Electric vehicle charging load space-time prediction method based on hybrid model driving

Abstract

The invention provides an electric vehicle charging load space-time prediction method based on hybrid model driving, and relates to the field of electric vehicle charging, and the method comprises the steps: building a user travel model based on a Markov chain; constructing a multi-angle traffic road network model and a load space-time prediction model based on multi-graph feature fusion; acquiring user historical travel data, electric vehicle charging load historical data, traffic network historical information and external historical data of the to-be-predicted area; generating a user MDP trip chain according to the user historical trip data of the to-be-predicted area through a user trip model; extracting a multi-angle spatial relation graph from the traffic road network historical information of the to-be-predicted area through the multi-angle traffic road network model; the electric vehicle charging load of the to-be-predicted area is predicted through the load space-time prediction model based on multi-graph feature fusion according to the user MDP trip chain, the multi-angle spatial relation graph and the external historical data, and the method has the advantage of improving the accuracy of electric vehicle charging load prediction.

Description

Technical Field

[0001] The invention relates to the field of electric vehicle charging, and in particular to a hybrid model-driven method for spatiotemporal prediction of electric vehicle charging load. Background Art

[0002] The gradual development of new energy in the transportation field promotes carbon emission reduction. Against this background, the large-scale popularization and application of electric vehicles has become an inevitable trend. While alleviating energy and environmental pressures, the large-scale access of electric vehicle loads has also affected the safety of power grid operation. Therefore, in order to reduce its negative impact, further research is needed to effectively predict the spatiotemporal distribution of electric vehicle charging loads.

[0003] The data-driven artificial intelligence method obtains a large amount of historical load data for iterative training, and the implementation process is relatively simple. However, due to the popularity of measurement equipment, the stability of measurement results and data confidentiality, it is difficult to obtain sufficiently reliable data. At the same time, the prediction model generated by this method is similar to a "black box", which makes it difficult to achieve quantitative analysis of relevant influencing factors and lacks interpretability. The model-driven method represents complex charging behavior by assuming a large number of model parameters. This method has strong interpretability, wide applicability, and the output results are more accurate and reliable, but its probability distribution is difficult to obtain and the computational complexity is high.

[0004] Therefore, it is necessary to provide a spatiotemporal forecasting method for electric vehicle charging load driven by a hybrid model to improve the accuracy of electric vehicle charging load forecasting. Summary of the invention

[0005] The present invention provides a hybrid model-driven spatiotemporal prediction method for electric vehicle charging load, comprising: constructing a user travel model based on a Markov chain; constructing a multi-angle traffic network model; constructing a load spatiotemporal prediction model based on multi-graph feature fusion; acquiring user historical travel data, electric vehicle charging load historical data, traffic network historical information and external historical data of a to-be-predicted area; generating a user MDP travel chain of the to-be-predicted area according to the user historical travel data of the to-be-predicted area through the user travel model; extracting a multi-angle spatial relationship graph from the traffic network historical information of the to-be-predicted area through the multi-angle traffic network model; and predicting the electric vehicle charging load of the to-be-predicted area according to the user MDP travel chain, the multi-angle spatial relationship graph and the external historical data of the to-be-predicted area through the load spatiotemporal prediction model fused with multi-graph features.

[0006] Furthermore, the construction of the user travel model based on the Markov chain includes: determining multiple spatial states, wherein the multiple spatial states at least include a stay state in a residential area, a stay state in a work area, a stay state in a commercial area, and a stay state in other areas; constructing a Markov chain state transfer matrix based on the multiple spatial states; defining multiple model parameters of the user travel model, wherein the multiple model parameters at least include the first trip time of each vehicle, the stay time at the destination, and the choice of the travel destination; constructing the user travel model based on the Markov chain state transfer matrix and the multiple model parameters.

[0007] Furthermore, the user travel model generates a user MDP travel chain for the area to be predicted based on the user historical travel data of the area to be predicted, including: S11, determining the probability distribution of the user's first travel time and the probability distribution of the stay time of various types of destinations based on the user's historical travel data of the area to be predicted, wherein the various types of destinations include residential areas, work areas, commercial areas and other areas; S12, selecting a user's starting destination; S13, extracting the first travel time based on the probability distribution function of the first travel time, and extracting the stay time at the current destination based on the probability distribution function of the stay time at the destination; S14, randomly selecting the next spatial state based on the state transfer matrix constructed according to the Markov chain, and recording the state transfer path, repeating the random selection and recording each state transfer path; S15, statistically calculating the frequency of each spatial state being selected based on each state transfer path, and judging the most suitable state transfer node based on the state transfer with the highest frequency; S16, repeating S13-S15 until transferring to the end point.

[0008] Furthermore, the probability distribution of residence time in residential areas is: The probability distribution of the stay time in the work area, commercial area or other areas is: Among them, f1(T mij ,k,λ) is the probability distribution of the residence time in the residential area, f2(T mij ,μ,σ,γ) is the probability distribution of the stay time in the work area, commercial area or other areas, T mij is the length of time that an electric vehicle stays at the destination; λ is the scale parameter of the probability distribution of the stay time in residential areas; k is the shape parameter of the probability distribution of the stay time in residential areas; μ is the location parameter of the probability distribution of the stay time in work areas, commercial areas or other areas; σ is the scale parameter of the probability distribution of the stay time in work areas, commercial areas or other areas; γ is the shape parameter of the probability distribution of the stay time in work areas, commercial areas or other areas, σ>0.

[0009] Furthermore, the multi-angle spatial relationship diagram includes a regional node distance diagram, a traffic network diagram and a flow density diagram.

[0010] Furthermore, the load spatiotemporal prediction model with multi-graph feature fusion includes a feature extraction module, a multi-graph feature fusion module, a spatiotemporal prediction module and a load prediction module, wherein the feature extraction module is used to extract external feature variables from the external historical data, the multi-graph feature fusion module is used to extract graph features of the multi-angle spatial relationship graph, the spatiotemporal prediction module is used to extract a multidimensional vector for load prediction from the external feature variables, the graph features of the multi-angle spatial relationship graph and the user MDP travel chain of the area to be predicted, and the load prediction module is used to predict the electric vehicle charging load in the area to be predicted based on the multidimensional vector for load prediction.

[0011] Furthermore, the feature extraction module extracts external feature variables from the external historical data, including: performing data preprocessing on the external historical data to obtain preprocessed external historical data; extracting external feature variables from the preprocessed external historical data, wherein the external feature variables include at least weather series, temperature series, date type series and economic feature series.

[0012] Furthermore, the multi-graph feature fusion module includes multiple graph attention network layers, multi-graph fusion layers and residual connection layers, wherein the multiple graph attention network layers are used to extract the features of the fused neighborhood information corresponding to the regional node distance graph, the traffic network graph and the flow density graph respectively, the multi-graph fusion layer is used to fuse the features of the fused neighborhood information corresponding to the regional node distance graph, the traffic network graph and the flow density graph respectively to generate a node representation, and the residual connection layer is used to generate the graph features of the multi-angle spatial relationship graph based on the node representation.

[0013] Furthermore, the spatiotemporal prediction module includes multiple LSTM layers and self-attention mechanism layers, wherein the multiple LSTM layers are used to extract initial multidimensional vectors from the external feature variables, the graph features of the multi-angle spatial relationship graph, and the user MDP travel chain of the area to be predicted, and the self-attention mechanism layer weights the initial multidimensional vector to generate the multidimensional vector for load forecasting.

[0014] Furthermore, the load prediction module predicts the electric vehicle charging load in the area to be predicted based on the multidimensional vector for load prediction, including: inputting the multidimensional vector for load prediction according to different weights; and predicting the electric vehicle charging load in the area to be predicted based on the multidimensional vector for load prediction input according to different weights through a fully connected layer.

[0015] Compared with the prior art, the hybrid model-driven spatiotemporal prediction method for electric vehicle charging load provided by the present invention has at least the following beneficial effects:

[0016] By combining the mechanism model and data-driven artificial intelligence algorithm, the interpretability of the spatiotemporal prediction method of electric vehicle charging load driven by the hybrid model is improved. First, a user travel model based on Markov chain is constructed to simulate the relationship between user travel patterns and spatiotemporal changes in load. Then, the artificial intelligence algorithm is used to extract the spatiotemporal characteristics of user travel in the model and the characteristics of other influencing factors in the historical data set for load forecasting, aiming to improve the prediction accuracy of the electric vehicle charging load, so that the entire method has stronger interpretability. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] This specification will be further described in the form of exemplary embodiments, which will be described in detail by the accompanying drawings. These embodiments are not restrictive, and in these embodiments, the same number represents the same structure, wherein:

[0018] Figure 1 It is a flow chart of a method for spatiotemporal prediction of electric vehicle charging load based on hybrid model driving according to some embodiments of this specification;

[0019] Figure 2 is a schematic diagram of a single-destination trip according to some embodiments of this specification;

[0020] Figure 3 is a schematic diagram of multi-destination travel according to some embodiments of this specification;

[0021] Figure 4 is a schematic diagram of a user MDP travel chain according to some embodiments of this specification;

[0022] Figure 5 It is a schematic diagram of the structure of the LSTM layer and the self-attention mechanism layer shown in some embodiments of this specification. DETAILED DESCRIPTION

[0023] In order to more clearly illustrate the technical solutions of the embodiments of this specification, the following is a brief introduction to the drawings required for the description of the embodiments. Obviously, the drawings described below are only some examples or embodiments of this specification. For ordinary technicians in this field, this specification can also be applied to other similar scenarios based on these drawings without creative work. Unless it is obvious from the language environment or otherwise explained, the same reference numerals in the figures represent the same structure or operation.

[0024] Figure 1is a flow chart of a hybrid model-driven spatiotemporal prediction method for electric vehicle charging load according to some embodiments of this specification, such as Figure 1 As shown, a spatiotemporal prediction method for electric vehicle charging load driven by a hybrid model may include the following steps.

[0025] Step 110, constructing a user travel model based on the Markov chain;

[0026] Step 120, constructing a multi-angle traffic network model;

[0027] Step 130, constructing a load spatiotemporal prediction model based on multi-graph feature fusion;

[0028] Step 140, obtaining historical user travel data, historical data on electric vehicle charging loads, historical information on traffic network, and external historical data in the area to be predicted;

[0029] Step 150, generating a user MDP travel chain of the area to be predicted based on the historical travel data of users in the area to be predicted by using a user travel model;

[0030] Step 160, extracting a multi-angle spatial relationship diagram from the historical information of the traffic network of the area to be predicted by using a multi-angle traffic network model;

[0031] Step 170 , predicting the electric vehicle charging load in the area to be predicted based on the user MDP travel chain, multi-angle spatial relationship graph and external historical data in the area to be predicted by using the load spatiotemporal prediction model based on multi-graph feature fusion.

[0032] In some embodiments, step 110 specifically includes:

[0033] Determine a plurality of spatial states, wherein the plurality of spatial states at least include a stay state in a residential area, a stay state in a work area, a stay state in a commercial area, and a stay state in other areas;

[0034] According to various spatial states, a Markov chain is constructed to construct a state transfer matrix;

[0035] Defining multiple model parameters of the user travel model, wherein the multiple model parameters at least include the first travel time of each vehicle, the destination stay time and the choice of the travel destination;

[0036] The state transfer matrix and multiple model parameters are constructed based on the Markov chain to build the user travel model.

[0037] Specifically, a Markov chain (MDP, Markov Decision Process) is a mathematical model that describes the process of a system transitioning between different states, and its future state depends only on the current state and has nothing to do with the past state.

[0038] According to the user's daily travel intention, the destinations of electric vehicles can be divided into four types, namely residential areas H, work areas W, commercial areas B, and other areas O. Among them, the arrival of electric vehicles in a certain area is regarded as a state, and the spatial transfer between states is the driving process of the electric vehicle. According to the characteristics of user travel, the following are established: Figure 2 and Figure 3 There are two types of travel chains shown: single-destination travel and multi-destination travel.

[0039] According to the four determined spatial states, a state transfer matrix is ​​constructed based on the Markov chain to describe the transfer rules of different spatial states of users driving electric vehicles. ij is the transition probability of the electric vehicle from state i to state j. According to the four spatial states divided above, the state transfer matrix N = 4 is obtained, which is specifically expressed as:

[0040]

[0041] Among them, m 11 、m 22 、m 33 and m 44 are the self-transition probabilities of the destination residential area, work area, commercial area and other areas, respectively, and the rest are the transition probabilities between the corresponding destinations, for example, m 12 is the transfer probability of electric vehicles from residential areas to work areas, m 34 is the transfer probability of electric vehicles from commercial areas to other areas, m 42 is the transfer probability of electric vehicles from other areas to the working area.

[0042] Transition probability m ij The following two constraints are met:

[0043]

[0044] Step 150 specifically includes:

[0045] S11, determining the probability distribution of the user's first trip time and the probability distribution of the stay time at various types of destinations based on the historical travel data of users in the area to be predicted, wherein the various types of destinations include residential areas, work areas, commercial areas and other areas;

[0046] S12, selecting a user starting destination;

[0047] S13, extracting the first travel time according to the probability distribution function of the first travel time, and extracting the stay time at the current destination according to the probability distribution function of the stay time at the destination;

[0048] S14. Construct a state transition matrix according to the Markov chain, randomly select the next spatial state, record the state transition path, and repeat the random selection multiple times and record the state transition path each time;

[0049] S15. According to the state transition path each time, statistically calculate the frequency of each spatial state being selected;

[0050] S16. Repeat steps S13 - S15 until reaching the end point.

[0051] Among them, the probability distribution of the residence time in the residential area is:

[0052]

[0053] The probability distribution of the residence time in the work area, commercial area or other areas is:

[0054]

[0055] Among them, f1(T mij , k, λ) is the probability distribution of the residence time in the residential area, f2(T mij , μ, σ, γ) is the probability distribution of the residence time in the work area, commercial area or other areas, T mij is the residence duration of the electric vehicle at the destination; λ is the scale parameter of the probability distribution of the residence time in the residential area; k is the shape parameter of the probability distribution of the residence time in the residential area; μ is the location parameter of the probability distribution of the residence time in the work area, commercial area or other areas; σ is the scale parameter of the probability distribution of the residence time in the work area, commercial area or other areas; γ is the shape parameter of the probability distribution of the residence time in the work area, commercial area or other areas, and σ > 0. The scale parameters λ, σ control the change in amplitude of the probability distribution function; the shape parameters k, γ control the change in the shape of the probability distribution function, that is, the shape of the distribution tail. When k > 1, the probability distribution function is right - skewed, and when 0 < k < 1, the probability distribution function is left - skewed; the location function determines the center and location characteristics of the probability distribution function; these parameters are all calculated by maximum likelihood estimation.

[0056] The state transfer matrix constructed above is used to dynamically simulate the user's travel path in real time. The travel chain records the state of electric vehicles at different time points, that is, the spatiotemporal distribution information of electric vehicles, which reflects the randomness of the spatial position movement of electric vehicles. Finally, the user travel model output can obtain the probability of transferring from the current node to all adjacent nodes, and the entire travel path is formed into a transfer density function with a total probability density of 1. Since the user's daily travel time and travel destination choices have a certain degree of randomness, they are converted into a probability model, and random sampling is performed through the Monte Carlo simulation method. Then, a large number of random samples are extracted from the probability distribution to simulate the relevant data of electric vehicles. Finally, the most suitable state transfer node is approximately calculated and the user MDP travel chain is obtained. The specific simulation generation steps are as follows:

[0057] Define the user travel model, including parameters such as the first trip time of each vehicle, the length of stay at the destination, the choice of travel destination, and the probability distribution of these parameters.

[0058] Initialization state. Select an initialization state and choose a user starting destination.

[0059] Monte Carlo random sampling. Use a pseudo-random number generator to generate random numbers. Based on the random numbers, extract all the destinations of this user's trip according to the Markov transition matrix. According to the fitted distribution probability function, extract the first trip time and the length of stay at the destination of the electric vehicle. The extraction is achieved by generating random numbers and mapping them to the fitted distribution probability function. According to the transition probability, randomly select the next state from the current state and record the path.

[0060] Simulation test. Repeat the random selection multiple times and record the state transition path each time.

[0061] Statistical analysis. By recording each state transition path, statistically calculating the frequency of each transfer node being selected, and judging the most appropriate state transition node based on the state transition with the highest frequency, a more accurate transition probability can be obtained and the transition probability can be optimized to ensure the reliability of the final generated travel chain.

[0062] After the transfer is completed, the above process is repeated until the electric vehicle is transferred to the residential area, thus forming a Figure 4 The user MDP travel chain shown in Figure 1 includes the user travel state sequence, travel state transition probability, first travel time, and destination stay duration.

[0063] In the process of electric vehicle time-space transfer, traffic conditions will indirectly affect travel time and the choice of user travel routes, which will further affect the time-space distribution of electric vehicle charging load. Therefore, it is necessary to consider the status of the entire traffic network in the area and use graph theory methods to consider from multiple angles to establish a complex traffic network model. The multi-angle spatial relationship diagrams in the traffic network include regional node distance diagrams, traffic network diagrams, and flow density diagrams.

[0064] G=(V,E,Q,W)

[0065] V={v i |i=1,2,3,…,n}

[0066] E={v ij |v i ∈V,v j ∈V,i≠j}

[0067] Q = {q ij | <v i ,v j >∈E}

[0068] W={w ij |v ij ∈E}

[0069] Among them, G is the traffic road network; V represents the set of all nodes in graph G, that is, the set of road intersections; E represents the set of all road sections in graph G; W is the adjacency matrix of road section weights, which includes the adjacency matrix W1 of road section distance and the adjacency matrix W2 of road section flow density; Q represents the road section <v i ,v j > is a set of functional area attributes, which combines the node functional area attributes q of the residential area H, work area W, commercial area B, and other areas O ij Assign values ​​to 1, 2, 3, and 0 respectively.

[0070] The intersection of road sections is regarded as a node, and the traffic flow of the road is determined by the traffic flow density between two nodes. The weight of the road section represents the travel cost of the road section, including the length of the road section and the road obstruction. According to the traffic light conditions at the road intersection and the different traffic capacities of the road section, the obtained road obstruction usually includes node impedance and section impedance, which can further obtain the traffic flow density of the road section in the area.

[0071] The adjacency matrix W1 of the road segment distance represents the distance between nodes, that is, the length of the road segment, as shown below:

[0072]

[0073] Among them, d ijis the path length between two nodes in the traffic network. If node v i With v j If they are not connected, then d ij Infinity.

[0074] The adjacency matrix W2 of the road section traffic density is used to describe the network topology of the traffic density in the area. The specific formula is as follows:

[0075]

[0076] Among them, e ij represents the traffic density between nodes i and j.

[0077] In some embodiments, the load spatiotemporal prediction model of multi-graph feature fusion includes a feature extraction module, a multi-graph feature fusion module, a spatiotemporal prediction module and a load prediction module, wherein the feature extraction module is used to extract external feature variables from external historical data, the multi-graph feature fusion module is used to extract graph features of a multi-angle spatial relationship graph, the spatiotemporal prediction module is used to extract a multidimensional vector for load prediction from external feature variables, graph features of a multi-angle spatial relationship graph and the user MDP travel chain of the area to be predicted, and the load prediction module is used to predict the charging load of electric vehicles in the area to be predicted based on the multidimensional vector for load prediction.

[0078] The feature extraction module extracts external feature variables from external historical data, including:

[0079] Perform data preprocessing on external historical data to obtain the preprocessed external historical data;

[0080] External characteristic variables are extracted from the preprocessed external historical data, wherein the external characteristic variables at least include a weather sequence, a temperature sequence, a date type sequence and an economic characteristic sequence.

[0081] Specifically, the feature extraction module can also be used to convert the user travel status sequence into a numerical form using one-hot encoding, and extract spatiotemporal features from the user's MDP travel chain, including the user's travel status sequence, travel status transition probability, user's first travel time, destination stay time, etc.

[0082] In order to consider the effect of each variable equally, all the extracted features are normalized. According to the following normalization equation, the feature value is controlled between [0, 1].

[0083]

[0084] Among them, x' and x represent the eigenvalues ​​before and after normalization, x min 、x max Respectively represent the minimum and maximum values ​​in the feature data.

[0085] The multi-graph feature fusion module includes multiple graph attention network layers, multi-graph fusion layers and residual connection layers. Among them, the multiple graph attention network layers are used to extract the features of the fused neighborhood information corresponding to the regional node distance map, the traffic network map and the flow density map respectively. The multi-graph fusion layer is used to fuse the features of the fused neighborhood information corresponding to the regional node distance map, the traffic network map and the flow density map respectively to generate node representation. The residual connection layer is used to generate graph features of the multi-angle spatial relationship map based on the node representation.

[0086] Specifically, in order to better handle traffic distance graphs and traffic network topology graphs with non-Euclidean structures. A masked multi-head attention mechanism is introduced to process traffic flow density graphs belonging to directed graphs. It can adaptively process adjacent nodes and assign different weight coefficients, and can better integrate the correlation between vertex features into the model, thereby improving the accuracy of load forecasting. A multi-graph fusion layer is added after the graph attention network layer to learn the weights between different views to obtain a global node representation. Multiple graph attention network layers are superimposed to capture the spatial features of the entire area. At the same time, in order to solve the overfitting problem caused by stacking graph attention network layers, a residual connection layer is introduced. Therefore, the multi-graph feature fusion module includes a graph attention network layer, a multi-graph fusion layer, and a residual connection layer.

[0087] Taking the region nodes as attention nodes, each view uses the graph attention network layer to learn the nodes separately, obtain a new feature vector, and parameterize the attention mechanism through the learnable weight vector. The specific process is as follows:

[0088] First, enter the graph attention network layer to calculate the attention coefficient between the target node i and the adjacent node j, and then normalize the attention coefficient and pay more attention to the adjacent nodes with larger attention coefficient values, so as to better explore the distance relationship, topological connection relationship and flow density relationship between regional nodes.

[0089]

[0090] Among them, e ij represents the similarity score between node i and its adjacent node j, || represents the concatenation of the transformed features of nodes i and j, and the LeakyReLU function is applied to map the connected high-dimensional features and weight vector a to obtain the similarity score, α ij represents the attention coefficient between node i and its adjacent node j and is normalized using the softmax function.

[0091] Then, according to the calculated attention coefficient, the calculated features are weighted and summed:

[0092]

[0093] Where σ is the activation function. Here, masked multi-head attention is used. This formula means that K independent attention mechanisms are executed on node i and the average is taken to obtain h′ which is output by the graph attention network layer and integrates the domain information. i .

[0094] Then enter the multi-graph feature fusion layer to perform feature fusion and generate a new set of node representations:

[0095]

[0096] Among them, β represents the weight of each graph, N represents the three traffic network graphs, and each graph is weighted and summed according to the weight for global representation, and then enters the residual connection layer to finally generate a set of new node vector representations and input them into the spatiotemporal prediction module.

[0097] The spatiotemporal prediction module includes multiple LSTM (Long Short-Term Memory) layers and self-attention mechanism layers. Among them, multiple LSTM layers are used to extract initial multidimensional vectors from external feature variables, graph features of multi-angle spatial relationship graphs and user MDP travel chains in the area to be predicted. The self-attention mechanism layer weights the initial multidimensional vectors to generate multidimensional vectors for load forecasting.

[0098] like Figure 5 As shown in the figure, the feature vector output by the multi-graph feature fusion module is input into the LSTM layer, and the load sequence features after the graph topology information is aggregated are modeled for time series through LSTM to extract its time features. Long short-term memory neural networks can transmit effective time information and have strong nonlinear mapping and self-learning capabilities. However, due to the presence of multiple different time patterns in the input data, the unidirectional algorithm will ignore the historical global information during the training process, reducing the generalization ability of the prediction model. Therefore, a temporal convolutional network Bi-LSTM is constructed by stacking multiple LSTM layers to obtain more complex features that are highly correlated with time series and effectively handle long-term dependencies in irregular prediction tasks. The structure consists of a forward LSTM and a reverse LSTM, presenting an independent forward and backward recurrent network, and introducing a gate structure to process information, so that at time t, both the information at time t-1 and the information at time t+1 can be used, with stronger time correlation.

[0099] The external feature variables, the features output by the multi-graph feature fusion module, and the spatiotemporal features extracted from the user's MDP travel chain are used as the time series input of the spatiotemporal prediction module, and multiple time series form a multidimensional vector. The input dimensional features are recorded as X = {X1, X2, X3...XN}, and the output is recorded as Y = {Y1, Y2, Y3...YN}, where N represents the number of multidimensional time series, that is, all features extracted from the input data, including travel state sequence, travel state transition probability, user's first travel time, stay time, historical power load in the region, weather, temperature, date type (holiday, weekday) features, and spatial features obtained from the output of the multi-graph feature fusion module. The length of each sequence is n, indicating that the data of the past n time steps are used for prediction. In the hidden layer, the hidden layer states of the two LSTMs are combined according to a specific combination method. In this layer, the input is multiplied by the weight W corresponding to the gate structure, and after being activated by the sigmoid activation function, it is output through the tanh activation function. There are two parts, B and B', in this part, which represent the forward output value and the reverse output value respectively. The final output value Y depends on B and B'. The specific formula is expressed as follows:

[0100] Y n =tanh(W o B n +W o 'B n ′+b Y )

[0101] B n ′=sigmoid(W f 'B n ′+W i 'X n +b B′ )

[0102] B n =sigmoid(W f B n +W i X n +b B )

[0103] Among them, W f , W i , W o They represent the corresponding weight matrices in the forward calculation. f ′、W i ′、W o ′ respectively represent the corresponding weights in the reverse calculation. Y , b B′ , b B Respectively represent the bias matrix during calculation. Tanh and sigmoid are activation functions

[0104] After passing through the LSTM layer, the self-attention mechanism is introduced to adaptively assign importance to different features, that is, to obtain the attention weights of the corresponding features, enhance the adaptability to irregular time series data, and further improve the model's ability to handle long-term dependencies. The self-attention mechanism is introduced into each time step of the LSTM layer to help the prediction model dynamically adjust the importance of different features in each time step, calculate and filter feature information. Specifically, the self-attention mechanism is introduced in each time step of the LSTM layer, and the attention weight vector of the time step is calculated, that is, the similarity between the current time step and the previous time step, and then weighted summed with the input vector (that is, the output vector Y of the LSTM layer) to obtain a weighted representation of the new output feature vector H. By analyzing how the model further processes the feature vector, the model can more flexibly use historical information to generate the current output, improve the accuracy of the model, and make the model more interpretable. The specific formula is expressed as:

[0105] α n =softmax(V T tanh(U·Y n +b))

[0106]

[0107] Among them, V and U represent weight matrices, b represents the bias term, and α n represents the attention weight vector, and H represents the feature vector of the new output.

[0108] When using this spatiotemporal prediction module to process data, there will be problems such as gradient disappearance or explosion, overfitting and increased training costs. Therefore, the dropout regularization technology is used and the residual connection is introduced to solve this series of problems.

[0109] The load forecasting module predicts the electric vehicle charging load in the area to be predicted based on the multidimensional vector used for load forecasting, including:

[0110] Input a multidimensional vector for load forecasting according to different weights;

[0111] The charging load of electric vehicles in the area to be predicted is predicted through a fully connected layer based on the multidimensional vector for load prediction input according to different weights.

[0112] Specifically, the new feature vector H output by the multi-graph feature fusion module is input into the load forecasting module according to different weights. The load forecasting module processes the new feature vector H output by the multi-graph feature fusion module through a fully connected layer, and maps it into the final output result.

[0113] The mean square error loss function is used to evaluate the accuracy of load forecasting, and the parameters of the loss function are adjusted to train the forecasting model so that it can achieve the forecasting accuracy required by the load. The specific loss function is:

[0114]

[0115] Among them, LOSS(β) is the loss, β represents all learnable parameters in the model; n represents the number of samples; y i Represents the output of the model (i.e., predicted value); Represents the expected output of the model (i.e., the true value).

[0116] Finally, it should be understood that the embodiments described in this specification are only used to illustrate the principles of the embodiments of this specification. Other variations may also fall within the scope of this specification. Therefore, as an example and not a limitation, alternative configurations of the embodiments of this specification may be considered consistent with the teachings of this specification. Accordingly, the embodiments of this specification are not limited to the embodiments explicitly introduced and described in this specification.

Claims

1. A spatiotemporal forecasting method for electric vehicle charging load based on hybrid model driving, characterized in that: include: Build a user travel model based on Markov chain; Construct a multi-angle traffic network model; Construct a load spatiotemporal prediction model based on multi-graph feature fusion; Obtain historical user travel data, electric vehicle charging load historical data, traffic network historical information and external historical data in the area to be predicted; Generate a user MDP travel chain for the area to be predicted based on the user historical travel data of the area to be predicted by using the user travel model; Extracting a multi-angle spatial relationship diagram from the historical information of the traffic network of the area to be predicted by using the multi-angle traffic network model; The load spatiotemporal prediction model based on the multi-graph feature fusion is used to predict the charging load of electric vehicles in the area to be predicted according to the user MDP travel chain, the multi-angle spatial relationship diagram and the external historical data in the area to be predicted.

2. The method for spatiotemporal prediction of electric vehicle charging load based on hybrid model driving according to claim 1 is characterized in that: The user travel model is constructed based on the Markov chain, including: Determine a plurality of spatial states, wherein the plurality of spatial states at least include a stay state in a residential area, a stay state in a work area, a stay state in a commercial area, and a stay state in other areas; According to the multiple spatial states, construct a Markov chain to construct a state transfer matrix; Defining a plurality of model parameters of the user travel model, wherein the plurality of model parameters at least include the first travel time of each vehicle, the destination stay duration and the choice of the travel destination; A state transfer matrix and the plurality of model parameters are constructed based on the Markov chain to construct the user travel model.

3. The hybrid model-driven spatiotemporal prediction method for electric vehicle charging load according to claim 2 is characterized in that: Generating a user MDP travel chain of the area to be predicted according to the historical travel data of users of the area to be predicted by using the user travel model, including: S11, determining the probability distribution of the user's first trip time and the probability distribution of the stay time at various types of destinations based on the historical travel data of users in the area to be predicted, wherein the various types of destinations include residential areas, work areas, commercial areas and other areas; S12, selecting a user starting destination; S13, extracting the first travel time according to the probability distribution function of the first travel time, and extracting the stay time at the current destination according to the probability distribution function of the stay time at the destination; S14, randomly select the next space state according to the state transfer matrix constructed by the Markov chain, and record the state transfer path, repeat the random selection multiple times and record the state transfer path each time; S15, according to each state transition path, statistically calculate the frequency of each space state being selected, and determine the most appropriate state transition node according to the state transition with the highest frequency; S16. Repeat S13-S15 until the transfer to the residential area.

4. The method for spatiotemporal prediction of electric vehicle charging load based on hybrid model driving according to claim 3 is characterized in that: The probability distribution of residence time in residential areas is: The probability distribution of the stay time in the work area, commercial area or other areas is: Among them, f1(T mij ,k,λ) is the probability distribution of the residence time in the residential area, f2(T mij ,μ,σ,γ) is the probability distribution of the stay time in the work area, commercial area or other areas, T mij is the length of time that an electric vehicle stays at the destination; λ is the scale parameter of the probability distribution of the stay time in residential areas; k is the shape parameter of the probability distribution of the stay time in residential areas; μ is the location parameter of the probability distribution of the stay time in work areas, commercial areas or other areas; σ is the scale parameter of the probability distribution of the stay time in work areas, commercial areas or other areas; γ is the shape parameter of the probability distribution of the stay time in work areas, commercial areas or other areas, σ>0.

5. A hybrid model-driven spatiotemporal prediction method for electric vehicle charging load according to any one of claims 1 to 4, characterized in that: The multi-angle spatial relationship diagram includes a regional node distance diagram, a traffic network diagram and a flow density diagram.

6. The hybrid model-driven spatiotemporal prediction method for electric vehicle charging load according to claim 5, characterized in that: The load spatiotemporal prediction model with multi-graph feature fusion includes a feature extraction module, a multi-graph feature fusion module, a spatiotemporal prediction module and a load prediction module, wherein the feature extraction module is used to extract external feature variables from the external historical data, the multi-graph feature fusion module is used to extract graph features of a multi-angle spatial relationship graph, the spatiotemporal prediction module is used to extract a multidimensional vector for load prediction from the external feature variables, the graph features of the multi-angle spatial relationship graph and the user MDP travel chain of the area to be predicted, and the load prediction module is used to predict the electric vehicle charging load in the area to be predicted based on the multidimensional vector for load prediction.

7. The hybrid model-driven spatiotemporal prediction method for electric vehicle charging load according to claim 6 is characterized in that: The feature extraction module extracts external feature variables from the external historical data, including: Performing data preprocessing on the external historical data to obtain the preprocessed external historical data; External characteristic variables are extracted from the preprocessed external historical data, wherein the external characteristic variables at least include a weather sequence, a temperature sequence, a date type sequence, and an economic characteristic sequence.

8. The hybrid model-driven spatiotemporal prediction method for electric vehicle charging load according to claim 6 is characterized in that: The multi-graph feature fusion module includes multiple graph attention network layers, multi-graph fusion layers and residual connection layers, wherein the multiple graph attention network layers are used to extract the features of the fused neighborhood information corresponding to the regional node distance graph, the traffic network graph and the flow density graph respectively; the multi-graph fusion layer is used to fuse the features of the fused neighborhood information corresponding to the regional node distance graph, the traffic network graph and the flow density graph respectively to generate node representation; the residual connection layer is used to generate the graph features of the multi-angle spatial relationship graph according to the node representation.

9. The hybrid model-driven spatiotemporal prediction method for electric vehicle charging load according to claim 6, characterized in that: The spatiotemporal prediction module includes multiple LSTM layers and a self-attention mechanism layer, wherein the multiple LSTM layers are used to extract initial multidimensional vectors from the external feature variables, the graph features of the multi-angle spatial relationship graph, and the user MDP travel chain of the area to be predicted, and the self-attention mechanism layer weights the initial multidimensional vector to generate the multidimensional vector for load prediction.

10. The hybrid model-driven spatiotemporal prediction method for electric vehicle charging load according to claim 7, characterized in that: The load prediction module predicts the electric vehicle charging load of the to-be-predicted area according to the multidimensional vector for load prediction, including: Inputting the multidimensional vector for load forecasting according to different weights; The charging load of the electric vehicle in the area to be predicted is predicted through a fully connected layer based on the multidimensional vector for load prediction input according to different weights.

Images