Method and system for estimating urban mobile source pollution emissions using incomplete information

By decomposing urban traffic data into complete and missing sections of information, and reconstructing the emission sequence using graph convolution and gated linear units, the problem of large errors in urban pollution emissions in the existing technology is solved, and more refined and accurate pollution emissions are achieved predict.

CN114036135BActive Publication Date: 2024-11-26UNIV OF SCI & TECH OF CHINA
View PDF 5 Cites 0 Cited by

Patent Information

Application Number
CN202111192115.1
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-10-13
Publication Date
2024-11-26
Estimated Expiration
2041-10-13

AI Technical Summary

Technical Problem

The existing traffic emission model has large errors and insufficient data volume in the refined emission estimation task within the city.

Method used

A method of using incomplete information is adopted to obtain data related to motor vehicle driving and urban traffic network data in urban traffic, which is divided into complete information and missing information sections, and the road network is extracted by using graph convolution units and gated linear units. Topological characteristics and timing dependence of motor vehicle emission sequences, reconstruct the emission sequence on the missing road section, and fill in the missing emission distribution information.

Benefits of technology

It realizes that the pollution value error is reduced when the data volume is small, and the refined estimation accuracy of urban mobile sources' pollution emissions can be improved, so as to accurately calculate the traffic emission sequence within a few hours of historical hours of all sections of the urban road network.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114036135B_ABST
    Figure CN114036135B_ABST
Patent Text Reader

Abstract

The present invention provides a method and system for estimating urban mobile source pollution emissions using incomplete information, including dividing the road sections into two categories according to the density of data obtained on the road sections: information-missing road sections that lack traffic mode information due to sparse data collection and information-complete road sections with sufficiently dense data; constructing a continuous velocity field and inferring flow values ​​for the information-complete road sections through the above-mentioned motor vehicle driving related data, and calculating the corresponding traffic emission values ​​through a micro-emission factor model; using graph convolution units and gated linear units to respectively extract the topological features of the road network and the temporal dependence of the motor vehicle emission sequence on the information-complete road sections, thereby reconstructing the emission sequence on the information-missing road sections and filling in the missing emission distribution information. The present invention uses less urban traffic data to mine more traffic modes and emission distribution characteristics, and realizes the task of refined estimation of traffic pollution emissions at the urban road section level.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of environmental detection and processing, and in particular to a method and system for estimating urban mobile source pollution emissions using incomplete information. Background Art

[0002] Traffic emission models are a major means widely used in the field of transportation to quantify motor vehicle emissions and calculate motor vehicle emission factors. According to the main application parameters, emission models are divided into emission models based on average speed and emission models based on driving conditions. The former mainly extracts a representative average speed value of motor vehicles from the overall driving characteristics contained in the driving data of motor vehicles of a specific model, and calculates the corresponding emission value based on this. Therefore, the model usually focuses on macro-level research and is used to estimate or predict the total traffic emissions in a specific area (usually a district or city) in a specific time period (usually a quarter or a year). Emission models based on driving conditions usually focus on micro-level research. This type of model mainly analyzes the emission of motor vehicles through the complete driving process based on the transient second-by-second data obtained when the motor vehicle is driving. It is usually suitable for analyzing the emission of a single vehicle or the emission calculation of a specific number of roads. However, the above two methods have defects in the task of refined emission estimation within the city, and there are challenges such as large errors or insufficient data volume. Summary of the invention

[0003] The present invention proposes a method and system for estimating urban mobile source pollution emissions using incomplete information, which can solve the problem that the existing method has large errors in estimating pollution values ​​when the amount of data is small.

[0004] To achieve the above object, the present invention adopts the following technical solutions:

[0005] A method for estimating urban mobile source pollution emissions using incomplete information comprises the following steps:

[0006] S10, obtaining the relevant data of motor vehicle driving in urban traffic and pre-processing the urban traffic network data and constructing the data space;

[0007] According to the density of data obtained on the road section, the road sections are divided into two categories: information-missing sections with missing traffic mode information due to sparse collected data and information-complete sections with sufficiently dense data;

[0008] S20, constructing a continuous speed field and inferring flow values ​​for the road section with complete information through the above motor vehicle travel related data, and calculating the corresponding traffic emission values ​​through a micro emission factor model;

[0009] S30. Graph convolutional units and gated linear units are used to respectively extract the topological features of the road network and the temporal dependencies of the emission sequences of motor vehicles on sections with complete information, thereby reconstructing the emission sequences on sections with missing information and filling in the missing emission distribution information.

[0010] Furthermore, the S10 obtains the motor vehicle driving related data in urban traffic and the urban traffic network data preprocessing and constructing the data space; according to the density of the data obtained on the road section, the road section is divided into two categories: information missing sections with missing traffic mode information due to sparse collected data and information complete sections with sufficiently dense data, specifically including:

[0011] S11: Urban road network data is constructed into a directed graph Given a road network, first take the set of all road forks as the edge ε connecting the road segments in the graph, then the vertex between two edges Represents a collection of N road segments, each of which has its own attributes, including the maximum lane speed value rv m , the number of lanes rn and the length of the road section r.len; the adjacency matrix W is abstracted into the adjacency correlation between vertices, and W is used to represent the connectivity between road sections: if road sections i and j are connected, then W i,j =1, otherwise W i,j =0;

[0012] S12: The driving data of the motor vehicle is the GPS trajectory of the sampled motor vehicle. Each GPS trajectory P contains a series of time-series GPS points, where each GPS point p contains a timestamp t and a geographic spatial coordinate g, i.e., p = (t, g). First, the instantaneous speed of the GPS point is calculated according to the change of geographic location and time: Given two GPS points p 1 and p 2 , then point p 1 The speed value is:

[0013]

[0014] Wherein dist(·) is a function for calculating the distance between two points, which is Manhattan distance, i.e., first-order norm in the present invention; each GPS point is matched to the corresponding road section using a map matching algorithm, so the attribute of each track point is expanded to p=(t, g, v, r);

[0015] S13: Next, define a concept sampling density; first count the total number of GPS points q recorded on the road section r within one hour r , then the sampling density of this section is d r :

[0016] d r =q r / r.len

[0017] Through analysis, sections with sampling density less than 20 are considered as sections with missing information. The road sections with sampling density greater than or equal to 20 are regarded as complete information sections.

[0018] Furthermore, the step S20, constructing a continuous speed field and inferring flow values ​​for the road section with complete information through the above motor vehicle driving related data, and calculating the corresponding traffic emissions through the micro emission factor model specifically includes:

[0019] S21: Given road segment The GPS point set {p 1 , p 2 ,…,p i ,…,p n |p i =(t i , g i , v i , r)} as input, firstly reconstruct the continuous velocity field V of the road section through the Gaussian adaptive sliding method r (t):

[0020]

[0021] Among them, the smoothing kernel function φ i The value of (t) decays as the time distance increases, that is, when the data point is far away from the target calculation point, the function will generate a smaller weight coefficient; therefore, the form of the kernel function is:

[0022]

[0023] Among them, τ is the time width. The larger the value of τ, the greater the influence of the data point on the target calculation point; the normalization function Φ(t) is the sum of all kernel functions:

[0024]

[0025] Therefore, all road segments in the urban road network have their continuous velocity fields constructed according to the above formula;

[0026] S22: Continuous velocity field V for a given section r r , the continuous emission rate TP is calculated using the EMIT micro emission factor model l (t); The EMIT model consists of two modules: engine emission module and tailpipe emission module; First, define the vehicle traction force P tract :

[0027]

[0028] The definitions of the parameters are as follows:

[0029] A: rolling resistance term; B: speed correction for rolling resistance term; C: air resistance term; M: engine mass; g: gravity constant; θ: road slope; l is defined as a common emission type (l = CO, HC, NO x ), then the engine emission rate EO of the emission l l The expression is:

[0030]

[0031] The parameter values ​​are as follows:

[0032] CO HC <![CDATA[NO x ]]> α 0.0316 0.00916 -0.00391 β 0.0 0.0 0.000305 δ 1.09e-07 7.55e-09 2.27e-08 ζ 0.00883 0.00111 0.00307 α′ 0.0261 0.00528 0.00323

[0033] The tailpipe emission rates TPl of different pollutants under the tailpipe emission module have different calculation formulas:

[0034]

[0035] The parameter values ​​are as follows:

[0036]

[0037]

[0038] S23: Given a continuous velocity field of road section r, estimate the corresponding traffic flow according to the relevant model of traffic theory: the traffic flow Q of road section r r (t) and density k r (t) and speed V r The relationship between (t) is:

[0039] Q r (t) = k r (t)×V r (t)

[0040] Since the density is unknown, the density value is inferred from the speed, and then the traffic flow value is calculated by the above formula; using the speed-density model:

[0041]

[0042] Among them, v m is the maximum speed, k c is the optimal density; therefore, the density According to the above formula, we can get:

[0043]

[0044] Therefore, all road sections in the urban road network have their traffic flows estimated according to the above formula;

[0045] S24: Continuous emission rate TP for a given road segment r l (t), traffic flow Q r (t) and the road section attributes, the change of motor vehicle emissions on the road section is described by the following function, and its emission value is:

[0046]

[0047] Furthermore, the S30 uses a graph convolution unit and a gated linear unit to respectively extract the topological features of the road network and the temporal dependency of the emission sequence of motor vehicles on the road section with complete information, thereby reconstructing the emission sequence on the road section with missing information and filling in the missing emission distribution information, specifically including:

[0048] S31: First, construct the vertex feature matrix of the graph structure in the Tth period The set of information-complete sections is The corresponding road segment in its vertex feature matrix The characteristic value calculation uses the emission value E obtained in step S20 r (t), calculate the emission values ​​of various pollutants on the road section:

[0049]

[0050] For the road sections with missing corresponding information in the vertex feature matrix Set its characteristic value to 0;

[0051] From the complete set of information Randomly delete 20% of the vertex eigenvalues ​​and set the eigenvalues ​​of the deleted vertices to 0 to obtain a new feature matrix Among them, the vertex set of deleted eigenvalues ​​is The set of vertices that remain is Since the network task is to estimate the historical emissions of motor vehicles in 6 time intervals, the network input feature matrix is ​​constructed The target output matrix is

[0052] S32: Given the topological structure and partial feature information of a traffic road network, a spatiotemporal data interpolation model based on a generative adversarial network is designed to generate real feature values, i.e., emission values, for sections with missing information in the road network. The adversarial model consists of two adversarial neural networks, i.e., a feature generator based on an autoencoder and a feature discriminator: the feature generator captures the data distribution and generates data as similar to the real samples as possible to confuse the discriminator; the discriminator distinguishes the real samples from those generated by the generator; both networks participate in the maximum-minimum game of adversarial learning to improve the authenticity of the generated graph features; after fine-tuning the network parameters through training, the generator provides complete and real node features as output based on the partial feature data input;

[0053] S33: In the autoencoder-based feature generator: First, the input graph feature matrix of the network Normalized to the range of [0, 1]; the generator is first composed of 3 consecutive spatiotemporal encoding blocks in series for feature learning; each spatiotemporal encoding block is composed of a gated linear unit and a graph convolution unit in series: where the gated linear unit is a convolution layer output without nonlinearization multiplied by a convolution layer output with nonlinear activation; therefore, given an input matrix The output of the gated linear unit operation *T for:

[0054]

[0055] in is the convolution kernel parameter;

[0056] At the same time, the graph convolution unit uses the connectivity of the graph, that is, the adjacency matrix, as a convolution filter for neighborhood mixing; based on the concept of spectral graph convolution, the graph convolution unit operation For the graph signal A nonlinear function whose output equal

[0057]

[0058] Among them, the Fourier basis is the eigenvector matrix of the normalized graph Laplacian matrix L; the Laplacian matrix in is the unit matrix, is a diagonal matrix with diagonal element values ​​d ii =∑ j W ij ; It is the eigenvalue diagonal matrix of matrix L, so Θ(Λ) is also a diagonal matrix, Θ is the graph convolution kernel parameter;

[0059] Given the above gated linear unit and graph convolution unit operations, the spatiotemporal encoding block is defined as Operation, that is, given the input matrix of the i-th spatiotemporal coding block and the kernel parameter Ξ j =[Γ j Θ j ], the output of the spatiotemporal coding block for:

[0060]

[0061] Among them, σ 1 (·), σ 2 (·) is the activation function; therefore, the output after the above three spatiotemporal encoding blocks is:

[0062]

[0063] Among them, all activation functions are elu activation functions; after three spatiotemporal encoding blocks, the input part of the graph features is converted into an N×512 intermediate feature matrix; the matrix is ​​then decoded by three spatiotemporal decoding blocks; the structure of the spatiotemporal decoding block is the same as that of the spatiotemporal encoding block, and its network structure is expressed by the formula:

[0064]

[0065] All the above activation functions are elu functions.

[0066]

[0067] The activation function σ in this step 2 is the sigmoid activation function; the graph features are reconstructed through three spatiotemporal decoding blocks;

[0068] S34: The graph feature discriminator is used to distinguish the graph feature matrix generated by the above feature generator and the real graph feature matrix X; the network starts with a spatiotemporal coding block, whose output is expanded into a one-dimensional matrix and then fed into three fully connected layers, that is, the spatiotemporal coding block is used to extract hidden graph features from the input data, and then the fully connected neurons in each fully connected layer work together to classify the extracted features to determine whether the input data is real; finally, a sigmoid function is used to generate a probability between 0 and 1, indicating the real probability of the discriminator's judgment of the input feature;

[0069] A LayerNorm layer is added after each operation and each fully connected layer in the spatiotemporal encoding block. Or X, the output of the spatiotemporal encoding block is:

[0070]

[0071] Where LN(·) represents the LayerNorm layer. Expand into a one-dimensional matrix It is used as the input of the next three fully connected layers. The activation function of each layer is the elu activation function and a LayerNorm operation is added. The activation function of the last fully connected layer is the sigmoid function. The output of the final discriminator is For input The output of the discriminator is For input X, the output of the discriminator is

[0072] S35: In the training of generative adversarial networks in adversarial learning, the training objective of the feature generator is expressed as:

[0073]

[0074] The training objective of the feature discriminator is expressed as:

[0075]

[0076] Furthermore, in step S35, a gradient penalty term is added to the loss function of the feature discriminator during the training process, so the loss function L of the feature discriminator is D for:

[0077]

[0078] Where λ is the tuning parameter.

[0079] Furthermore, in step S35, λ is an adjustment parameter, and its value is 10.

[0080] Furthermore, in step S35, an estimated mean square error term is added to the loss function of the generator, so the loss function L of the feature generator is G for:

[0081]

[0082] The Adam optimizer is used to alternately train the feature generator and the feature discriminator until the two loss functions converge to a stable state.

[0083] On the other hand, the present invention also discloses a system for estimating urban mobile source pollution emissions using incomplete information, comprising the following units:

[0084] Data acquisition and processing unit, used to acquire motor vehicle driving related data in urban traffic and pre-process urban traffic network data and construct data space;

[0085] The traffic emission value calculation unit is used to divide the road sections into two categories according to the density of data obtained on the road sections: information-missing road sections with missing traffic mode information due to sparse data collection and information-complete road sections with sufficiently dense data; construct a continuous speed field and infer flow value for the information-complete road sections through the above-mentioned motor vehicle driving related data, and calculate the corresponding traffic emission value through the micro emission factor model;

[0086] The emission distribution information filling unit is used to use the graph convolution unit and the gated linear unit to respectively extract the topological characteristics of the road network and the temporal dependencies of the motor vehicle emission sequence on the information-complete section, so as to reconstruct the emission sequence on the information-missing section and fill in the missing emission distribution information.

[0087] In a third aspect, the present invention further discloses a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes the steps of the above method.

[0088] It can be seen from the above technical scheme that the method and system for estimating urban mobile source pollution emissions using incomplete information of the present invention aims to use less urban traffic data to mine more traffic patterns and emission distribution characteristics, so as to overcome the shortcomings of existing methods and realize the task of refined estimation of traffic pollution emissions at the urban road section level.

[0089] In this embodiment, the road sections are divided into two categories according to the completeness of the motor vehicle driving information obtained, and different ideas are designed for calculating the motor vehicle emissions of the two types of road sections. For the road sections with complete information, a relatively accurate continuous speed field and flow change can be reconstructed based on the obtained data, and the motor vehicle emission value of the road section can be calculated as accurately as possible using the micro emission factor calculation model; for the road sections with missing information, the emission sequence of the road section is estimated by using the spatial topological characteristics of the road network and the temporal dependence of the motor vehicle emissions in the road sections with complete information, thereby calculating the traffic emission sequence of all road sections in the urban road network within a few hours in history. BRIEF DESCRIPTION OF THE DRAWINGS

[0090] Figure 1 and Figure 2 It is the urban mobile source emission estimation method flow;

[0091] Figure 3 It is a generative adversarial network based on graph convolution and gated linear units. DETAILED DESCRIPTION

[0092] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments.

[0093] like Figure 1 and Figure 2 As shown, the method for estimating urban mobile source pollution emissions using incomplete information described in this embodiment includes the following steps:

[0094] Step 1: Preprocess the data related to motor vehicle driving in urban traffic and urban traffic network data and construct the data space. According to the density of data obtained on the road section, the road section is divided into two categories: information-missing sections with missing traffic mode information due to sparse data collection and information-complete sections with sufficiently dense data.

[0095] Step 2: Use the above motor vehicle driving related data to construct a continuous speed field and infer flow values ​​for the complete information section, and use this to calculate the corresponding traffic emissions through the micro emission factor model.

[0096] Step 3: In order to reconstruct the emission distribution of several hours of history on sections with missing information, the idea of ​​generative adversarial learning is used to use graph convolutional units and gated linear units to extract the topological features of the road network and the temporal dependencies of the emission sequences of motor vehicles on sections with complete information, thereby reconstructing the emission sequences on sections with missing information and filling in the missing emission distribution information.

[0097] The following are specific instructions:

[0098] Furthermore, the above step S1: pre-processing and constructing data space for motor vehicle driving related data and urban traffic road network data in urban traffic. According to the density of data obtained on the road section, the road section is divided into two categories: information-missing road sections with missing traffic mode information due to sparse collected data and information-complete road sections with sufficiently dense data points, which specifically include the following subdivision steps S11 to S13:

[0099] S11: Urban road network data can be constructed into a directed graph Given a road network, first take the set of all road forks as the edge ε connecting the road segments in the graph, then the vertex between two edges Represents a collection of N road segments, each of which has its own attributes, including the maximum lane speed value rv m , the number of lanes rn and the length of the road section r.len. The adjacency matrix W can be abstracted into the adjacency correlation between vertices. In the present invention, W is used to represent the connectivity between road sections: if road sections i and j are connected, then W i,j =1, otherwise W i,j =0.

[0100] S12: The vehicle driving related data is the GPS track of the sampled vehicle. Each GPS track P contains a series of time-series GPS points, where each GPS point p contains a timestamp t and a geographic spatial coordinate g, i.e., p = (t, g). First, the instantaneous speed of the GPS point is calculated based on the change in geographic location and time: Given two GPS points p 1 and p 2 , then point p 1 The speed value is:

[0101]

[0102] Where dist(·) is a function for calculating the distance between two points, which is Manhattan distance, i.e., first-order norm, in the present invention. Each GPS point is matched to the corresponding road section using a map matching algorithm, so the attributes of each track point can be expanded to p = (t, g, v, r).

[0103] S13: Next, we define a concept: sampling density. First, we count the total number of GPS points q recorded on the road section r within one hour. r , then the sampling density of this section is d r :

[0104] d r =q r / r.len

[0105] Through analysis, sections with sampling density less than 20 are considered as sections with missing information. The road sections with sampling density greater than or equal to 20 are regarded as complete information sections.

[0106] The above step S2: constructs a continuous speed field and infers flow value for the complete road section through the above motor vehicle driving related data, and calculates the corresponding traffic emissions through the micro emission factor model. Specifically includes the following subdivision steps S21 to S24:

[0107] S21: Given road segment The GPS point set {p 1 , p 2 ,…,p i ,…,p n |p i =(t i , g i , v i , r)} as input, firstly reconstruct the continuous velocity field V of the road section through the Gaussian adaptive sliding method r (t):

[0108]

[0109] Among them, the smoothing kernel function φ i The value of (t) decays as the time distance increases, that is, when the data point is far away from the target calculation point, the function will generate a smaller weight coefficient. Therefore, the form of the kernel function is:

[0110]

[0111] Among them, τ is the time width. The larger the value of τ, the greater the influence of the data point on the target calculation point. The normalization function Φ(t) is the sum of all kernel functions:

[0112]

[0113] Therefore, all road sections in the urban road network can construct their continuous velocity fields according to the above formula.

[0114] S22: Continuous velocity field V for a given section r r , the continuous emission rate TP is calculated using the EMIT micro emission factor model l (t). The EMIT model consists of two modules: engine emission module and tailpipe emission module. First, define the vehicle traction force P tract :

[0115]

[0116] The definitions of the parameters are as follows:

[0117] A: rolling resistance term; B: speed correction for rolling resistance term; C: air resistance term; M: engine mass; g: gravity constant; θ: road slope; l is defined as a common emission type (l = CO, HC, NO x ), then the engine emission rate EO of the emission l l The expression is:

[0118]

[0119] The parameter values ​​are as follows:

[0120] CO HC <![CDATA[NO x ]]> α 0.0316 0.00916 -0.00391 β 0.0 0.0 0.000305 δ 1.09e-07 7.55e-09 2.27e-08 ζ 0.00883 0.00111 0.00307 α′ 0.0261 0.00528 0.00323

[0121] Tail gas emission rate TP of different pollutants under the tailpipe emission module l There are different calculation formulas:

[0122]

[0123] The parameter values ​​are as follows:

[0124]

[0125] S23: Given a continuous velocity field of a road section r, the corresponding traffic flow can be estimated based on the relevant models of traffic theory: the traffic flow Q of road section r r (t) and density k r (t) and speed V r The relationship between (t) is:

[0126] Q r (t) = k r (t)×V r (t)

[0127] Since the density is unknown, we first need to infer the density value through the speed, and then calculate the traffic flow value through the above formula. In the present invention, the speed-density model is used:

[0128]

[0129] Among them, v m is the maximum speed, k c is the optimal density. Therefore, the density According to the above formula, we can get:

[0130]

[0131] Therefore, the traffic flow of all road sections in the urban road network can be estimated according to the above formula.

[0132] S24: Continuous emission rate TPl(t) and traffic flow Q for a given road section r r (t) and the road section attributes, the change of motor vehicle emissions on the road section is described by the following function, and its emission value is:

[0133]

[0134] Furthermore, the above step S3: in order to reconstruct the emission distribution of several hours of history on the road section with missing information, the idea of ​​adversarial generative learning is used to use graph convolution units and gated linear units to respectively extract the topological features of the road network and the temporal dependencies of the emission sequence of motor vehicles on the road section with complete information, so as to reconstruct the emission sequence on the road section with missing information and fill in the missing emission distribution information. Specifically, it includes the following subdivision steps S31 to S35:

[0135] S31: First, construct the vertex feature matrix of the graph structure in the Tth period (each period is 1 hour) The set of information-complete sections is The corresponding road segment in its vertex feature matrix The characteristic value of is calculated using the emission value E obtained in step 2 r (t), calculate the emission values ​​of various pollutants on the road section:

[0136]

[0137] For the road sections with missing corresponding information in the vertex feature matrix Set its eigenvalue to 0.

[0138] In order to train the network, we need to Randomly delete 20% of the vertex eigenvalues ​​and set the eigenvalues ​​of the deleted vertices to 0 to obtain a new feature matrix Among them, the vertex set of deleted eigenvalues ​​is The set of vertices that remain is Since the network task is to estimate the historical emissions of motor vehicles in 6 time intervals, the network input feature matrix is ​​constructed The target output matrix is

[0139] S32: Given the topological structure and partial feature information of a traffic road network, a spatiotemporal data interpolation model based on a generative adversarial network is designed to generate real feature values ​​(emission values) for sections with missing information in the road network. The adversarial model consists of two adversarial neural networks, namely a feature generator based on an autoencoder and a feature discriminator: the feature generator captures the data distribution and generates data as similar to the real samples as possible to confuse the discriminator; the discriminator distinguishes the real samples from the samples generated by the generator. Both networks participate in the max-min game of adversarial learning to improve the authenticity of the generated graph features. After fine-tuning the network parameters through training, the generator provides complete and real node features as output based on the partial feature data of the input.

[0140] S33: In the autoencoder-based feature generator: First, the input graph feature matrix of the network Normalized to the range [0, 1]. The generator is first composed of 3 consecutive spatiotemporal encoding blocks in series for feature learning. Each spatiotemporal encoding block is composed of a gated linear unit and a graph convolution unit in series: (1) The principle of the gated linear unit is to multiply the output of a convolution layer without nonlinearization by the output of a convolution layer with nonlinear activation. Therefore, given an input matrix The output of the gated linear unit operation *T for:

[0141]

[0142] in is the convolution kernel parameter; (2) The graph convolution unit uses the connectivity of the graph, namely the adjacency matrix, as the convolution filter for neighborhood mixing. Based on the concept of spectral graph convolution, the graph convolution unit operates It can be expressed as for the graph signal A nonlinear function whose output equal

[0143]

[0144] Among them, the Fourier basis is the eigenvector matrix of the normalized graph Laplacian matrix L. The Laplacian matrix in is the unit matrix, is a diagonal matrix with diagonal element values ​​d ii =∑ j W ij ; It is the eigenvalue diagonal matrix of matrix L, so Θ(Λ) is also a diagonal matrix, and Θ is the graph convolution kernel parameter. In the graph convolution operation, usually I=M.

[0145] Given the above gated linear unit and graph convolution unit operations, the spatiotemporal encoding block can be defined as Operation, that is, given the input matrix of the i-th spatiotemporal coding block and the kernel parameter Ξ j =[Γ j Θ j ], the output of the spatiotemporal coding block for:

[0146]

[0147] Among them, σ 1 (·), σ 2 (·) is the activation function. Therefore, the output after the above three spatiotemporal encoding blocks is:

[0148]

[0149] Among them, all activation functions are elu activation functions. After three spatiotemporal encoding blocks, the input partial graph features are converted into an N×512 intermediate feature matrix. The matrix is ​​then decoded by three spatiotemporal decoding blocks. The structure of the spatiotemporal decoding block is the same as that of the spatiotemporal encoding block, and its network structure is expressed by the formula:

[0150]

[0151] All the above activation functions are elu functions.

[0152]

[0153] The activation function σ in this step 2is the sigmoid activation function. The graph features are reconstructed through three spatiotemporal decoding blocks. In essence, the feature generator consisting of three spatiotemporal encoding blocks and three spatiotemporal decoding blocks is an autoencoder that encodes part of the input data into high-dimensional feature maps and then decodes them into a complete graph feature matrix.

[0154] S34: The purpose of the graph feature discriminator is to distinguish the graph feature matrix generated by the above feature generator and the true graph feature matrix X. The network starts with a spatiotemporal coding block, whose output is expanded into a one-dimensional matrix and then fed into three fully connected layers, that is, the spatiotemporal coding block is used to extract hidden graph features from the input data, and then the fully connected neurons in each fully connected layer work together to classify the extracted features to determine whether the input data is true. Finally, a sigmoid function is used to generate a probability between 0 and 1, indicating the true probability of the discriminator's judgment of the input feature.

[0155] A LayerNorm layer is added after each operation and each fully connected layer in the spatiotemporal encoding block to solve the problem of gradient vanishing and parameter polarization in adversarial learning. Or X, the output of the spatiotemporal encoding block is:

[0156]

[0157] Where LN(·) represents the LayerNorm layer. Expand into a one-dimensional matrix It is used as the input of the next three fully connected layers. The activation function of each layer is the elu activation function and a LayerNorm operation is added. The activation function of the last fully connected layer is the sigmoid function. The output of the final discriminator is For input The output of the discriminator is For input X, the output of the discriminator is

[0158] S35: In the training of generative adversarial networks in adversarial learning, the training objective of the feature generator is usually expressed as:

[0159]

[0160] The training objective of the feature discriminator is usually expressed as:

[0161]

[0162] In the actual training process, in order to overcome the problem of gradient disappearance, a gradient penalty term is added to the loss function of the feature discriminator, so the loss function L of the feature discriminator isD for:

[0163]

[0164] Where λ is an adjustment parameter, and its value in this patent is 10. In addition, in order to improve the quality of samples generated by the feature generator, an estimated mean square error term is added to the loss function of the generator, so the loss function of the feature generator L G for:

[0165]

[0166] The Adam optimizer is used to alternately train the feature generator and the feature discriminator until the two loss functions converge to a stable state.

[0167] The main features of the embodiments of the present invention are as follows:

[0168] 1. In order to solve the problem of spatial imbalance of traffic status information, the road sections in the road network are divided into two types of sections according to the density of the collected GPS data: sections with complete information and sections with missing information. Different ideas are used to estimate the traffic emission values ​​of the two types of sections.

[0169] 2. For sections with complete information: In order to overcome the information discontinuity and error noise in the time series, the Gaussian adaptive smoothing model is used to construct a continuous speed domain through discrete GPS points. The speed at each moment is corrected by the time series correlation between other sampling points and the speed value at that moment, thereby offsetting the error noise. After that, the traffic flow of the section is estimated based on its speed domain, and the corresponding traffic emission sequence is calculated using the micro emission model.

[0170] 3. For sections with missing information: Due to the connectivity in the urban road network, the traffic states of the two sections are interdependent, and have similar rules in traffic patterns and emission distribution. Therefore, a generative adversarial network based on spatiotemporal autoencoders is proposed, and a spatiotemporal encoding block and a spatiotemporal decoding block composed of graph convolutional layers and gated linear units are designed to construct the temporal dependency features of the traffic emission sequence of the above-mentioned sections with complete information and the spatial proximity feature space of the urban road network, thereby achieving the task of estimating the traffic emission values ​​on sections with incomplete information through the powerful feature learning ability of the generative adversarial network.

[0171] In summary, in this embodiment, the road sections are divided into two categories according to the completeness of the motor vehicle driving information obtained, and different ideas are designed for calculating the motor vehicle emissions of the two types of road sections. For the road sections with complete information, a relatively accurate continuous velocity field and flow change can be reconstructed based on the obtained data, and the motor vehicle emission value of the road section can be calculated as accurately as possible using the micro emission factor calculation model; for the road sections with missing information, the emission sequence of the road section is estimated by using the spatial topological characteristics of the road network and the temporal dependence of the motor vehicle emissions in the road sections with complete information, thereby calculating the traffic emission sequence of all road sections in the urban road network within a few hours in history.

[0172] On the other hand, the present invention also discloses a system for estimating urban mobile source pollution emissions using incomplete information, comprising the following units:

[0173] Data acquisition and processing unit, used to acquire motor vehicle driving related data in urban traffic and pre-process urban traffic network data and construct data space;

[0174] The traffic emission value calculation unit is used to divide the road sections into two categories according to the density of data obtained on the road sections: information-missing road sections with missing traffic mode information due to sparse data collection and information-complete road sections with sufficiently dense data; construct a continuous speed field and infer flow value for the information-complete road sections through the above-mentioned motor vehicle driving related data, and calculate the corresponding traffic emission value through the micro emission factor model;

[0175] The emission distribution information filling unit is used to use the graph convolution unit and the gated linear unit to respectively extract the topological characteristics of the road network and the temporal dependencies of the motor vehicle emission sequence on the information-complete section, so as to reconstruct the emission sequence on the information-missing section and fill in the missing emission distribution information.

[0176] In a third aspect, the present invention further discloses a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes the steps of the above method.

[0177] It is understandable that the system provided by the embodiment of the present invention corresponds to the method provided by the embodiment of the present invention, and the explanation, examples and beneficial effects of the relevant contents can refer to the corresponding parts in the above method.

[0178] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0179] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0180] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0181] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0182] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for estimating urban mobile source pollution emissions using incomplete information, characterized in that: The following steps are included: S10, obtaining the relevant data of motor vehicle driving in urban traffic and pre-processing the urban traffic network data and constructing the data space; According to the density of data obtained on the road section, the road sections are divided into two categories: information-missing sections with missing traffic mode information due to sparse collected data and information-complete sections with dense data; S20, constructing a continuous speed field and inferring flow values ​​for the road section with complete information through the above motor vehicle travel related data, and calculating the corresponding traffic emission values ​​through a micro emission factor model; S30, using a graph convolutional unit and a gated linear unit to respectively extract the topological features of the road network and the temporal dependency of the emission sequence of motor vehicles on the road section with complete information, thereby reconstructing the emission sequence on the road section with missing information and filling in the missing emission distribution information; The S20 specifically includes: S21: Given road segment The GPS point set {p1, p2, ..., p i ,…,p n |p i =(t i , g i , v i , r)} as input, firstly reconstruct the continuous velocity field V of the road section through the Gaussian adaptive sliding method r (t): Among them, the value of the smooth kernel function φi(t) decays as the time distance increases. Therefore, the form of the kernel function is: Among them, τ is the time width. The larger the value of τ, the greater the influence of the data point on the target calculation point; the normalization function Φ(t) is the sum of all kernel functions: Therefore, all road segments in the urban road network have their continuous velocity fields constructed according to the above formula; S22: Continuous velocity field V for a given section r r , the continuous emission rate TP is calculated using the EMIT micro emission factor model l (t); The EMIT model consists of two modules: engine emission module and tailpipe emission module; First, define the vehicle traction force P tract : The definitions of the parameters are as follows: A: rolling resistance term; B: speed correction for rolling resistance term; C: air resistance term; M: engine mass; g: gravity constant; θ: road slope; Definition, for common emission types, l is CO, HC, NO x , then the emission, the engine emission rate EO l The expression is: The parameter values ​​are as follows: The tailpipe emission rates TPl of different pollutants under the tailpipe emission module have different calculation formulas: The parameter values ​​are as follows: S23: Given a continuous velocity field of road section r, estimate the corresponding traffic flow according to the relevant model of traffic theory: the traffic flow Q of road section r r (t) and density k r (t) and speed V r The relationship between (t) is: Q r (t)=k r (t)×V r (t) Since the density is unknown, the density value is inferred from the speed, and then the traffic flow value is calculated; using the speed-density model: Among them, v m is the maximum speed, k c is the optimal density; therefore, the density V r Substitute (t) into Q r (t) can be obtained: Therefore, traffic flows are estimated for all road segments in the urban road network; S24: Continuous emission rate TP for a given road segment r l (t), traffic flow Q r (t) and the road section attributes, the motor vehicle emission value of the road section is: The S30 specifically includes: S31: First, construct the vertex feature matrix of the graph structure in the Tth period The set of information-complete sections is The corresponding road segment in its vertex feature matrix The characteristic value calculation uses the emission value E obtained in step S20 r (t), calculate the emission values ​​of various pollutants on the road section: For the road sections with missing corresponding information in the vertex feature matrix Set its characteristic value to 0; From the complete set of information Randomly delete 20% of the vertex eigenvalues ​​and set the eigenvalues ​​of the deleted vertices to 0 to obtain a new feature matrix Among them, the vertex set of deleted eigenvalues ​​is The set of vertices that remain is Since the network task is to estimate the historical emissions of motor vehicles in 6 time intervals, the network input feature matrix is ​​constructed The target output matrix is S32: Given the topological structure and partial feature information of a traffic road network, a spatiotemporal data interpolation model based on a generative adversarial network is designed to generate real feature values, i.e., emission values, for sections with missing information in the road network. The adversarial model consists of two adversarial neural networks, a feature generator based on an autoencoder and a feature discriminator. The feature generator captures the data distribution and generates data similar to real samples to confuse the discriminator. The discriminator distinguishes real samples from samples generated by the generator. Both networks participate in the maximum-minimum game of adversarial learning to improve the authenticity of the generated graph features. After fine-tuning the network parameters through training, the generator provides complete and real node features as output based on partial feature data of the input. S33: In the autoencoder-based feature generator: First, the input graph feature matrix of the network Normalized to the range of [0,1]; the generator is first composed of 3 consecutive spatiotemporal encoding blocks in series for feature learning; each spatiotemporal encoding block is composed of a gated linear unit and a graph convolution unit in series: where the gated linear unit is a convolution layer output without nonlinearization multiplied by a convolution layer output with nonlinear activation; therefore, given an input matrix Gated Linear Unit Operation Output for: in is the convolution kernel parameter; At the same time, the graph convolution unit uses the connectivity of the graph, that is, the adjacency matrix, as a convolution filter for neighborhood mixing; based on the concept of spectral graph convolution, the graph convolution unit operation For the graph signal A nonlinear function whose output equal Among them, the Fourier basis is the eigenvector matrix of the normalized graph Laplacian matrix L; the Laplacian matrix in is the unit matrix, is a diagonal matrix with diagonal element values ​​d ii =∑ j W ij ; It is the eigenvalue diagonal matrix of matrix L, so Θ(Λ) is also a diagonal matrix, Θ is the graph convolution kernel parameter; Given the above gated linear unit and graph convolution unit operations, the spatiotemporal encoding block is defined as Operation, that is, given the input matrix of the jth spatiotemporal coding block and the kernel parameter Ξ j =[Γ j Θ j ], the output of the spatiotemporal encoding block for: Among them, σ1(·), σ2(·) are activation functions; therefore, the output after the above three spatiotemporal encoding blocks is: Among them, all activation functions are elu activation functions; after three spatiotemporal encoding blocks, the input part of the graph features is converted into an N×512 intermediate feature matrix; the matrix is ​​then decoded by three spatiotemporal decoding blocks; the structure of the spatiotemporal decoding block is the same as that of the spatiotemporal encoding block, and its network structure is expressed by the formula: All the above activation functions are elu functions. The activation function σ2 in this step is a sigmoid activation function; the graph features are reconstructed through three spatiotemporal decoding blocks; S34: The graph feature discriminator is used to distinguish the graph feature matrix generated by the above feature generator and the real graph feature matrix X; the network starts with a spatiotemporal coding block, whose output is expanded into a one-dimensional matrix and then fed into three fully connected layers, that is, the spatiotemporal coding block is used to extract hidden graph features from the input data, and then the fully connected neurons in each fully connected layer work together to classify the extracted features to determine whether the input data is real; finally, a sigmoid function is used to generate a probability between 0 and 1, indicating the real probability of the discriminator's judgment of the input feature; A LayerNorm layer is added after each operation and each fully connected layer in the spatiotemporal encoding block. Or X, the output of the spatiotemporal encoding block is: Where LN(·) represents the LayerNorm layer. Expand into a one-dimensional matrix It is used as the input of the next three fully connected layers. The activation function of each layer is the elu activation function and a LayerNorm operation is added. The activation function of the last fully connected layer is the sigmoid function. The output of the final discriminator is For input The output of the discriminator is For input X, the output of the discriminator is S35: In the training of generative adversarial networks in adversarial learning, the training objective of the feature generator is expressed as: The training objective of the feature discriminator is expressed as:

2. The method for estimating urban mobile source pollution emissions using incomplete information according to claim 1 is characterized by: S10 obtains the motor vehicle driving related data in urban traffic and pre-processes the urban traffic network data and constructs the data space; according to the density of the data obtained on the road section, the road section is divided into two categories: information missing sections that lack traffic mode information due to sparse collected data and information complete sections with dense data, specifically including: S11: Urban road network data is constructed into a directed graph Given a road network, first take the set of all road forks as the edge ε connecting the road segments in the graph, then the vertex between two edges Represents a collection of N road segments, each of which has its own attributes, including the maximum lane speed value rv m , the number of lanes rn and the length of the road section r.len; the adjacency matrix W is abstracted into the adjacency correlation between vertices, and W is used to represent the connectivity between road sections: if road sections i and j are connected, then W i,j =1, otherwise W i,j =0; S12: The vehicle driving related data is the GPS trajectory of the sampled vehicle. Each GPS trajectory P contains a series of time-series GPS points, where each GPS point p contains a timestamp t and a geographic spatial coordinate g, i.e., p = (t, g). First, the instantaneous speed of the GPS point is calculated according to the change of geographic location and time: given two GPS points p1 and p2, the speed value of point p1 is: where dist(·) is a function that calculates the distance between two points. A map matching algorithm is used to match each GPS point to the corresponding road segment, so the attributes of each track point are expanded to p = (t, g, v, r). S13: Count the total number of GPS points q recorded on the road section r within one hour r , then the sampling density of this section is d r : d r =q r / r.len Through analysis, sections with sampling density less than 20 are considered as sections with missing information. The road sections with sampling density greater than or equal to 20 are regarded as complete information sections.

3. The method for estimating urban mobile source pollution emissions using incomplete information according to claim 2 is characterized by: In step S35, a gradient penalty term is added to the loss function of the feature discriminator during the training process, so the loss function L of the feature discriminator is D for: Where λ is the tuning parameter.

4. The method for estimating urban mobile source pollution emissions using incomplete information according to claim 3 is characterized by: In step S35, λ is an adjustment parameter, and its value is 10.

5. The method for estimating urban mobile source pollution emissions using incomplete information according to claim 4 is characterized by: In step S35, the estimated mean square error term is added to the loss function of the generator, so the loss function L of the feature generator is G for: The Adam optimizer is used to alternately train the feature generator and the feature discriminator until the two loss functions converge to a stable state.

6. A system for estimating urban mobile source pollution emissions using incomplete information, used to implement the method for estimating urban mobile source pollution emissions using incomplete information as described in any one of claims 1 to 5, characterized in that: It includes the following units: Data acquisition and processing unit, used to acquire motor vehicle driving related data in urban traffic and pre-process urban traffic network data and construct data space; The traffic emission value calculation unit is used to divide the road sections into two categories according to the density of data obtained on the road sections: information-missing road sections with missing traffic mode information due to sparse data collection and information-complete road sections with sufficiently dense data; construct a continuous speed field and infer flow value for the information-complete road sections through the above-mentioned motor vehicle driving related data, and calculate the corresponding traffic emission value through the micro emission factor model; The emission distribution information filling unit is used to use the graph convolution unit and the gated linear unit to respectively extract the topological characteristics of the road network and the temporal dependencies of the motor vehicle emission sequence on the information-complete section, so as to reconstruct the emission sequence on the information-missing section and fill in the missing emission distribution information.

Citation Information

Patent Citations

  • Remote sensing monitoring system for automotive exhaust emission of urban road network

    CN106845371A

  • Method and system for predicting air pollution trend by using motor vehicle GPS track

    CN111932874A

  • Road network emission list establishment method and system based on space-time velocity domain reconstruction

    CN113256966A

  • Method for estimating and predicting short-term traffic circulation state of urban road network

    WO2021109318A1

  • KR20200084750A