and n is the total number of time points. MSE calculates the average of the squared difference between the predicted value and the actual value, which can emphasize larger errors.

[0108] After obtaining the error, the prediction values ​​of the online simulation deduction module and the deep neural network training module are combined by weighted summation to obtain the comprehensive prediction value; the expression is as follows:

[0109] P total (t) = ω S ·S(t)+ω D ·D(t)

[0110] In the formula, the weight of the online simulation module is ω S , the weight of the deep neural network training module is ω D , then the comprehensive prediction value P total (t);

[0111] The purpose of updating weights is to ensure that the prediction results in different time periods and under different conditions can reflect the latest prediction accuracy. After each cycle (one week), the MSE is recalculated based on the latest actual measurement values ​​and historical prediction results, and the weights are adjusted.

[0112] According to the evaluation results, the weights of the online simulation deduction module and the deep neural network training module are adjusted to enable the present invention to achieve accurate prediction.

[0113] Beneficial effects of the present invention

[0114] 1. The present invention combines offline simulation deduction and online real-time monitoring to overcome the problem of insufficient prediction accuracy of traditional methods in dynamic environments.

[0115] 2. The present invention utilizes an online simulation module to make the prediction results closer to the actual situation, and can provide accurate data support for the decision-making of the transportation and energy departments.

[0116] 3. The present invention introduces a deep neural network training module to improve the accuracy of charging load prediction.

[0117] 4. The present invention proposes a data compensation method to solve the problem of failure or incomplete data caused by communication failure and environmental influence of some detectors.

[0118] 5. The present invention effectively combines the two prediction methods and dynamically optimizes the prediction results through weighted averaging and weight estimation mechanism methods to improve the prediction accuracy and stability of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0119] Figure 1 is a flow chart of the present invention;

[0120] Figure 2 Schematic diagram of the working of the modules of the present invention, wherein part (a) is a schematic diagram of the working of the offline simulation deduction module; part (b) is a schematic diagram of the working of the deep neural network training module; part (c) is a schematic diagram of the working of the online simulation deduction module; and part (d) is a schematic diagram of the working of the data analysis module;

[0121] Figure 3 It is the topological model of Shandong highway network;

[0122] Figure 4 The following are four charging stations with different traffic distributions throughout the day. DETAILED DESCRIPTION

[0123] The present invention is further described in detail below in conjunction with specific embodiments;

[0124] The ETC gantry vehicle passing records within 24 hours from April 4, 2012 to April 5, 2012 on Shandong Expressway Network were used. Since the data of charging stations in service areas are difficult to obtain in China, the open data set on the official website of California Institute of Technology was used to obtain the corresponding charging stations in the service areas in the region.

[0125] like Figure 1 As shown, an electric vehicle charging load analysis system based on deep learning and real-time traffic simulation includes: an offline simulation deduction module, a deep neural network training module, an online simulation deduction module, and a data analysis module;

[0126] like Figure 2 As shown in part (a), the operations performed by the offline simulation module include:

[0127] Collect the electric vehicle (EV) starting toll station record data, ETC gantry vehicle passing information, and service area charging station data on provincial highways; based on the collected data, build a highway network topology model and a revised energy consumption model;

[0128] In order to ensure the data quality and stability of the offline simulation module, data verification, exception handling and real-time monitoring measures are adopted;

[0129] Data verification is performed during the data collection phase, i.e., multiple verifications are performed on the ETC gantry vehicle passing information, EV starting toll station records, and service area data to ensure the accuracy and completeness of the data; multiple verifications include: ETC gantry vehicle passing information should contain vehicle ID, timestamp, and location, and the timestamp should conform to the standard time format (ISO 8601); vehicle speed data cannot exceed a physically reasonable upper limit, and the toll amount cannot be a negative value; compare the ETC gantry vehicle passing information and toll station records to ensure that the time and vehicle ID information of each vehicle passing through the toll station and the ETC gantry are consistent. If there is any inconsistency, it is marked as potential erroneous data;

[0130] Abnormality handling: Introduce abnormality detection algorithms to monitor the collected data in real time, discover and handle abnormal values ​​and missing values; use regression analysis technology to fill in missing data to ensure data continuity and consistency;

[0131] Real-time monitoring: Establish a data monitoring system to monitor any abnormal situations in the data collection and processing process in real time, and issue warnings and handle them in a timely manner (mark abnormal data and filter abnormal data) to ensure the stability and reliability of data flow;

[0132] The total length of Shandong's expressway network is more than 4,700 kilometers; there are 49 interchanges; there are 309 toll booths that have access to the networked toll collection system; and there are 172 service areas and parking lots.

[0133] like Figure 3 As shown in Figure 1, the highway network topology model includes: ETC gantry vehicle passing information, spatial relationship and connectivity between toll stations and service areas, as well as the distance and path between each node (gantry, toll station, service area);

[0134] The steps to build the revised energy consumption model are as follows:

[0135] 1. A macro traffic flow model (Lighthill-Whitham-Richards, LWR) is established to reflect the real-time status of the energy consumption model; that is, when the macro traffic flow model shows that the vehicle is in a congested state, the low speed or frequent parking of the vehicle will increase the energy consumption, while when the vehicle is in a high-speed driving state, the energy consumption will be relatively stable; the expression of the macro traffic flow model is:

[0136]

[0137] In the formula, represents the vehicle density on road section x at time t, Indicates traffic flow;

[0138] 2. Establish an energy consumption model, the expression is as follows:

[0139] P消耗 =F 阻力总 ×v

[0140] Where P 消耗 is the energy consumption (power) per unit time (one hour), F 阻力总 is the total resistance of the vehicle, v is the speed of the vehicle; where F 阻力总 is the resultant force of all resistances, including rolling resistance F 滚动 , air resistance F 空气 ; Rolling resistance F 滚动 The expression is as follows:

[0141] F 滚动 =C r ×m×g

[0142] In the formula, C r is the rolling resistance coefficient, m is the vehicle mass, and g is the acceleration due to gravity;

[0143] Air resistance F 空气 The expression is as follows:

[0144]

[0145] Where ρ is the air density, A is the effective cross-sectional area of ​​the vehicle, and C d is the aerodynamic drag coefficient, v is the speed of the vehicle;

[0146] 3. Further optimize the energy consumption model through electric vehicle (EV) dynamic equations;

[0147] Specifically, the electric vehicle (EV) dynamics equation is expressed as follows:

[0148] F 净 =F 驱动 -F 阻力总

[0149] In the formula, F 净 is the net force acting on the EV, causing the EV to accelerate; F 驱动 is the driving force provided by the motor, which can be expressed as the motor torque T 电机 The traction force is divided by the wheel radius r and is expressed as follows:

[0150]

[0151] In summary, the basic driving equation of electric vehicles (EVs) can be obtained:

[0152]

[0153] In the formula, m×a represents the resistance, and a is the acceleration of the vehicle;

[0154] The basic driving equation of electric vehicles (EV) can be used to analyze the energy consumption and performance characteristics of electric vehicles (EV) under different speeds, road conditions and driving conditions, and provide an important theoretical basis for the design, optimization and energy management of electric vehicles (EV);

[0155] 4. In order to consider the impact of the energy recovery system of electric vehicles (EVs) on the basic driving equation, an additional coefficient is introduced to represent the efficiency of the energy recovery system;

[0156] Energy recovery system efficiency: Electric vehicles (EVs) are equipped with a braking energy recovery system to convert the kinetic energy generated during braking into electrical energy and store it in the battery. However, the recovery system is not 100% efficient, so the efficiency of energy recovery η needs to be considered. 回收 ; If the efficiency of braking energy recovery is η 回收 , then the actual recovered energy is the braking energy multiplied by η 回收 ;

[0157] Therefore, the revised energy consumption model is expressed as:

[0158] P 消耗 =F 阻力总 ×v+P 加速 +P 制动能量回收

[0159] Where P 加速 is the energy required for vehicle acceleration, P 制动能量回收 is the energy recovered by braking energy;

[0160] The offline simulation module simulates the driving trajectory of electric vehicles (EVs) according to the road network topology model and the revised energy consumption model, including the departure point, destination, route sections and stop time, and obtains the energy consumption and corresponding charging requirements of each electric vehicle (EV) during driving;

[0161] Finally, the offline simulation module generates charging load demand data for each service area based on the simulated driving trajectory, the energy consumption of electric vehicles (EVs) during driving, and the corresponding charging demand. This data is used to reflect the amount of charging services that the service area should provide in different time periods. This data is used as training data for the deep learning model and can also be used for charging load prediction in the subsequent online simulation module.

[0162] Further, such as Figure 2 As shown in part (b), the operations performed by the deep neural network training module include: data preparation and preprocessing and designing the neural network model. The specific process is as follows:

[0163] Combined with the revised energy consumption model, the Simulation of Urban MObility (SUMO) software is used to simulate the driving process of electric vehicles under different road sections, speeds and traffic flow conditions. By inputting the speed, acceleration, road conditions and slope information of the simulated EV, the energy consumption forecast of each vehicle is obtained. Then, by integrating the data of all vehicles, the charging load demand of each service area is calculated.

[0164] Specifically, data preparation and preprocessing require that 80% of the energy consumption data generated by the modified energy consumption model be used for the training set and 20% for the test set; the training set is used to train the model, and by adjusting the model parameters, it can learn the relationship between the input data and the target output, and the test set is used to test the generalization ability of the model, that is, to test the performance of the model on data that has never been seen;

[0165] In the data preparation stage, deviation standardization (Min-Max normalization) is performed, that is, all numerical features are normalized to the interval [0,1] or [-1,1] to ensure that features of different dimensions can participate in model training together, so as to be input into the neural network model for training; the expression is as follows:

[0166]

[0167] Where X is any value in the energy consumption data, X norm is the normalized data value, X min is the minimum value in the energy consumption data, X max is the maximum value in the energy consumption data;

[0168] Furthermore, a deep neural network model is designed through the attention mechanism to predict charging load and identify electric vehicle (EV) types. Multi-Head Attention and Temporal Fusion Transformer (TFT) are used to improve the performance of the model, as follows:

[0169] The expression of the multi-head attention mechanism is as follows:

[0170]

[0171] Where Q, K, and V are query, key, and value matrices, respectively. is the scaling factor;

[0172] The outputs of all attention heads are concatenated and the final multi-head attention output is obtained through linear transformation. The expression is as follows:

[0173] MultiHead(Q,K,V)=Concat(head1,head2,...,head h )W O

[0174] Where W O is the output linear transformation weight matrix;

[0175] The goal of the temporal fusion transformer (TFT) is to transform energy consumption data into a format suitable for model processing and capture the changing trend in the time dimension. The specific operations include: temporal encoding of energy consumption data, adding timestamp, date, and hour information to enhance the model's perception of time characteristics;

[0176] Temporal Fusion Transformer (TFT) includes: input feature processing layer, separation feature encoding layer, long-term and short-term attention layer, and dynamic convolution layer; the details are as follows:

[0177] 1. Input feature processing layer: the time series encoded data will be used as input and enter the input feature processing stage;

[0178] Specifically, the input multidimensional data features (vehicle speed, vehicle density, traffic flow, and energy consumption data) are preprocessed to meet the requirements of the model;

[0179] Preprocessing includes: standardization, normalization, and denoising preprocessing operations on multidimensional data features;

[0180] 2. Separate feature encoding layer: split the processed input features and encode them independently, so that the model can better learn the specific information of different features;

[0181] Specifically, the processed input features are separated using an embedding layer to capture the pattern of each feature;

[0182] Furthermore, the function of the embedding layer is to map the features of the separation feature encoding layer into dense vectors that can capture the semantic relationship between the features. The specific steps are as follows:

[0183] Define the basic parameters of the embedding layer, including: vocabulary size and embedding dimension:

[0184] Vocabulary Size: The total number of categories, which is the same as the number of electric vehicles (EVs) collected;

[0185] Embedding Dimension: The dimension of the vector space to which each category is mapped. The embedding dimension is selected to be the same as the mapping vector of each category.

[0186] Initialize the embedding matrix;

[0187] The core of the embedding layer is an embedding matrix, which is randomly initialized at the beginning of training and optimized by gradient descent during training. The dimension expression of the embedding matrix is ​​as follows:

[0188] EmbeddingMatrix Size=Vocabulary Size×EmbeddingDimension

[0189] Where, Embedding Matrix Size represents the size of the embedding matrix, Vocabulary Size represents the vocabulary size, and Embedding Dimension represents the embedding dimension;

[0190] Map each electric vehicle (EV) category to the corresponding vector in the embedding matrix to achieve category indexing; that is, each category is assigned a unique index, and the embedding vector is searched for corresponding to different electric vehicle (EV) types using the numbers 0, 1, 2, 3, ..., m (m is a natural number), and so on. The corresponding embedding vector is searched in the embedding matrix through the index, and the embedding vector corresponding to the number can be represented as the corresponding row of the embedding matrix;

[0191] Train the embedding layer;

[0192] During the model training process, the embedding matrix is ​​optimized through the back-propagation algorithm, and the weights of the embedding layer (embedding matrix) are updated according to the gradient of the loss function to better capture the potential relationship of the category features: Forward propagation, in the forward propagation stage, the category index is passed through the embedding layer to find the corresponding embedding vector, and these embedding vectors are used as the input of the model;

[0193] Back-propagation algorithm, that is, in the back-propagation phase, the gradient of the loss function is calculated and the weights in the embedding matrix are updated;

[0194] Application embedding layer;

[0195] After training is completed, the embedding layer can be used to convert new category features into their corresponding embedding vectors, so that the model can process new data; in the prediction stage, the new category features are searched for corresponding embedding vectors through the embedding layer, and these embedding vectors are provided as input to the model for prediction;

[0196] In order to improve the generalization ability of the model, after the initial embedding layer processing, further feature extraction can be performed through multiple embedding layers. Each layer will be processed by a nonlinear activation function to capture deeper feature associations.

[0197] 3. Long-term and short-term attention layer; that is, the feature sequence after separation feature encoding enters the long-term and short-term attention layer to focus on which time points are most important for current energy consumption or traffic prediction;

[0198] Specifically, for the encoded independent features, the attention mechanism is used to capture the dependencies of different time steps. The long-term and short-term attention mechanisms are combined to focus on the important information in the input sequence, better handle long-term dependencies, and enhance the understanding of the association between sequences.

[0199] The long and short-term attention expressions are as follows:

[0200]

[0201] Where P is the position encoding matrix;

[0202] 4. Dynamic convolution layer: further extract local features from independent features and dynamically adjust the convolution kernel according to the context to improve the response ability to time series changes;

[0203] Specifically, dynamic convolution is used to extract the temporal features from the independent features. The size and weight of the convolution kernel will be dynamically adjusted according to different moments to adapt to the characteristics of different time periods.

[0204] The expression is as follows:

[0205] Fused Features=Concat(Temporal Features,Static Features)W F

[0206] Where W F is the linear transformation matrix of feature fusion;

[0207] 5. Output the final prediction result through the fully connected layer; the output of the embedding layer is connected to the fully connected layer (FullyConnected Layer), and the final model output is generated through nonlinear transformation; for example, energy consumption prediction value or traffic flow prediction value;

[0208] 6. Model optimization: adjust the parameters of the embedding layer and subsequent network layers by calculating the error and back-propagating the error, and gradually optimize the prediction results;

[0209] like Figure 2As shown in part (c), in response to the existing data failure or incompleteness, the online simulation module performs the following operations: Based on the existing OD traffic, the ETC gantry vehicle passing information and the data of the failed charging pile are simulated and predicted, thereby filling the gaps in data missing, so that the prediction deviation is not too large or divergent, and the integrity and reliability of the prediction model are ensured. The online simulation module can effectively reduce the prediction deviation caused by data failure and improve the stability and accuracy of the system in practical applications;

[0210] The online simulation module takes into account vehicle type, driving trajectory, and road congestion to achieve accurate charging demand prediction. It uses real-time data and prediction results to conduct highly refined real-time traffic simulation to simulate the actual driving process of EVs on the road network. By dynamically adjusting the traffic flow and charging demand model, it can accurately predict and reflect the charging demand of each service area, and realize real-time monitoring and optimization of road network traffic flow.

[0211] In order to cope with the growing amount of data and frequently updated environmental conditions, the online simulation module adopts the following measures:

[0212] Distributed computing architecture: Introducing distributed computing architecture and using Hadoop big data processing technology to ensure that the system can efficiently process massive amounts of data;

[0213] Real-time data processing: Apache Flink technology is used to achieve rapid processing and response of real-time data, ensuring the real-time and stability of the system;

[0214] Elastic scalability: The system is designed with good elastic scalability, and can dynamically adjust computing resources according to actual needs to ensure efficient operation during peak hours;

[0215] Fault-tolerance mechanism: Introduce a fault-tolerance mechanism to ensure that the system can still run stably in the event of partial node failure or data loss, thereby improving system reliability;

[0216] Compared with traditional methods, the online simulation module is innovative and efficient. It uses deep learning technology and real-time data processing to achieve high accuracy and real-time performance of charging load prediction.

[0217] like Figure 4 As shown in the figure, four charging stations with different traffic distribution throughout the day are selected for visualization; it can be seen from the figure:

[0218] The daily traffic volume of this section of Service Area 1 is 11,322 vehicles. The surrounding roads are simple and there is no highway entrance or exit between the two traffic nodes. The average daily charging demand of this charging station is 617kW, with a peak period of 16:00-18:00 and a low period of 0:00-6:00.

[0219] The No. 2 service area has a long distance and a daily traffic volume of 18,088 vehicles. The daily charging demand of this charging station fluctuates greatly. 10:00-18:00 is the peak period, with the peak charging demand reaching up to 2,100kW; 1:00-6:00 is the trough period, with the charging demand being less than 900kW.

[0220] The distance of this section of service area No. 3 is relatively short, with a daily traffic of 20,318 vehicles, and there are multiple confluence and diversion nodes around it. The average daily charging demand of this charging station is 1637kW, with a peak period of 8:00-24:00, with a peak demand of up to 2000kW, and a low period of 1:00-5:00;

[0221] The distance of this section of service area No. 4 is relatively short, so the traffic volume on this section is relatively small. The average daily demand of this charging station is only 198kW, and the off-peak period is between 16:00 and 18:00.

[0222] Finally, the prediction results provided by the online simulation module provide important decision-making support for traffic management and energy planning departments, which can be used to optimize the supply strategy of charging facilities, improve the utilization efficiency and service level of charging infrastructure, and promote the development of intelligent transportation systems towards smart cities.

[0223] like Figure 2 As shown in part (d), the data analysis module performs the following tasks:

[0224] Perform weighted analysis on the prediction results of the online simulation module and the prediction data of the deep neural network training module, and introduce a feedback mechanism to evaluate the prediction result deviation and obtain a comprehensive prediction value;

[0225] Specific methods include: prediction result weighting method and feedback mechanism;

[0226] The expression of the prediction result weighting method is as follows:

[0227] Weighted analysis = ω 仿真 ×Prediction value 仿真 +ω 深度学习 ×Prediction value 深度学习

[0228] In the formula, ω 仿真 is the weight of the online simulation module; ω 深度学习 is the weight of the deep neural network training module; the predicted value 仿真It is the predicted value obtained by the online simulation module; predicted value 深度学习 is the predicted value obtained by the deep neural network training module;

[0229] In the present invention, a dynamic weight adjustment mechanism based on prediction error is used to evaluate the deviation of historical prediction results through actual detection values, and the confidence (weight) of the feedback mechanism is estimated according to the deviation value; the specific steps are as follows:

[0230] 1. Save the prediction results of the online simulation module and the deep neural network training module as historical values;

[0231] 2. After a certain time interval (one week), collect the actual measurement values ​​of the charging load in the actual service area;

[0232] 3. Compare the EV starting toll station record data, ETC gantry vehicle passing information, and service area charging station data collected by the deep neural network training module with the actual measured values ​​of EV starting toll station record data, ETC gantry vehicle passing information, and service area charging station data collected by the online simulation deduction module to calculate the prediction deviation;

[0233] The purpose of calculating the prediction deviation is to evaluate the difference between the predicted values ​​of the online simulation module and the deep neural network training module and the actual measured values, and then adjust the weights of the two to optimize the future prediction accuracy;

[0234] The prediction deviation is calculated by the mean square error (MSE). First, the charging load measurement value of the actual service area is collected, which is recorded as A(t); the comprehensive prediction value of the online simulation deduction module is recorded as S(t); and the comprehensive prediction value of the deep neural network training module is recorded as D(t).

[0235] Calculation error: The error of the online simulation module and the deep neural network training module is calculated by the mean square error (MSE), and the formula is as follows:

[0236]

[0237] Where P(t) is the predicted value, i.e., the online simulation module S(t) or the deep neural network training module D(t), and A(t) is the actual value, t is the time, and n is the total number of time points. MSE calculates the average of the squared difference between the predicted value and the actual value, which can emphasize larger errors.

[0238] The errors of the online simulation deduction module and the deep neural network training module are shown in Table 1 below:

[0239] Table 1 Errors of online simulation module and deep neural network training module

[0240]

[0241] As can be seen from the above table, the MSE of the online simulation module is similar to that of the deep neural network training module. The online simulation module achieves the lowest error, with an MSE of 0.085. Through experiments on the data set and comparison of prediction results, the model has the smallest error, which proves the feasibility and superiority of this method.

[0242] After obtaining the error, the prediction values ​​of the online simulation deduction module and the deep neural network training module are combined by weighted summation to obtain the comprehensive prediction value; the expression is as follows:

[0243] P total (t) = ω S ·S(t)+ω D ·D(t)

[0244] In the formula, the weight of the online simulation module is ω S , the weight of the deep neural network training module is ω D , then the comprehensive prediction value P total (t);

[0245] The purpose of updating weights is to ensure that the prediction results in different time periods and under different conditions can reflect the latest prediction accuracy. After each cycle (one week), the MSE is recalculated based on the latest actual measurement values ​​and historical prediction results, and the weights are adjusted.

[0246] Through the data analysis module, the present invention demonstrates how to effectively combine the online simulation deduction module and the deep neural network training module prediction method and dynamically optimize the prediction results through a dynamic weight adjustment mechanism based on prediction error, so as to further improve the prediction accuracy and stability of the system.

[0247] According to the evaluation results, the weights of the online simulation deduction module and the deep neural network training module are adjusted to enable the present invention to achieve accurate prediction.

Claims

1. An electric vehicle charging load analysis system based on deep learning and real-time traffic simulation, characterized in that: include: Offline simulation module, deep neural network training module, online simulation module, data analysis module; The offline simulation module is used to collect the electric vehicle starting toll station record data, ETC gantry vehicle passing information, and service area charging station data on the highway, and build a highway network topology model and a revised energy consumption model based on the collected data; The deep neural network training module is used to generate a training set and a test set from the energy consumption data generated by the modified energy consumption model, preprocess the data of the training set and the test set through deviation standardization, input the preprocessed data set into the improved attention mechanism model, and output energy consumption prediction data; The online simulation module is used to calculate the ETC gantry vehicle passing information and the data of the charging pile failure part according to the existing starting and ending point traffic through simulation and prediction, and output the energy consumption prediction result; The data analysis module is used to perform weighted analysis on the energy consumption prediction results of the online simulation deduction module and the energy consumption prediction data of the deep neural network training module, and introduce a feedback mechanism to evaluate the prediction result deviation and obtain a comprehensive prediction value.

2. The electric vehicle charging load analysis system based on deep learning and real-time traffic simulation according to claim 1 is characterized in that: The offline simulation module includes: data verification, exception handling and real-time monitoring functions, which are executed in the collection stage; among them, data verification adopts a multiple verification method, which includes: verifying whether the timestamp conforms to the standard time format; verifying whether the vehicle speed data exceeds a reasonable upper limit, and verifying whether the toll amount is a negative value; comparing the vehicle passing information of the ETC gantry and the toll station records to ensure that the time and vehicle ID information of each vehicle passing through the toll station and the ETC gantry are consistent. If there is any inconsistency, it is marked as potential erroneous data.

3. The electric vehicle charging load analysis system based on deep learning and real-time traffic simulation according to claim 1 is characterized in that: The modified energy consumption model expression is as follows: P 消耗 =F 阻力总 ×v+P 加速 +P 制动能量回收 Where P 加速 is the energy required for vehicle acceleration, P 制动能量回收 is the energy of braking energy recovery, P 消耗 is the energy consumption per unit time, F 阻力总 is the total resistance experienced by the vehicle, and v is the speed of the vehicle.

4. The electric vehicle charging load analysis system based on deep learning and real-time traffic simulation according to claim 1 is characterized in that: The expression of the deviation standardization is as follows: Where X is any value in the energy consumption data, X norm is the normalized data value, X min is the minimum value in the energy consumption data, X max It is the maximum value in the energy consumption data.

5. The electric vehicle charging load analysis system based on deep learning and real-time traffic simulation according to claim 1 is characterized in that: The improved attention mechanism model is specifically improved as follows: a deep neural network model is designed through the attention mechanism to perform charging load prediction and electric vehicle type identification, and a multi-head attention mechanism and time series coding are used to improve the performance of the model; wherein, the time series coding includes: an input feature processing layer, a separation feature coding layer, a long- and short-term attention layer, a dynamic convolution layer, and outputs the final prediction result through a fully connected layer.

6. The electric vehicle charging load analysis system based on deep learning and real-time traffic simulation according to claim 5 is characterized in that: The separation feature encoding layer separates the processed input features and uses the embedding layer to capture the pattern of each feature; the construction of the embedding layer includes: defining basic parameters of the embedding layer, initializing the embedding matrix, training the embedding layer and applying the embedding layer.

7. The electric vehicle charging load analysis system based on deep learning and real-time traffic simulation according to claim 1 is characterized in that: The online simulation module uses distributed computing architecture, real-time data processing, elastic scaling and fault-tolerant mechanisms to solve the problems of ever-increasing data volumes and frequently updated environmental conditions.

8. The electric vehicle charging load analysis system based on deep learning and real-time traffic simulation according to claim 1 is characterized in that: The expression of the weighted analysis is as follows: Weighted analysis = ω 仿真 ×Prediction value 仿真 +ω 深度学习 ×Prediction value 深度学习 In the formula, ω 仿真 is the weight of the online simulation module; ω 深度学习 is the weight of the deep neural network training module; Predicted value 仿真 It is the predicted value obtained by the online simulation module; predicted value 深度学习 is the predicted value obtained by the deep neural network training module; The specific steps of the feedback mechanism are as follows: The prediction results of the online simulation module and the deep neural network training module are saved as historical values; After one week, actual measurements of charging load in the actual service area were collected; The prediction deviation is calculated by comparing the EV starting toll station record data, ETC gantry vehicle passing information, and service area charging station data collected by the deep neural network training module and the actual measured values ​​of EV starting toll station record data, ETC gantry vehicle passing information, and service area charging station data collected by the online simulation deduction module; the prediction deviation is obtained by mean square error calculation.

9. The electric vehicle charging load analysis system based on deep learning and real-time traffic simulation according to claim 1 is characterized in that: The prediction result deviation is calculated using mean square error; The expression of the comprehensive prediction value is as follows: P total (t)=ω S ·S(t)+ω D ·D(t) In the formula, the comprehensive weight of the online simulation module is ω S , the comprehensive weight of the deep neural network training module is ω D , then the comprehensive prediction value P total (t), the comprehensive prediction value of the online simulation module is recorded as S(t); The comprehensive prediction value of the deep neural network training module is denoted as D(t).

Images