Traffic flow and charging station load prediction method based on deep learning

Abstract

The invention provides a traffic flow and charging station load prediction method based on deep learning, and aims to improve the precision of traffic flow prediction and accurately predict the load condition of an electric vehicle charging station. The method comprises the following steps of: firstly, predicting future traffic flow and traffic speed by adopting a space-time graph convolutional network model based on an attention mechanism; and secondly, based on the probability distribution of the daily travel distance of the vehicle, calculating the charging demand of the electric vehicle by using a hybrid model and an expectation maximization algorithm, and predicting the energy consumption of the electric vehicle through a polynomial regression model. And finally, based on the driving behaviors of the electric vehicles, an electric vehicle queuing model is constructed, the model comprehensively considers the behaviors of capacity limitation, overlong waiting time, intolerant leaving and the like of the charging station, and the load prediction of the charging station is calculated through a random Markov chain analysis method. The method can effectively capture the time-space characteristics of the traffic flow and the charging demand of the electric vehicle, and improves the accuracy of load prediction of the charging station.

Description

Technical Field

[0001] The present invention relates to the technical field of traffic flow prediction and electric vehicle charging load prediction, and in particular to a traffic flow and electric vehicle charging load prediction method based on deep learning. Background Art

[0002] As global climate change and environmental pollution become increasingly serious, electric vehicles (EVs) have gradually become the focus of attention due to their environmental protection and energy-saving advantages. Compared with traditional fuel vehicles, electric vehicles reduce dependence on fossil fuels and reduce greenhouse gas emissions. However, the large-scale popularization of electric vehicles has also brought some new challenges, one of which is how to effectively predict traffic flow and charging load to optimize traffic management and the layout and operation of charging stations.

[0003] Traditional traffic flow forecasting methods usually rely on historical data and statistical models, such as time series analysis, regression analysis, etc. Although these methods perform well in some scenarios, they often have difficulty capturing the complex spatiotemporal dependencies in traffic flow, especially in the face of emergencies and changes in traffic patterns, and the forecasting accuracy is low.

[0004] Similarly, for the prediction of EV charging load, traditional methods mainly rely on simple demand forecasting models, which usually assume that charging demand is evenly distributed and ignore the complexity and randomness of EV charging behavior. For example, EV charging behavior is not only affected by the vehicle battery status and remaining power, but also by multiple factors such as traffic flow, driver behavior, and charging station service capacity.

[0005] In recent years, with the development of artificial intelligence technology, especially the advantages of deep learning in dealing with complex nonlinear problems, more and more studies have tried to introduce deep learning into the prediction of traffic flow and charging load. For example, convolutional neural networks (CNN) and recurrent neural networks (RNN) perform well in processing image and sequence data, but these models alone are still difficult to fully capture the spatiotemporal correlation in traffic data.

[0006] Deep learning models based on attention mechanisms, such as the combination of graph convolutional networks (GCNs) and attention mechanisms, have shown great potential in capturing global and local dependencies in data. These models can dynamically adjust the correlation weights between nodes and can more accurately capture the spatiotemporal dynamic characteristics of traffic flow and charging behavior. Summary of the invention

[0007] The present invention aims to solve the problems existing in the prior art and proposes a traffic flow and charging station load prediction method based on deep learning. By combining the spatial-temporal graph convolutional network model and the hybrid model with the attention mechanism, the traffic flow and electric vehicle charging load are accurately predicted, providing support for intelligent traffic management and optimization of charging infrastructure.

[0008] To achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0009] A spatial-temporal graph convolutional network model based on an attention mechanism is used to predict future traffic flow and traffic speed. The spatial-temporal graph convolutional network model based on the attention mechanism defines a traffic network as an undirected graph, and uses the feature value of each node in the traffic network in a historical time slice to predict traffic flow and traffic speed in multiple future time slices.

[0010] According to the actual data, a hybrid model is used to fit the distribution of daily driving distance of electric vehicles and the distribution of daily driving of electric vehicles, and the parameters of the hybrid model corresponding to the distribution of daily driving distance of electric vehicles and the distribution of daily driving of electric vehicles are optimized respectively by the expectation maximization algorithm;

[0011] According to the actual energy consumption data, a polynomial regression model is used to fit the mapping relationship between electric vehicle speed and energy consumption;

[0012] Using the above-mentioned predicted future traffic flow and traffic speed, the distribution of daily driving distance of electric vehicles, the distribution of daily driving of electric vehicles, and the mapping relationship between electric vehicle speed and energy consumption, combined with electric vehicle battery capacity and electric vehicle penetration rate data, the arrival rate of electric vehicles at charging stations is estimated;

[0013] An electric vehicle charging queue model is established that takes into account service capacity limitations and driver charging behavior, and the estimated arrival rate of electric vehicles at charging stations is corrected to obtain the corrected average arrival rate and average departure rate.

[0014] Combining the average arrival rate and the average departure rate, a random Markov chain is used to calculate the probability distribution of the number of electric vehicles in the charging station;

[0015] The overall load of the charging station is calculated by combining the probability distribution of the number of electric vehicles in the charging station and the power of the charging piles.

[0016] Preferably, the attention-based spatial-temporal graph convolutional network model includes three independent components, which are used to model the temporal correlation of the recent period, daily cycle and weekly cycle respectively. Each component includes several spatiotemporal blocks and a fully connected layer, and each spatiotemporal block consists of a spatiotemporal attention module and a spatiotemporal convolution module.

[0017] Preferably, the spatiotemporal attention module is used to calculate the spatial correlation between different nodes in the traffic network, can calculate the relative importance of each node to other nodes, extract the node features of the input data through the fully connected layer, and calculate the spatial attention weights between nodes based on the node feature vectors.

[0018] Preferably, the spatiotemporal convolution module is used to perform graph convolution operations on the node features output by the spatiotemporal attention module, capturing the spatial relationships in the transportation network through the relative importance weights between nodes.

[0019] Preferably, the estimating the arrival rate of electric vehicles at charging stations includes, based on the distribution of electric vehicles in daily driving, estimating the distance traveled in a day using the driving distance of the electric vehicle at a certain moment; based on the relationship between the charging state of the electric vehicle, the driving energy consumption of the electric vehicle, the battery capacity of the electric vehicle and the driving distance, estimating the daily driving distance of the electric vehicle to reach the charging station at this moment; then estimating the probability of the electric vehicle choosing to charge when passing the charging station at this moment based on the distribution of the daily driving distance of the electric vehicle; and finally calculating the arrival rate of the electric vehicle at the charging station in combination with the predicted results of traffic flow and the penetration rate of the electric vehicle.

[0020] Preferably, the electric vehicle charging queue model comprehensively considers three behaviors of electric vehicles leaving due to capacity limitation of charging stations, long waiting queues and impatience, corrects the arrival rate of electric vehicles at charging stations, and obtains the average arrival rate and average leaving rate of electric vehicles at charging stations.

[0021] Preferably, the use of random Markov chains to calculate the probability distribution of the number of electric vehicles in the charging station includes analyzing state transitions in the charging station based on random process theory, the states including arrival and departure, establishing a balance equation, and calculating the probability distribution of the number of electric vehicles in the charging station in combination with the average arrival rate and average departure rate of electric vehicles at the charging station.

[0022] The present invention proposes a traffic flow and charging station load prediction method based on deep learning, which has the following features:

[0023] Beneficial effects:

[0024] A spatial-temporal graph convolutional network model based on the attention mechanism is adopted to more accurately capture the spatiotemporal dependency of traffic flow and traffic speed and improve the prediction accuracy.

[0025] By utilizing the spatial and temporal attention mechanism, the correlation weights of nodes and time slices in the traffic network are dynamically adjusted to enhance the robustness and adaptability of the spatial-temporal graph convolutional network model based on the attention mechanism.

[0026] The spatial-temporal graph convolutional network model based on the attention mechanism contains three components: recent, daily periodicity, and weekly periodicity, which comprehensively captures the changes in traffic flow and traffic speed at different time scales.

[0027] An electric vehicle charging queuing model is established that takes into account service capacity limitations and driver charging behavior to more accurately simulate the arrival and departure behavior of electric vehicles at charging stations and improve the accuracy of charging load prediction.

[0028] The mapping relationship between electric vehicle speed and energy consumption is established through a polynomial regression model. Combined with the traffic flow and traffic speed prediction results, the energy consumption of electric vehicles is accurately estimated, providing reliable data for charging load prediction.

[0029] Combining deep learning with probabilistic models improves the accuracy and stability of traffic flow and charging load predictions, achieving an effective integration of the two, and has high practical value and promotion prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 Flowchart of the traffic flow and charging station load prediction method based on deep learning and probabilistic modeling.

[0031] Figure 2 It is a spatial-temporal graph convolutional network model framework based on the attention mechanism. DETAILED DESCRIPTION

[0032] The present invention will be further described in detail below in conjunction with the accompanying drawings.

[0033] like Figure 1 As shown, the present invention provides a method for predicting traffic flow and charging station load based on deep learning, comprising the following steps:

[0034] Step S1: Use the spatial-temporal graph convolutional network model based on the attention mechanism to predict future traffic flow and traffic speed. The traffic network is defined as an undirected graph G = (V, E, A), where V is a set of nodes, |V| = N nodes; E is a set of edges, indicating the connectivity between nodes; Represents the adjacency matrix of the undirected graph G. Each node on the traffic network undirected graph G detects F measurements with the same sampling frequency, that is, each node generates a feature vector of length F in each time slice, including traffic flow, traffic occupancy, traffic speed, etc. Assume that the fth and sth time series recorded by each node on the traffic network undirected graph G are the traffic flow sequence and traffic speed sequence, respectively, f,s∈(1,...,F). Use represents the cth eigenvalue of node i at time t, Represents all feature values ​​of node i at time t. Represents all eigenvalues ​​of all nodes at time t. represents all the eigenvalues ​​of all nodes in the past τ time slices. In addition, let represents the traffic flow of node i at future time t, let represents the traffic speed of node i at future time t. The following will introduce how to use the historical measurements in the past τ time slices to predict the traffic speed of all nodes in the traffic network after T p Future traffic flow in time slices and traffic speed in represents the future traffic flow of node i starting from τ+1, represents the future traffic speed of node i starting from τ+1.

[0035] The spatial-temporal graph convolutional network model framework based on the attention mechanism is as follows Figure 2 As shown, it consists of three independent components with the same structure, which are used to model the recent, daily periodicity and weekly periodicity dependencies of historical data respectively.

[0036] Assume that the sampling frequency is q times per day and the current time is t 0 , the size of the prediction window is T p . The length of the intercept along the time axis is T h , T d and T w The three time series segments of are used as inputs for the recent, daily and weekly components, respectively, where T h , T d and T w All T p An integer multiple of .

[0037] Recent inputs are is the historical time series segment directly adjacent to the forecast period. The input for the daily period is It consists of segments from the past few days with the same time period as the forecast period. The input for the weekly period is It consists of segments with the same weekly attributes and time intervals in the past few weeks. These three components share the same network structure, each of which contains several spatiotemporal blocks and a fully connected layer. Each spatiotemporal block has a spatiotemporal attention module and a spatiotemporal convolution module. The following is a detailed explanation of the spatial attention module, temporal attention module, and graph convolution module.

[0038] In order to effectively model the spatial correlation between different nodes, a spatial attention module is designed, which can calculate the relative importance of each node in the traffic network to all other nodes. Specifically, the calculation process of the spatial attention module is as follows:

[0039] First, the input sequence X passes through the first fully connected layer Projected into a d-dimensional space to extract higher-level features. Then, the spatial attention weights between nodes are calculated

[0040]

[0041] Among them, the intermediate variables are calculated by the input feature sequence and the fully connected layer, H S =XW S .H S,j Indicates H S Finally, A S Perform weighted summation with the input sequence X to obtain the weighted node features:

[0042] Z S =A S X

[0043] In order to capture the temporal dependency in traffic data, a temporal attention module is introduced. The calculation process of the temporal attention module is as follows:

[0044] First, the input sequence X is passed through the second fully connected layer Projected into a d-dimensional space to extract higher-level features. Then, the temporal attention weights between different time slices are calculated.

[0045]

[0046] Among them, H T =XW T , H T,j Indicates H T Finally, A T Perform weighted summation with the input sequence X to obtain the weighted time slice features:

[0047] Z T =A T X

[0048] In order to further extract the spatial relationship between nodes, a graph convolution module is introduced after the spatial attention module. Specifically, the calculation process of the graph convolution module is as follows:

[0049] First, the weighted node feature Z S Through a graph convolution layer Projected into a d-dimensional space. Then, perform graph convolution on the node features:

[0050] Z G =A S ZS W G

[0051] Among them, A S is the relative importance weight between nodes calculated in the spatial attention module. Through graph convolution operations, the spatial relationship between nodes in the traffic network can be better captured.

[0052] After processing multiple spatiotemporal blocks, the final feature representation is passed through the third fully connected layer Project it into a scalar space to get the traffic flow prediction value of each node in the future time slice:

[0053] Y=Z G W F

[0054] The mean square error loss function is used to train the spatial-temporal graph convolutional network model based on the attention mechanism. After the training is completed, the traffic flow and traffic speed are predicted by inputting the corresponding recent, daily cycle and weekly cycle inputs.

[0055] Step S2: Calculate the probability distribution of the number of electric vehicles in the charging station.

[0056] In order to obtain the probability distribution of the actual daily driving distance of electric vehicles, the distribution of the daily driving distance of electric vehicles is fitted according to the actual data using a method based on a mixed model. The mixed model uses a combination of selected probability density functions to model random variables:

[0057]

[0058] Where U is the set of random variables, Θ is the probability density parameter, and π m is the weight, M is the number of probability density function types selected, and p m (u i ∣θ m ) is the chosen probability density function.

[0059] Using the expectation maximization algorithm, the parameters of the mixed model are obtained through two steps of iteration to maximize the expectation of the likelihood:

[0060]

[0061] Where N is the number of data samples, V is the set of auxiliary variables, and v m,i is an indicator variable (often called a latent variable) that indicates the i-th data point u i Probability from the mth distribution.

[0062] Before applying the expectation maximization algorithm, the parameters and auxiliary variable set V are initialized using the K-means clustering method. The number of probability density functions M is another key indicator that can be obtained using the Akaike Information Criterion (AIC) and the Bayesian Information Criterion (BIC), which are expressed as follows:

[0063] AIC=2M-2lnargmax Θ L(Θ|U)

[0064] BIC=Mln(N)-2lnargmax Θ L(Θ|U)

[0065] The expectation-maximization algorithm is performed in an iterative manner, selecting the best M by increasing M one by one until BIC(M)-BIC(M+1) and AIC(M)-AIC(M+1) are less than the preset termination criteria.

[0066] In the qth iteration, the first step is to calculate the current estimate Θ (q) The expected value of the lower likelihood function is:

[0067] Q(Θ|Θ (q) )=E[lnL(Θ|U,V)|(V,Θ (q) )]

[0068] The second step is to determine the parameters to maximize the expected value in this iteration:

[0069] Θ (q+1) = argmax Θ Q(Θ|Θ (q) )

[0070] The two-step algorithm is executed in an iterative manner until the preset stopping criterion is reached, and the parameters are optimally estimated using the proposed hybrid model to obtain the probability distribution of the daily travel distance of the EV.

[0071] The same method can also be used to obtain the daily driving distribution of electric vehicles, that is, the probability of whether the electric vehicle is driving at the current moment.

[0072] Through the above steps, the probability distribution of the daily travel distance of electric vehicles and the distribution of vehicles in daily driving are obtained. Next, the arrival rate of electric vehicles at charging stations is estimated.

[0073] The willingness of EV drivers to use charging stations along their driving route determines the arrival rate and is highly correlated with the state of charge (SOC) of the EV, while the charging time μ is determined by the SOC and the charging power. ev It is a key data source for estimating the SOC of electric vehicles, and has the following relationship:

[0074] SOC=1-p ev D ev / E ev

[0075] Among them, E ev represents the battery capacity of electric vehicles, p ev Indicates the driving energy consumption of electric vehicles.

[0076] Driving energy consumption of electric vehicles ev Related to the speed v of the electric vehicle, a polynomial regression model is used to model the mapping relationship between the speed of the electric vehicle and the energy consumption. The energy consumption data of the electric vehicle at different speeds is collected through the on-board data recording system of the electric vehicle, special test experiments or data provided by the manufacturer. The collected energy consumption data is cleaned to remove outliers and noise; the data is smoothed to reduce data fluctuations; the data is normalized to process the data on a uniform scale. Select the polynomial regression model and assume that the energy consumption p ev With speed v speed The relationship between can be expressed as:

[0077] p ev =a 0 +a 1 v speed +a 2 v speed 2 +...+a n v speed n

[0078] Where (a 0 ,a 1 ,…,a n ) are constants of the polynomial.

[0079] The polynomial regression model is trained using the training set data, and the model parameters are estimated using the least squares method. The model parameters are further adjusted through gradient descent to minimize the model fitting error. Finally, p ev With speed v speed The relationship between the traffic speed and the energy consumption of the electric vehicle can be predicted by the traffic speed predicted in step S1.

[0080] Use the travel distance D at time t ev,t Estimated daily driving distance

[0081]

[0082] In order to obtain the energy consumption p ev Distance from day After that, the estimated daily distance of the electric vehicle to reach the charging station at time t can be calculated. Assume that EV drivers prefer to operate at SOC lower than the standard SOC min , and their average SOC is SOC m (0≤SOC min ≤SOC min ). The estimated daily driving distance of an electric vehicle arriving at the charging station at time t is for:

[0083]

[0084] The probability P of an electric vehicle choosing to charge when it passes a charging station at time t t The daily driving distance distribution p(X) can be used to estimate:

[0085]

[0086] Therefore, at time t, the arrival rate of electric vehicles at the charging station is:

[0087]

[0088] in is the penetration rate of electric vehicles in the total traffic flow; f t is the predicted result of traffic flow at time t.

[0089] The arrival rate obtained at this time is not completely accurate. The present invention proposes an electric vehicle charging queuing model that considers service capacity limitations and driver charging behavior. Sometimes electric vehicles leave the charging station without charging. This behavior of electric vehicle drivers may be caused by insufficient total number of available chargers, capacity limitations of charging stations, etc. Assuming that the total number of available chargers at the charging station is C and the maximum capacity of the charging station is K, the root causes of electric vehicle leaving behavior are divided into three types: forced leaving; refusal to join; impatient leave.

[0090] 1. Forced departure: If the charging station already accommodates K EVs, the other EVs will be forced to leave the charging station because there is no space in the queue.

[0091] 2. Refusing to join: When arriving at a charging station, an EV finds the waiting queue too long and therefore refuses to join the queue, even if the total number of EVs in the charging station is less than K. According to categories 1 and 2, the probability that an EV chooses to queue is related to the number of EVs w in the charging station and is assumed to be

[0092]

[0093] where the exponential function indicates that the probability of selecting an electric vehicle in the queue decreases faster when the number of electric vehicles w in the charging station increases, and σ is a parameter that defines the probability decrease rate.

[0094] 3. Impatient departure: Even if the waiting queue is not too long, the EVs in the queue may still choose to leave the charging station when they are too impatient. The number of EVs remaining in the queue at each time interval is related to the number of EVs in the charging station w and is assumed to be

[0095]

[0096] Considering categories 1-3, the average arrival rate and average departure rate can be expressed as λ w =α w λ and μ w =μ-b λ .

[0097] Based on the described electric vehicle charging queue model, the charging station load can be obtained using random Markov chain analysis. According to the theory of random processes, the steady-state probability of the Markov chain in state e should satisfy the equilibrium equation:

[0098]

[0099] The formula indicates that the sum of all state probabilities should be 1. In addition, the probability of entering and exiting each state should be equal, q eg is the transition rate from state e to g, and S is the state space.

[0100] The state transitions in a charging station are of only two types: arrival and departure. Therefore, the charging station balance equation can be expressed as

[0101] (λ w +μ w ) w =λ w-1 P w-1 +μ w+1 P w+1

[0102] Combining the above electric vehicle departure behavior and charging station balance equation, the probability that a charging station has w electric vehicles is

[0103]

[0104] in,

[0105]

[0106] Among them, ρ=λ / μ, η=δ / μ.

[0107] Therefore, the number of charging electric vehicles in the charging station can be expressed as

[0108]

[0109] Step S3: Calculate the overall load of the charging station based on the probability distribution of the number of electric vehicles charged in the charging station and the charging power of each vehicle.

[0110] The overall load of the charging station P FCS Through N ch and charging pile power p EV To determine:

[0111] P FCS =N ch p EV .

Claims

1. A traffic flow and charging station load prediction method based on deep learning, characterized in that: The steps include: A spatial-temporal graph convolutional network model based on an attention mechanism is used to predict future traffic flow and traffic speed. The spatial-temporal graph convolutional network model based on the attention mechanism defines a traffic network as an undirected graph, and uses the feature value of each node in the traffic network in a historical time slice to predict traffic flow and traffic speed in multiple future time slices. According to the actual data, a hybrid model is used to fit the distribution of daily driving distance of electric vehicles and the distribution of daily driving of electric vehicles, and the parameters of the hybrid model corresponding to the distribution of daily driving distance of electric vehicles and the distribution of daily driving of electric vehicles are optimized respectively by the expectation maximization algorithm; According to the actual energy consumption data, a polynomial regression model is used to fit the mapping relationship between electric vehicle speed and energy consumption; Using the above-mentioned predicted future traffic flow and traffic speed, the distribution of daily driving distance of electric vehicles, the distribution of daily driving of electric vehicles, and the mapping relationship between electric vehicle speed and energy consumption, combined with electric vehicle battery capacity and electric vehicle penetration rate data, the arrival rate of electric vehicles at charging stations is estimated; An electric vehicle charging queue model is established that takes into account service capacity limitations and driver charging behavior, and the estimated arrival rate of electric vehicles at charging stations is corrected to obtain the corrected average arrival rate and average departure rate. Combining the average arrival rate and the average departure rate, a random Markov chain is used to calculate the probability distribution of the number of electric vehicles in the charging station; The overall load of the charging station is calculated by combining the probability distribution of the number of electric vehicles in the charging station and the power of the charging piles.

2. The method for predicting traffic flow and charging station load based on deep learning according to claim 1, characterized in that: The attention-based spatial-temporal graph convolutional network model includes three independent components, which are used to model the temporal correlation of recent, daily and weekly cycles respectively. Each component includes several spatiotemporal blocks and a fully connected layer, and each spatiotemporal block consists of a spatiotemporal attention module and a spatiotemporal convolution module.

3. The method for predicting traffic flow and charging station load based on deep learning according to claim 2, characterized in that: The spatiotemporal attention module is used to calculate the spatial correlation between different nodes in the traffic network. It can calculate the relative importance of each node to other nodes, extract the node features of the input data through the fully connected layer, and calculate the spatial attention weights between nodes based on the node feature vectors.

4. The method for predicting traffic flow and charging station load based on deep learning according to claim 2, characterized in that: The spatiotemporal convolution module is used to perform graph convolution operations on the node features output by the spatiotemporal attention module, and capture the spatial relationships in the transportation network through the relative importance weights between nodes.

5. The method for predicting traffic flow and charging station load based on deep learning according to claim 1, characterized in that: The method for estimating the arrival rate of electric vehicles at charging stations includes estimating the distance traveled in a day by using the travel distance of the electric vehicle at a certain moment based on the distribution of the electric vehicle in daily travel; estimating the daily travel distance of the electric vehicle to reach the charging station at this moment based on the relationship between the charging state of the electric vehicle, the driving energy consumption of the electric vehicle, the battery capacity of the electric vehicle and the travel distance; then estimating the probability of the electric vehicle choosing to charge when passing the charging station at this moment based on the distribution of the daily travel distance of the electric vehicle; and finally calculating the arrival rate of the electric vehicle at the charging station in combination with the predicted results of the traffic flow and the penetration rate of the electric vehicle.

6. The method for predicting traffic flow and charging station load based on deep learning according to claim 1, characterized in that: The electric vehicle charging queue model comprehensively considers three behaviors of electric vehicles leaving due to capacity limitation of charging stations, long waiting queues and impatience, corrects the arrival rate of electric vehicles at charging stations, and obtains the average arrival rate and average leaving rate of electric vehicles at charging stations.

7. The method for predicting traffic flow and charging station load based on deep learning according to claim 1, characterized in that: The method of using a random Markov chain to calculate the probability distribution of the number of electric vehicles in a charging station includes analyzing the state transition in the charging station according to random process theory, wherein the state includes arrival and departure, establishing a balance equation, and calculating the probability distribution of the number of electric vehicles in the charging station in combination with the average arrival rate and the average departure rate of the electric vehicles at the charging station.

Images