Vehicle fuel consumption prediction method and system based on spatiotemporal attention mechanism graph convolutional network

Abstract

The application discloses a vehicle fuel consumption prediction method and system based on a space-time attention mechanism graph convolutional network, comprising: dynamically learning the correlation between each input factor, and constructing the correlation graph structure between each input factor and fuel consumption; adopting a neighbor attention graph convolutional network to learn the space-time characteristics of fuel consumption, so as to capture the spatial dependency between each input factor; the output of the neighbor attention graph convolutional network sequentially passes through a gated recurrent unit (GRU) and a Transformer layer, so as to capture the local dependency and global dependency of historical vehicle driving state data in the time dimension; and based on the output of the Transformer layer and a trained multi-layer feedforward neural network model, a fuel consumption prediction result of a future set period is obtained. According to the application, the fuel consumption of the next time period is predicted according to the driving state data of the driver, so that the driver can correct the driving behavior in time according to the prediction value, and the purpose of saving fuel in the whole driving process is achieved.

Description

Technical Field

[0001] This invention relates to the field of vehicle fuel consumption prediction technology, and in particular to a method and system for predicting vehicle fuel consumption based on a spatiotemporal attention-based graph convolutional network. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] With the continuous increase in vehicle usage, the fuel consumption of gasoline and diesel-powered vehicles not only increases the demand for petroleum resources, putting greater pressure on the energy sector, but also contributes significantly to air pollution through vehicle exhaust emissions. Therefore, establishing reasonable fuel consumption analysis and prediction models to save on vehicle fuel consumption is of great practical importance.

[0004] Most current fuel consumption prediction methods are based on historical vehicle driving data. They first establish the influence of factors on fuel consumption by analyzing the strength of the linear relationship between factors and fuel consumption. Then, they select highly correlated factors to build multiple linear regression models or time series prediction models to predict fuel consumption for different routes. For example, convolutional neural networks (CNNs) are used to effectively extract spatial features from grid data, or graph convolutional neural networks (GCNs) are used to describe the spatial correlation of graph-based data. However, these methods rely solely on linear methods to determine which factors affect fuel consumption, which is inaccurate. In reality, there is a complex and nonlinear correlation between fuel consumption and influencing factors. Furthermore, these methods cannot simultaneously model the spatiotemporal characteristics and dynamic correlations of vehicle operating status data, resulting in inaccurate fuel consumption predictions.

[0005] Meanwhile, some studies directly focus on driving behavior, establishing relationship models between different driving behaviors and corresponding fuel consumption characteristics, and using deep learning methods to analyze and predict fuel consumption. While these methods have achieved some results, none of them comprehensively consider the factors influencing fuel consumption in vehicle driving status data; secondly, most of them directly consider the impact of each factor on fuel consumption, while ignoring the mutual influence between factors that may indirectly affect fuel consumption. Summary of the Invention

[0006] To address the aforementioned issues, this invention proposes a vehicle fuel consumption prediction method and system based on a spatiotemporal attention-based graph convolutional network. This method predicts fuel consumption for the next time period based on vehicle driving status data, enabling drivers to adjust their driving behavior in a timely manner based on the predicted value, thereby achieving fuel savings throughout the entire driving process.

[0007] In some implementations, the following technical solutions are adopted:

[0008] A vehicle fuel consumption prediction method based on spatiotemporal attention-based graph convolutional networks includes:

[0009] Acquire vehicle driving status data, determine input factors, dynamically learn the correlation between each input factor, and construct a correlation graph structure between each input factor and fuel consumption;

[0010] Based on the aforementioned association graph structure and the time series data of vehicle driving status over the past T time periods, a graph convolutional network based on neighbor attention is used to learn spatiotemporal features in order to capture the spatial dependence between each input factor and fuel consumption.

[0011] The output of the neighbor-attention-based graph convolutional network passes through a gated recurrent unit (GRU) and a Transformer layer in sequence to capture the local and global dependencies of historical vehicle driving state data in the time dimension.

[0012] Based on the output of the Transformer layer and the trained multi-layer feedforward neural network model, the fuel consumption prediction result for a future set period is obtained.

[0013] In other embodiments, the following technical solutions are adopted:

[0014] A vehicle fuel consumption prediction system based on a spatiotemporal attention-based graph convolutional network includes:

[0015] The fuel consumption influencing factor association graph representation module is used to acquire vehicle driving status data, determine input factors, dynamically learn the correlation between each input factor, and construct the association graph structure between each input factor and fuel consumption;

[0016] The fuel consumption spatiotemporal feature learning module is used to learn spatiotemporal features based on the association graph structure and the time series data of vehicle driving status over the past T time periods. It employs a neighbor-attention-based graph convolutional network to capture the spatial dependencies between each input factor and fuel consumption. The output of the neighbor-attention graph convolutional network is then passed sequentially through a gated recurrent unit (GRU) and a Transformer layer to capture the local and global dependencies of historical vehicle driving status data in the time dimension.

[0017] The fuel consumption prediction output module obtains the fuel consumption prediction results for a future set period based on the output of the Transformer layer and the trained multi-layer feedforward neural network model.

[0018] Compared with the prior art, the beneficial effects of the present invention are:

[0019] (1) The present invention predicts fuel consumption for the next time period based on driver driving status data, so that the driver can correct his driving behavior in a timely manner based on the predicted value, thereby achieving the purpose of saving fuel throughout the driving process.

[0020] In the vehicle driving status data obtained by this invention, fuel consumption and other factors are given equal importance, which can be closer to the real situation and facilitate the next step of better exploring the spatial dependence between factors and between factors and fuel consumption.

[0021] (2) In order to better explore the nonlinear relationship between different influencing factors, this invention introduces an embedding vector to represent each input factor. These embedding vectors will be randomly initialized and trained together with the rest of the model. The similarity between these embedding vectors can initially indicate the similarity between factors, so these embedding vectors can be used to determine which factors are related when learning graph association structures.

[0022] (3) This invention employs a Neighbor Attention-Based Graph Convolutional Network (NAGCN) module to capture the complex spatial dependencies between various factors. Addressing the shortcomings of traditional fuel consumption prediction models in the temporal feature learning process, a temporal dependency modeling module composed of a gated recurrent unit (GRU) and a Transformer layer is proposed to capture the local and global dependencies of historical vehicle driving state data in the temporal dimension.

[0023] Other features and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the vehicle fuel consumption prediction method based on spatiotemporal attention mechanism graph convolutional network in an embodiment of the present invention.

[0025] Figure 2 This is a schematic diagram of a graph convolutional network structure based on neighbor attention in an embodiment of the present invention;

[0026] Figure 3 This is a schematic diagram of the Transformer layer structure in an embodiment of the present invention. Detailed Implementation

[0027] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0028] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0029] Example 1

[0030] As pointed out in the background art, existing fuel consumption analysis and prediction methods based on vehicle driving conditions and road characteristics do not fully consider the interaction between influencing factors and their direct and indirect impact on fuel consumption. Furthermore, existing fuel consumption prediction methods ignore the local and global correlation of historical vehicle driving status data over time.

[0031] Based on this, in one or more embodiments, a fuel consumption prediction method based on a spatiotemporal attention-based graph convolutional network is disclosed. To address the lack of a prior graph showing the relationship between fuel consumption and influencing factors, this method first utilizes a graph structure learning layer to dynamically learn the relationships between various factors from the input data. Then, based on this graph structure, graph convolutional networks based on first-order and second-order neighbor attention mechanisms are designed to comprehensively capture the spatial correlations among fuel consumption influencing factors. Subsequently, gated recurrent units (GRUs) and Transformer layers are used to capture the local and global dependencies of vehicle driving state data in the temporal dimension. Finally, a prediction module accurately predicts the fuel consumption value for the next time period.

[0032] Specifically, in combination Figure 1 The fuel consumption prediction method based on spatiotemporal attention mechanism graph convolutional network in this embodiment specifically includes the following process:

[0033] (1) Obtain vehicle driving status data, determine input factors, dynamically learn the correlation between each input factor, and construct the correlation graph structure between each input factor and fuel consumption;

[0034] In this embodiment, data such as vehicle speed, engine speed, torque, water temperature, coolant temperature, acceleration, fuel tank level, and fuel consumption are used as input factors.

[0035] In the absence of a correlation graph between fuel consumption and its influencing factors defined by prior knowledge, the interaction relationships between various input factors are dynamically learned based on the input vehicle driving status data. This allows for the identification of factors affecting fuel consumption and the initial construction of a correlation graph structure for fuel consumption influencing factors. Specifically, this involves the following steps:

[0036] ① Embedded vector representation of influencing factors.

[0037] In vehicle operation status data, different factors exhibit different characteristics, and complex nonlinear relationships may exist between them. To better explore the nonlinear relationships between different influencing factors, firstly, to flexibly represent the features of each input factor, an embedding vector is introduced for each factor, i.e.:

[0038] v i ∈R d for i∈{1,2,...N}

[0039] These embedding vectors will be randomly initialized and trained along with the rest of the model. We believe that the similarity between these embedding vectors can initially indicate the similarity between factors, so we use these embedding vectors during graph structure learning to determine which factors are related to each other.

[0040] (2) Determine the correlation between various factors (including fuel consumption) based on the embedding vector similarity calculation.

[0041] The association structure between influencing factors is determined by calculating the similarity between their embedding vectors. Specifically, an undirected graph is used to describe the relationships between factors, where nodes represent fuel consumption and other influencing factors, edges represent dependencies between factors, and this undirected graph is represented by an adjacency matrix A. Since prior knowledge about the correlations between factors is missing from the original data, for each factor i, its possible candidate nodes are the complete set of nodes excluding itself, denoted as Ci.

[0042] Furthermore, the dependency relationship between factor i and other factors is represented by calculating the similarity between the embedding vector of i and the embedding vector of its candidate vector j:

[0043]

[0044] ③ Construct a relationship diagram structure between each input factor and fuel consumption.

[0045] Repeat the above steps for each factor, first calculating e. ij That is, the embedding vector of factor i and its candidate vector j∈C i The normalized dot products between factors are used. However, since the similarity calculated between some factors is extremely small and can be ignored, the Top K normalized dot products are selected, and the value of k can be defined by the sparsity level. This yields the final adjacency matrix A, which is the relationship graph structure between fuel consumption and other factors.

[0046] A ij =e ij {j∈TopK({e ik :k∈Ci})}

[0047] Furthermore, the fuel consumption analysis module does not focus on fuel consumption as the central point; it only analyzes the correlation between other factors and fuel consumption. We believe that only by placing fuel consumption and other factors on an equal footing in the obtained vehicle driving status data can we more closely reflect reality and facilitate further exploration of the spatial dependencies between factors and between factors and fuel consumption.

[0048] Based on the above three steps, a correlation diagram reflecting the relationship between input factors reflecting the vehicle's driving status and the relationship between input factors and fuel consumption will be obtained.

[0049] (2) Based on the aforementioned association graph structure and the time series data of vehicle driving status in the past T time periods, a graph convolutional network based on neighbor attention is used to learn the spatiotemporal features of fuel consumption in order to capture the spatial dependence between various input factors.

[0050] This embodiment combines a correlation graph of fuel consumption influencing factors, which is dynamically learned from the input data. It uses spatiotemporal feature learning to fuse the spatiotemporal dependencies between influencing factors and fuel consumption in historical vehicle driving status data. The specific process is as follows:

[0051] ① Spatial feature learning based on neighbor attention graph convolutional networks.

[0052] Based on the graph representation layer, the relationships between factors are initially learned by the similarity between the feature vectors of the input factors. However, since the vehicle's operating status data is dynamically and non-linearly changing, accurate fuel consumption prediction requires a deeper modeling of the dynamic spatial correlations between the factors. Therefore, we further propose a Neighbor Attention-Based Graph Convolutional Network (NAGCN) module to capture the complex spatial dependencies between factors.

[0053] Specifically, the input to the spatial feature learning module based on the neighbor attention graph convolutional network is the time series data of vehicle driving status {X1,X2,…,X} over the past T time periods. T The adjacency matrix A obtained through graph structure learning in the previous section is used to input the data vector corresponding to each time period and the adjacency matrix A into a NAGCN module for feature learning.

[0054] like Figure 2 As shown, the NAGCN module consists of two independent parts: a first-order neighbor attention graph convolutional network (1-NAttGCN) and a second-order neighbor attention graph convolutional network (2-NattGCN), which are used to capture the direct influence of first-order neighbors and the indirect influence of second-order neighbors between factors and fuel consumption, respectively. The specific implementation steps are as follows:

[0055] 1) First-order neighbor attention graph convolutional network module

[0056] In the first-order neighbor attention graph convolutional network module, for the first-order neighbors in the learned graph structure that directly affect fuel consumption, the attention weights between factors obtained from the graph structure learning module are corrected using the original input data. This fuses the information between nodes with their first-order neighbor information to further capture the direct influence relationships between factors. Furthermore, to better represent nodes, the factor embedding vectors obtained from the graph structure learning layer are integrated into the attention mechanism. Therefore, the aggregated representation h of factor i... i as follows:

[0057]

[0058] Where, x i ∈R w It is the input feature of the i-th factor, N (i) ={j|A ij >0} is the set of first-order neighbor factors of factor i in the adjacency matrix A obtained from the graph structure learning module, where W1∈R d×w This is a trainable weight matrix applied to each node in the graph. Note the coefficient α in the above formula. i,j The calculation formula is:

[0059]

[0060]

[0061]

[0062] in This indicates a connection, therefore g i The factor embedding vector v i and related transformation features W1x i The attention coefficients are calculated using the LeakyReLU function as a non-linear activation function and normalized using the softmax function.

[0063] 2) Second-order neighbor attention graph convolutional network module

[0064] like Figure 2As shown, under real-world conditions, some factors are not directly related to fuel consumption (we call them second-order neighbors), but they indirectly affect fuel consumption through factors that directly influence it (first-order neighbors). For example, the coolant temperature variance factor has a very small direct correlation with fuel consumption, but it has a strong influence on the velocity variance, a first-order neighbor of fuel consumption. Therefore, we believe that the coolant temperature variance should also have an indirect impact on fuel consumption. Based on this, this embodiment uses a 2-NAttGCN layer to analyze the influence of the node's second-order neighbors. The factor representation h' i It can be expressed like this:

[0065]

[0066] π(i,p)=π(i,j)×π(j,p)

[0067]

[0068]

[0069] Where M(i)={p|A ip =0,A jp >0} is the set of second-order neighbors of factor i obtained from the learned adjacency matrix A, and W2 is also a trainable weight matrix.

[0070] (3) The output of the neighbor attention graph convolutional network passes through the gated recurrent unit (GRU) and the Transformer layer in sequence to capture the local and global dependencies of historical vehicle driving status data in the time dimension.

[0071] To address the shortcomings of traditional fuel consumption prediction models in the learning of time features, this embodiment proposes a time dependency modeling module consisting of a gated recurrent unit (GRU) and a Transformer layer, which captures the local and global dependencies of historical vehicle driving status data in the time dimension.

[0072] 1) GRU captures local time dependencies

[0073] Gated Recurrent Units (GRUs) are used to process time series information. First, a hidden vector representation is maintained for each time step. When representing the hidden vector at time step t, the hidden state of the GRU at the previous time step t-1 and the output of the 2-NAttGCN convolutional module at time step t are used as input to obtain the output at the current time step t. Furthermore, the GRU operation is applied individually to each node, and the GRU parameters are shared across all nodes. Specifically, for each node v with timestamp t, given the output X of the previous layer... t The hidden vector Z from the previous time stept-1 The operation of GRU can be represented as follows:

[0074] u t =σ(W u [f(A, X) t ), z t-1 ]+b u )

[0075] r t =σ(W r [f(A, X) t ), z t-1 ]+b r )

[0076] C t =tanh(W c [f(A, X) t ), (r t *z t-1 )]+b c )

[0077] z t =u t *z t-1 +(1-u t )*c t

[0078] Among them W u W r W c The parameter to be learned, z t It serves as the output for the current time step and also as the input for the next time step.

[0079] 2) The Transformer layer captures global time dependencies.

[0080] The GRU layer can capture time information within a local range. However, in the problem of vehicle fuel consumption prediction, time information may not only be sequentially dependent, but should also be analyzed from a global perspective. Therefore, after the GRU layer, we use a Transformer layer to directly capture global dependencies in the time dimension.

[0081] Specifically, similar to the GRU layer, the Transformer layer is also applied individually to each time point. For each node v, we will use the output of the GRU layer. As input to the Transformer layer. Because the Transformer layer's attention mechanism treats different positions equally while ignoring relative positional information when calculating the attention function, a positional encoding e is first applied to each position. t By Z vCombined with positional encoding, a new representation Z' is generated. t ,Right now:

[0082] Z' t =Z v +e t

[0083]

[0084] Use a sequence with positional encoding as input to the Transformer layer. For example... Figure 3 As shown, the Transformer layer consists of a multi-head self-attention layer, a shared feedforward neural working layer, and a batch normalization layer between them. First, the input to the self-attention function includes a query of dimension d. k The key and the dimension of all positions in the time series are d. v The value of . Calculate the dot product of the given query with all keys, and divide each dot product by . Then, the softmax function is applied to obtain the attention score for each position. These attention scores will be used as weights to aggregate information from the corresponding values. We will simultaneously calculate the attention score for the query at all positions:

[0085]

[0086] Where, Q, This represents the queries, keys, and values ​​of all nodes. The output sequence of the GRU layer is organized in matrix form by stacking the elements according to their sequence positions. The resulting matrix is ​​then represented as Z. v ∈R T×d The superscript v i This indicates that he corresponds to node v. i Then matrix Z v Linear projection onto the query, key, and value is shown below:

[0087] Q v =Z v W Q K v =Z v W K V = Z v W V

[0088] Among them, W Q W K W V This is the projection matrix that needs to be learned and is shared by all nodes. Therefore, the attention calculation function can be rewritten as:

[0089]

[0090] A multi-head attention model is employed to aggregate representation information from different subspaces to enhance the model's representational power. In multi-head attention, a total of K sets of projection matrices are used to project Z... v Projected onto K different sets of queries, keys, and values. The result of multi-head attention is a concatenation of the outputs of each single-head attention function, which can be specifically represented as:

[0091] Multihead(Z v = Concat(head1, ..., head) S W O

[0092]

[0093] in, W is the projection matrix of the s-th attention head. O It is another type of linear output projection.

[0094] After the multi-head attention layer, the output state is passed to a pointwise feedforward neural network layer. For example... Figure 3 As shown, after each sub-layer, there is a residual connection and a normalization layer. Finally, the output of the Transformer layer is obtained, which can be represented as H. v ∈R T×d .

[0095] (4) Based on the output of the Transformer layer and the trained multi-layer feedforward neural network model, the fuel consumption prediction result for a future set period is obtained.

[0096] Combining the information learned from the preceding spatiotemporal feature training, the average fuel consumption of the vehicle in the next time period is output. Specifically, after the Transformer layer, {H} is used. v Using |v∈V} as input, a multi-layer feedforward network is used to predict fuel consumption in future time periods. During training, the vehicle state data for historical time periods is first split into a training component Y. train Test section Y test And the verification part Y valid Then, in the training data Y train A window of length T+T' is slid upwards to generate training sequences. For each generated training sequence, T elements are used as input, and the remaining T' elements are used as ground truth. The mean absolute error is then used as the loss, which can be expressed as:

[0097]

[0098]

[0099]

[0100] Here, f() is the prediction model, which can be trained by minimizing the loss formula. The prediction part adopts iterative prediction, that is, after the model is input with the observations of T time steps, it predicts the value of the next time step T+1, and then concatenates the predicted value with the previous T-1 observations as input to predict the value of the next time step, and so on.

[0101] Example 2

[0102] In one or more embodiments, a vehicle fuel consumption prediction system based on a spatiotemporal attention mechanism graph convolutional network is disclosed, comprising:

[0103] The fuel consumption influencing factor association graph representation module is used to acquire vehicle driving status data, determine input factors, dynamically learn the correlation between each input factor, and construct the association graph structure between each input factor and fuel consumption;

[0104] The fuel consumption spatiotemporal feature learning module is used to learn fuel consumption spatiotemporal features based on the association graph structure and the time series data of vehicle driving status over the past T time periods, using a neighbor attention graph convolutional network to capture the spatial dependencies between various input factors. The output of the neighbor attention graph convolutional network is then passed through a gated recurrent unit (GRU) and a Transformer layer to capture the local and global dependencies of historical vehicle driving status data in the time dimension.

[0105] The fuel consumption prediction output module obtains the fuel consumption prediction results for a future set period based on the output of the Transformer layer and the trained multi-layer feedforward neural network model.

[0106] It should be noted that the specific implementation methods of the above modules have been described in Example 1, and will not be detailed here.

[0107] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A vehicle fuel consumption prediction method based on a spatio-temporal attention mechanism graph convolutional network, characterized in that, The input factors at least include vehicle speed, engine speed, torque, water temperature, coolant temperature, acceleration, fuel tank level and fuel consumption. The vehicle driving state data is acquired, input factors are determined, the correlation between each input factor is dynamically learned, and a correlation graph structure between each input factor and fuel consumption is constructed, specifically including: learning an embedding vector representation for each input factor, i.e. , using the embedding vector to determine which factors are correlated with each other during graph structure learning; determining the correlation between each factor based on the similarity between the embedding vectors; constructing a correlation graph structure between fuel consumption and other factors based on the correlation, introducing an embedding vector for each factor to represent; representing the dependency relationship between i the embedding vector of the candidate vector j of the embedding vector of the candidate vector i between each factor and other factors by calculating the similarity between the embedding vector of the candidate vector Select The value of k can be defined by the sparsity level; the final adjacency matrix A is obtained, which is the graph structure between fuel consumption and other factors. ; based on the association graph structure and vehicle driving state time series data of a past time period, a neighbor attention-based graph convolution network is used for spatio-temporal feature learning to capture the spatial dependency between each input factor and fuel consumption; the output of the neighbor attention-based graph convolution network is sequentially subjected to a gated recurrent unit (GRU) and a Transformer layer to capture the local dependency and global dependency of historical vehicle driving state data in the time dimension; based on the output of the Transformer layer and a trained multi-layer feedforward neural network model, a fuel consumption prediction result for a future set period is obtained. The neighbor-attention-based graph convolution network includes a first-order neighbor-attention graph convolution network and a second-order neighbor-attention graph convolution network, which are respectively used to capture direct influence of first-order neighbors and indirect influence of second-order neighbors between the input factors and the fuel consumption; in the first-order neighbor-attention graph convolution network, for the first-order neighbors having direct influence on the fuel consumption in the associated graph structure, the attention weights between the factors obtained from the associated graph structure are corrected by the original input data, the information between the nodes is fused with the first-order neighbor information, and the direct influence relationship between the input factors is further captured; in order to better represent the nodes, the factor embedding vectors obtained from the associated graph structure are integrated into the attention mechanism. The associated graph structure between the fuel consumption and other factors is constructed based on the correlation, and specifically includes: 2.The vehicle fuel consumption prediction method based on the spatio-temporal attention mechanism graph convolutional network according to claim 1, wherein, The training process of the multi-layer feedforward neural network model specifically includes: computing a normalized dot product between the embedding vector of the input factor and its candidate vector ​​ selecting one normalized dot product, constructing an adjacency matrix A; wherein, for each factor dependent candidate nodes, i.e. the complete set except itself, the value is determined according to the sparsity level. 3.The vehicle fuel consumption prediction method based on the spatio-temporal attention mechanism graph convolutional network according to claim 1, wherein, The gating recurrent unit GRU reserves a hidden vector representation for each time step; when representing the hidden vector of the first time step, the hidden state of the GRU at the previous time step and the output of the second-order neighbor attention graph convolution module at the time step are jointly taken as inputs to obtain the output of the current time step. ​ 4.The vehicle fuel consumption prediction method based on the spatio-temporal attention mechanism graph convolutional network according to claim 3, wherein, For each node , the output of the GRU is taken as input to a Transformer layer, for each position first employing a positional encoding , generating a sequence with positional encoding by combining with the positional encoding, using the sequence with positional encoding as input to the Transformer layer. 5.The vehicle fuel consumption prediction method based on the spatio-temporal attention mechanism graph convolutional network according to claim 1, wherein, The average absolute error is taken as the loss, and the loss formula is minimized for training; vehicle state data of a historical time period is split into a training portion, a test portion, and a validation portion; a window of length is slid over the training portion data to generate training sequences; For each generated training sequence, one element is used as input, the rest elements are used as truth values; After the model inputs the observation values of T time steps, the value of the next time step T+1 is predicted, then the predicted value and the previous T-1 observation values are spliced as input to predict the value of the next time step, and the cycle is repeated. The oil consumption influencing factor associated graph representation module is configured to acquire vehicle driving state data, determine input factors, dynamically learn the correlation between the input factors, and construct an associated graph structure between the input factors and the fuel consumption.

6. A vehicle fuel consumption prediction system based on a spatio-temporal attention mechanism graph convolution network, adopting a vehicle fuel consumption prediction method based on a spatio-temporal attention mechanism graph convolution network according to any one of claims 1-5, characterized in that, The output of the neighbor-attention-based graph convolution network is sequentially subjected to a gated recurrent unit (GRU) and a Transformer layer, so as to capture local dependence and global dependence of the historical vehicle driving state data in the time dimension. The fuel consumption prediction result output module obtains the fuel consumption prediction result of a future set period based on the output of the Transformer layer and the trained multi-layer feedforward neural network model. The fuel consumption spatiotemporal feature learning module is used to learn fuel consumption based on the aforementioned association graph structure and past data. We used a neighbor-attention-based graph convolutional network to learn spatiotemporal features from time-series data of vehicle driving status over several time periods in order to capture the spatial dependence between various input factors and fuel consumption. ​ ​

Images