A dynamic assessment method for urban road network emissions at a large-scale segment scale

Abstract

A method for dynamic evaluation of urban road network emissions at a large-scale road section scale belongs to the technical field of urban road network emission prediction. In order to accurately consider the comprehensive impact of multi-scale spatial factors on carbon emissions, the present invention constructs a road network topology structure, obtains emission history data of all road sections for expansion, and then constructs a multi-scale spatial feature extraction module, including an area feature extraction layer, a road section feature extraction layer, a gated fusion layer, a time feature extraction module, and an emission prediction module. Then, based on the emission history data of all road sections, a dynamic prediction of urban road network emissions at a large-scale road section scale is performed. The present invention organically integrates data at all levels, fully explores the potential connections between data, and comprehensively and meticulously reflects the spatial distribution differences of road carbon emissions, laying a solid foundation for carbon emission analysis and prediction, improving the ability to characterize the spatial characteristics of road carbon emissions, and providing strong support for accurate prediction and the formulation of effective emission reduction strategies.

Description

Technical Field

[0001] The present invention belongs to the technical field of urban road network emission prediction, and in particular relates to a dynamic evaluation method for urban road network emissions oriented to large-scale road sections. Background Art

[0002] Against the backdrop of global efforts to address climate change, the importance and urgency of dynamic assessments of road carbon emissions are becoming increasingly prominent. Data released by the International Energy Agency (IEA) in 2023 showed that carbon emissions from the transportation sector accounted for approximately 24% of total global energy-related carbon emissions, of which road transport accounted for over 70%, and this proportion is still rising.

[0003] Road carbon emissions are influenced not only by factors like vehicle type and fuel quality, but also by the road and surrounding environment. For example, carbon emissions per unit time on congested urban roads, due to frequent vehicle starts and stops, can be 30% to 50% higher than on smoother roads. Furthermore, different road surrounding environments, such as commercial, industrial, and residential areas, significantly influence traffic flow and driving patterns, thus altering carbon emission levels.

[0004] Current assessments of road carbon emissions are largely incomplete and lack dynamism. According to a 2023 survey by relevant research institutions, over 60% of existing assessment methods consider only a single or limited set of factors, failing to accurately reflect the combined impact of multi-scale spatial factors and the interconnected influences of road network connectivity on carbon emissions. This results in a lack of accurate data support for the formulation of energy conservation and emission reduction policies, making it difficult to effectively reduce road carbon emissions.

[0005] Existing methods for studying road carbon emissions are significantly deficient in considering multi-scale spatial factors. For one thing, they lack in-depth and systematic integration of road attributes, such as road grade and peak-hour congestion, failing to fully explore the impact of these factors on carbon emissions at different scales. Furthermore, emission characteristics, which can have nonlinear and intertwined effects, are often analyzed in isolation in existing research, making it difficult to fully capture the comprehensive impact of their complex interactions on carbon emissions.

[0006] Carrying out dynamic assessments of road carbon emissions and comprehensively and meticulously considering multi-scale spatial factors have become key tasks in responding to climate change and achieving sustainable development. Summary of the Invention

[0007] The problem to be solved by the present invention is to accurately consider the comprehensive impact of multi-scale spatial factors on carbon emissions, and propose a dynamic assessment method for urban road network emissions at a large-scale road section scale.

[0008] To achieve the above object, the present invention is implemented through the following technical solutions:

[0009] A dynamic emission assessment method for urban road networks at a large-scale road section scale includes the following steps:

[0010] S1. Construct a road network topology, defining the transportation network based on roads and their connections. Collect carbon emission characteristic data for each road segment over the past t time steps and its corresponding road network topology data to obtain historical emission data for all road segments.

[0011] S2. Input the emission history data for all road sections obtained in step S1 into a 1×1 convolutional layer for data expansion to obtain the emission history data with expanded features.

[0012] S3. Construct a multi-scale spatial feature extraction module, including a region feature extraction layer, a road segment feature extraction layer, and a gated fusion layer. The expanded emission history dataset obtained in step S2 is input into the region feature extraction layer and the road segment feature extraction layer, respectively, for feature extraction to obtain region and road segment attention scores. The obtained region and road segment attention scores are then input into the gated fusion layer for feature fusion to obtain emission feature data after spatial feature extraction.

[0013] S4. Build a temporal feature extraction module. Based on the Transformer architecture, design an attention score between time steps. Perform temporal feature extraction on the emission feature data extracted from the spatial features in step S3 to obtain spatiotemporally extracted emission feature data.

[0014] S5. Construct an emission prediction module to construct a method for considering the impact of other road sections on the spatiotemporal extracted emission characteristic data obtained in step S4, and obtain emission prediction results for the urban road network at a large-scale road section scale.

[0015] Furthermore, the road network topology constructed in step S1 is recorded as G = (V', E', J'), where V' represents the set of regional road segments, covering all road segments in the study area; E' is used to represent the connection relationship between road segments; J' is the adjacency matrix J'∈R constructed from the Euclidean distance between roads. N*N , N is the number of elements in the set V of regional road sections, R represents the set of real numbers, indicating that the elements in the adjacency matrix are all real numbers, which is used in mathematics to accurately describe the properties of the matrix and the value range of the elements. i The road section in the past t Carbon emission characteristic data of time steps x i,t ∈R,1≤ i ≤N, get the emission history data X of all road sections in ∈R t×N .

[0016] Furthermore, the expression of the expansion process of step S2 is:

[0017] X = Conv (X in ) ;

[0018] Among them, Conv represents the convolution operation, X is the emission history data of all road sections after the expanded features, X∈R C×T×N , C represents the number of feature channels;

[0019] Then, the learnable spatial position embedding matrix is ​​used to learn the longitude and latitude of the road sections within the range. The emission history data of all road sections after position embedding is: X ', the expression of the embedding process is:

[0020] X '=F(X,Rarea);

[0021] Where F(·) is a 1×1 convolutional layer, Rarea is the adjacency matrix represented by longitude and latitude coordinates, and Rarea∈R N×N .

[0022] Furthermore, the specific implementation method of step S3 includes the following steps:

[0023] S3.1. Constructing the region feature extraction layer: First, construct the region query subspace , regional key subspace and region-valued subspace , the expression is:

[0024] ;

[0025] ;

[0026] ;

[0027] in, 、 、 They are regional query subspaces , regional key subspace and region-valued subspace The learnable weight matrix of

[0028] Calculate the i Regional attention score between road segments , the expression is:

[0029] ;

[0030] in, Matrix form representing past time; It is i The corresponding value of each road segment in the regional query subspace, d k is the preset parameter, A i is the first scaling parameter optimized through training, B i is the second scaling parameter optimized through training;

[0031] Then for the i The regional attention scores between the road sections are weighted to obtain the adjusted i Regional attention score of road sections The expression is:

[0032] ;

[0033] in, It is Hadamard. is a learnable tensor, softmax Function that converts a real vector into a probability distribution;

[0034] Introducing Re into the feature transformation layer LU The function performs preliminary nonlinear transformation, and the calculation expression is:

[0035] ;

[0036] in, is the feature transformation matrix;

[0037] Then, the multi-layer perceptron generation function is used to process the regional features after the preliminary transformation to obtain the processed regional features The expression is:

[0038] ;

[0039] in, is the multilayer perceptron generating function;

[0040] Then for the i The regional attention scores of each road section are fused with the processed regional features to obtain the improved i The regional global characteristics of a road section are expressed as:

[0041] ;

[0042] in, represents the feature concatenation operation, is the projection matrix;

[0043] S3.2. Construct the road segment feature extraction layer, including constructing the i road segment query subspace , road segment key subspace and segment value subspace , the expression is:

[0044] ;

[0045] ;

[0046] ;

[0047] in, 、 、 They are road segment query subspaces , road segment key subspace and segment value subspace The learnable weight matrix of

[0048] The road segment query subspace , road segment key subspace Perform dot product operation to calculate the i Road segment attention score , the expression is:

[0049] ;

[0050] S3.3. Constructing the gated fusion layer.

[0051] First, dimension adaptive weight calculation, road section feature dimension adaptive weight The calculation formula is

[0052] ;

[0053] in, Function for getting the dimensions of the input feature vector;

[0054] Adaptive weight of area feature dimension The calculation formula is:

[0055] ;

[0056] The expression for setting the gate signal is:

[0057] ;

[0058] in, Indicates the i The gate signal of each road section is The activation function is calculated and has a value range of [0,1], which is used to perform weighted selection processing on the input data; They are functions for linear transformation of road section features and area features, respectively, and map the multi-dimensional feature vector into a one-dimensional vector through a specific weight matrix;

[0059] Then, feature fusion is performed based on the gated signal to obtain the first i Emission characteristic data of road sections , the expression is:

[0060] .

[0061] Furthermore, the specific implementation method of step S4 includes the following steps:

[0062] S4.1. Designing attention scores between time steps based on the Transformer architecture , based on each x i,t The expression for data containing T time steps is:

[0063] ;

[0064] in, is represented as a learnable weight matrix for the temporal training query subspace, is represented as a learnable weight matrix in the time-key subspace, represents the corresponding weight vector;

[0065] Then the attention scores between time steps and the value subspace between time steps are Fusion, add the residual connection layer normalization operation LN after feature extraction, and obtain the first i Emission characteristic data of road sections , the calculation expression is:

[0066] .

[0067] Furthermore, the specific implementation method of step S5 includes the following steps:

[0068] S5.1. Considering other road segment features for the first i The emission characteristic data of each road section is affected, and the road section feature fusion layer is constructed. i The emission characteristic data of each road section is fused with its own characteristics and the characteristics of other road sections at the same time to obtain the fused feature vector The expression is:

[0069] ;

[0070] in, For other road features, represents the feature concatenation operation, For road sections j To the road section i The Euclidean distance of j Relative to the i Other sections of the road;

[0071] S5.2. Convert the two-dimensional input feature vector Map to high-dimensional space and construct high-dimensional feature matrix , the expression is:

[0072] ;

[0073] in, is the kernel function;

[0074] Then, the high-dimensional feature matrix is ​​added to the regularization term to obtain the urban road network emission prediction result for large-scale road sections, which is expressed as:

[0075] ;

[0076] in, express T +1 time i Emission forecast for urban road network for each road section, It's all from The matrix form of W is the sample weight matrix, I is the identity matrix.

[0077] Beneficial effects of the present invention:

[0078] The method for dynamic emission assessment of urban road networks at a large-scale road section scale described in the present invention breaks through the limitations of traditional research methods and is no longer limited to single-dimensional considerations. Instead, it comprehensively analyzes the rich information at the levels of road basic attributes, surrounding environment, and traffic flow characteristics. In terms of road basic attributes, it covers factors such as roads themselves, which directly affect the energy consumption and emissions of vehicles during driving. Surrounding environmental factors are conducive to the fact that different regional functions will lead to significant differences in traffic demand and travel patterns, which in turn have different degrees of effect on carbon emissions. On this basis, the present invention uses scientific and reasonable data fusion and feature engineering technologies to carefully construct a multi-scale spatial feature data set. Through a rigorous data processing process, the data at all levels are organically integrated, and the potential connections between the data are fully explored, reflecting the spatial distribution differences of road carbon emissions and the complex influencing factors behind them in an extremely comprehensive and detailed manner, laying a solid foundation for subsequent carbon emission analysis and prediction, and is one of the core technologies of the present invention.

[0079] The present invention describes a method for dynamic emission assessment of urban road networks at a large-scale road section scale, which uses a variety of advanced time series analysis methods to conduct in-depth trend and cycle analysis of carbon emission time series. During the trend analysis process, the overall trend of carbon emissions over time can be accurately identified, whether it is rising, falling or fluctuating, it can be clearly presented. Cycle analysis helps us discover the hidden periodic patterns in carbon emission data, such as whether there are seasonal fluctuations, differences between weekdays and weekends, etc. At the same time, this technology also has the ability to handle the impact of special events. Through effective processing technology, the effect of these special events on carbon emissions can be accurately quantified, so as to more comprehensively grasp the temporal variation patterns of carbon emissions. This technology provides crucial time dimension information for the subsequent prediction model, and is a key link in improving prediction accuracy, making the prediction results more in line with the actual carbon emissions situation.

[0080] The dynamic emission assessment method for urban road networks at a large-scale road section scale described in this invention requires the selected model to possess strong feature learning and pattern recognition capabilities, taking into account the complexity and multidimensionality of carbon emission data. After the model is constructed, the model parameters are rationally adjusted to continuously explore the model's potential and better adapt it to the characteristics of carbon emission data. At the same time, cross-validation technology is used to divide and train the dataset multiple times, effectively avoiding model overfitting and improving the model's generalization ability. Using a variety of evaluation indicators, such as mean square error, mean absolute error, and coefficient of determination, the model performance is comprehensively evaluated from different perspectives to ensure the accuracy and reliability of the model.

[0081] The dynamic emission assessment method for urban road networks at a large-scale, segment-scale, and described in this paper ensures that the prediction model fully utilizes extracted spatial and temporal features to achieve highly accurate carbon emission forecasts. This not only provides a powerful tool for the study of road carbon emissions but also provides a scientific basis for relevant departments to formulate energy-saving and emission reduction policies and plan transportation infrastructure. By systematically integrating multi-source information, the ability to characterize the spatial characteristics of road carbon emissions is enhanced, providing strong support for accurate predictions and the development of effective emission reduction strategies, and is of great significance for promoting the green transformation of the transportation industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] Figure 1 This is a flow chart of a large-scale road section-scale urban road network emission dynamic assessment method described in the present invention. DETAILED DESCRIPTION

[0083] In order to make the objectives, technical solutions, and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only intended to explain the present invention and are not intended to limit the present invention. That is, the specific embodiments described herein are only some embodiments of the present invention, not all embodiments. Generally, the components of the specific embodiments of the present invention described and illustrated in the drawings herein can be arranged and designed in various different configurations, and the present invention can also have other embodiments.

[0084] Therefore, the following detailed description of the specific embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention as claimed, but is merely representative of selected specific embodiments of the present invention. All other specific embodiments obtained by those skilled in the art based on the specific embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0085] In order to further understand the content, features and effects of the present invention, the following specific embodiments are given as examples, and the attached Figure 1 The detailed instructions are as follows:

[0086] Example 1:

[0087] A dynamic emission assessment method for urban road networks at a large-scale road section scale includes the following steps:

[0088] S1. Construct a road network topology, defining the transportation network based on roads and their connections. Collect carbon emission characteristic data for each road segment over the past t time steps and its corresponding road network topology data to obtain historical emission data for all road segments.

[0089] Furthermore, the road network topology constructed in step S1 is recorded as G = (V', E', J'), where V' represents the set of regional road segments, covering all road segments in the study area; E' is used to represent the connection relationship between road segments; J' is the adjacency matrix J'∈R constructed from the Euclidean distance between roads. N*N , N is the number of elements in the set V of regional road segments, R represents a real number set, and the first i The road section in the past t Carbon emission characteristic data of time steps x i,t ∈R,1≤ i ≤N, get the emission history data X of all road sections in ∈R t×N .

[0090] Furthermore, V' covers all road segments in the study area. E' is used to represent the connection relationship between road segments. If two road segments are connected to each other, there is a corresponding connection relationship record in E'.

[0091] S2. Input the emission history data for all road sections obtained in step S1 into a 1×1 convolutional layer for data expansion to obtain the emission history data with expanded features.

[0092] Furthermore, the expression of the expansion process of step S2 is:

[0093] X = Conv (X in )

[0094] Among them, Conv represents the convolution operation, X is the emission history data of all road sections after the expanded features, X∈R C×T×N , C represents the number of feature channels;

[0095] Then, the learnable spatial position embedding matrix is ​​used to learn the longitude and latitude of the road sections within the range. The emission history data of all road sections after position embedding is: X ', the expression of the embedding process is:

[0096] X =F(X, Rarea)

[0097] Where F(·) is a 1×1 convolutional layer, Rarea is the adjacency matrix represented by longitude and latitude coordinates, and Rarea∈R N×N .

[0098] Furthermore, the convolutional layer plays a crucial role in the entire data processing system. Its core function is to cleverly and accurately incorporate dynamically changing positional information into the input data. During the actual operation, it meticulously calculates and processes each component of the input data according to pre-defined convolution rules. By performing convolution operations with specific convolution kernels, it integrates the positional information into the data's feature representation in a way that can be effectively utilized by subsequent models. The positionally embedded data X' is then fed into the regional multi-head self-attention unit for spatial feature extraction.

[0099] S3. Construct a multi-scale spatial feature extraction module, including a region feature extraction layer, a road segment feature extraction layer, and a gated fusion layer. The expanded emission history dataset obtained in step S2 is input into the region feature extraction layer and the road segment feature extraction layer, respectively, for feature extraction to obtain region and road segment attention scores. The obtained region and road segment attention scores are then input into the gated fusion layer for feature fusion to obtain emission feature data after spatial feature extraction.

[0100] For the calculation of the area feature extraction layer of road section i, it means taking road i as the center and taking all connected road sections within a certain range as an area. The certain range is taken as a circular range with a radius of 5 kilometers as an example. The road sections within the range are circled through E' in the data and the corresponding connection relationship records.

[0101] Furthermore, the specific implementation method of step S3 includes the following steps:

[0102] S3.1. Constructing the region feature extraction layer: First, construct the region query subspace , regional key subspace and region-valued subspace , the expression is:

[0103] ;

[0104] ;

[0105] ;

[0106] in, 、 、 They are regional query subspaces , regional key subspace and region-valued subspace The learnable weight matrix of

[0107] Calculate the i Regional attention score between road segments , the expression is:

[0108] ;

[0109] in, Matrix form representing past time; It is i The corresponding value of each road segment in the regional query subspace, d k is the preset parameter, A i is the first scaling parameter optimized through training, B i is the second scaling parameter optimized through training;

[0110] In traditional traffic flow prediction and assessment, the original formula maps the road segment attention score to the range [0, 1] using a softmax function and then multiplies it with the linearly transformed features to calculate the global characteristics of the road segment. However, unlike traffic volume or speed, emissions are a continuously changing process and can be significantly affected by surrounding road characteristics, resulting in nonlinear increases. This is because frequent acceleration and deceleration typically result in higher emissions than those generated by constant speed driving. The characteristics of the adjacent road network have complex nonlinear correlations. For example, while congestion guidelines may linearly decrease vehicle speed, at certain critical points, frequent vehicle stops and starts can cause emissions to vary exponentially.

[0111] A sophisticated weighting mechanism is introduced for calculating attention scores. By element-by-element multiplication of the softmax-processed attention score with the nonlinear transformation of the score using a learnable tensor, this innovation changes the previous simplistic approach. It fully considers the differential impact of complex factors on carbon emissions across different road sections, transcending the limitations of traditional attention calculations. This significantly enhances the accuracy of road section importance assessments, enabling more precise focus on key sections and avoiding the omission of critical information due to irrational attention allocation. The adjusted regional attention score for road section i is:

[0112] Then for the i The regional attention scores between the road sections are weighted to obtain the adjusted i Regional attention score of road sections The expression is:

[0113] ;

[0114] in, It is Hadamard. is a learnable tensor, softmax Function that converts a real vector into a probability distribution;

[0115] At the feature transformation level, single linear transformations previously struggled to extract complex carbon emission-related features. This improvement introduces nonlinearity through the ReLU function and increases feature diversity through a generating function. This change takes into account the high complexity and nonlinearity of carbon emission characteristics, transcending the simplistic approach of traditional linear transformations. It enhances the model's ability to learn complex features, enabling it to understand carbon emission data from a richer perspective. This will enable more comprehensive and in-depth feature extraction and avoid evaluation bias caused by insufficient feature extraction.

[0116] Introducing Re into the feature transformation layer LU The function performs preliminary nonlinear transformation, and the calculation expression is:

[0117] ;

[0118] in, is the feature transformation matrix;

[0119] Then, the multi-layer perceptron generation function is used to process the regional features after the preliminary transformation to obtain the processed regional features The expression is:

[0120] ;

[0121] in, It is a multi-layer perceptron generating function; it generates a feature vector with the same dimension as the feature after the initial transformation and adds it to the feature after the initial transformation, breaking through the simple thinking of traditional linear transformation and enhancing the model's ability to learn complex features.

[0122] Next, when it comes to information fusion, traditional methods often focus solely on single-dimensional features, ignoring the synergistic effects between features from different dimensions. This improvement utilizes a feature concatenation method, concatenating features with their projected transformations before multiplying them with the adjusted attention score. This fully considers the combined impact of multidimensional features on carbon emissions, overcomes the drawbacks of isolated information processing, and enhances the model's ability to integrate multidimensional information. This will achieve a more comprehensive feature representation and avoid biased assessments caused by insufficient information fusion.

[0123] Then for the i The regional attention scores of each road section are fused with the processed regional features to obtain the improved i The regional global characteristics of a road section are expressed as:

[0124] ;

[0125] in, represents the feature concatenation operation, is the projection matrix;

[0126] S3.2. Construct the road segment feature extraction layer, including constructing the i road segment query subspace , road segment key subspace and segment value subspace , the expression is:

[0127] ;

[0128] ;

[0129] ;

[0130] in, 、 、 They are road segment query subspaces , road segment key subspace and segment value subspace The learnable weight matrix of

[0131] Secondly, the correlation scores between sensor nodes are calculated. Since the core goal of road-level feature extraction is to search for roads with similar characteristics globally, and then perform feature extraction, a sparse self-attention mechanism is used to extract spatial features in order to effectively reduce computational complexity without sacrificing accuracy.

[0132] The road segment query subspace , road segment key subspace Perform dot product operation to calculate the i Road segment attention score , the expression is:

[0133] ;

[0134] S3.3. Constructing the gated fusion layer.

[0135] The gated fusion layer plays a vital role in the entire model architecture. Its main function is to organically integrate dynamic features at multiple spatial scales and static spatial features with the help of the gating mechanism.

[0136] In existing prediction and evaluation model architectures, the original gated fusion layer fails to fully account for the dimensional differences between the segment and patch feature extraction layers when fusing multi-scale features. Segment features focus more on local details, while patch features cover more macroscopic information. This dimensionality mismatch makes it difficult to accurately weigh the importance of different features during gating calculations, resulting in inadequate information fusion.

[0137] This improvement overcomes this challenge by introducing a dimensionality-adaptive adjustment mechanism to improve the calculation of gating signals. This mechanism dynamically adjusts feature weights based on the dimensionality of road section and area features, avoiding information imbalance caused by dimensionality differences and enhancing the model's ability to integrate features at multiple spatial scales.

[0138] First, dimension adaptive weight calculation, road section feature dimension adaptive weight The calculation formula is

[0139] ;

[0140] in, Function for getting the dimensions of the input feature vector;

[0141] Adaptive weight of area feature dimension The calculation formula is:

[0142] ;

[0143] The expression for setting the gate signal is:

[0144] ;

[0145] in, Indicates the i The gate signal of each road section is The activation function is calculated and has a value range of [0,1], which is used to perform weighted selection processing on the input data; They are functions for linear transformation of road section features and area features, respectively, and map the multi-dimensional feature vector into a one-dimensional vector through a specific weight matrix;

[0146] Then, feature fusion is performed based on the gated signal to obtain the first i Emission characteristic data of road sections , the expression is:

[0147] .

[0148] S4. Build a temporal feature extraction module. Based on the Transformer architecture, design an attention score between time steps. Perform temporal feature extraction on the emission feature data extracted from the spatial features in step S3 to obtain spatiotemporally extracted emission feature data.

[0149] The extracted spatial features are fused with temporal features. Time-dependent features, encompassing both spatiotemporal and temporal dependencies, need to be extracted from a time series consisting of traffic data at multiple time points. Given that traditional extraction methods tend to lose relative position information, a traffic temporal feature extraction module is proposed. This module selectively extracts features based on traffic characteristics and relative position information in the time series.

[0150] Each road segment input sequence has a clear temporal order. In terms of temporal dependency, the input sequence exhibits a clear and orderly temporal order, a property that plays a crucial role in accurate traffic data prediction. However, the original Transformer architecture has limitations in learning positional information from time series, making it difficult to effectively capture this critical information. To address this challenge, the process of extracting spatiotemporal features closely incorporates the unique properties of the traffic system and incorporates relative position information into the sequence. Furthermore, a parameter that is continuously optimized through training is specifically introduced to represent relative position during the subsequent calculation of the attention score.

[0151] Each x i,t Considering the high temporal fluidity of traffic data, which means that future time information does not affect the trend of data from previous time steps, a weight matrix is ​​generated for each of the T different relative position relationships. A relative position weight is added to each road i.

[0152] Furthermore, the specific implementation method of step S4 includes the following steps:

[0153] S4.1. Designing attention scores between time steps based on the Transformer architecture , based on each x i,t The expression for data containing T time steps is:

[0154] ;

[0155] in, is represented as a learnable weight matrix for the temporal training query subspace, is represented as a learnable weight matrix in the time-key subspace, represents the corresponding weight vector;

[0156] In the process of traffic time feature extraction, after a series of preliminary calculation and processing steps, it enters the key time series dependency mining stage. At this time, the attention scores between all time steps are It has been calculated that these scores reflect the closeness of the correlation between data at different time steps. At the same time, the corresponding value subspace Fusion. Next, the attention score is multiplied with the value subspace. This multiplication operation is not a simple numerical calculation, but has profound significance. In essence, it is achieved by right The information in the dataset is weighted and integrated to mine and determine the dependencies between time series. After feature extraction, a layer normalization operation LN with residual connection is added;

[0157] Then the attention scores between time steps and the value subspace between time steps are Fusion, add the residual connection layer normalization operation LN after feature extraction, and obtain the first i Emission characteristic data of road sections , the calculation expression is:

[0158] .

[0159] This operation effectively extracts the dependencies between data points in various dimensions. These dependencies are crucial for understanding how traffic data changes over time and predicting future traffic conditions. This allows the model to more accurately capture and utilize the inherent patterns in time series, thereby improving the accuracy and reliability of predictions.

[0160] S5. Construct an emission prediction module, construct a method to analyze the impact of other road sections on the emission characteristic data obtained after spatiotemporal extraction in step S4, and obtain emission prediction results for urban road networks at a large-scale road section scale.

[0161] First, a segment feature fusion layer is constructed. The carbon emissions value of segment i at time t is fused with its own characteristics and the characteristics of other segments at the same time. Current methods do not fully consider the impact of segment characteristics and other segment characteristics. This step shifts the focus from isolated carbon emissions to integrating multiple aspects of information, overcoming the issue of incomplete information utilization and providing a richer feature foundation for subsequent predictions.

[0162] Furthermore, the specific implementation method of step S5 includes the following steps:

[0163] S5.1. Considering other road segment features for the first i The emission characteristic data of each road section is affected, and the road section feature fusion layer is constructed. i The emission characteristic data of each road section is fused with its own characteristics and the characteristics of other road sections at the same time to obtain the fused feature vector The expression is:

[0164] ;

[0165] in, For other road features, represents the feature concatenation operation, road section j To the road section i The Euclidean distance of j Relative to the i Other sections of the road;

[0166] S5.2. Convert the two-dimensional input feature vector Map to high-dimensional space and construct high-dimensional feature matrix , the expression is:

[0167] ;

[0168] in, is the kernel function;

[0169] Then, the high-dimensional feature matrix is ​​added to the regularization term to obtain the urban road network emission prediction result for large-scale road sections, which is expressed as:

[0170] ;

[0171] in, express T +1 time i Emission forecast for urban road network for each road section, It's all from The matrix form of W is the sample weight matrix, I is the identity matrix.

[0172] Furthermore, W is used to adjust the importance of different samples in the calculation, taking into account the different impacts of different samples on the prediction results, and avoiding the problem of all samples having the same impact weight on the results. is in inverse matrix form. By adding regularization term , prevent the model from overfitting and ensure that the model still has good generalization ability under complex calculations.

[0173] for Forecast value, we need to talk about the forecast value Substitute the input parameters into the above steps and repeat the calculation.

[0174] It should be noted that relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus comprising the element.

[0175] Although the present application has been described above with reference to specific embodiments, various modifications may be made thereto and components may be substituted with equivalents without departing from the scope of the present application. In particular, as long as there are no structural conflicts, the various features of the embodiments disclosed herein may be combined with each other in any manner, and the omission of an exhaustive description of these combinations in this specification is solely for the sake of space and resource conservation. Therefore, the present application is not limited to the specific embodiments disclosed herein, but includes all technical solutions within the scope of the claims.

Claims

1. A dynamic emission assessment method for urban road networks at a large-scale road section scale, characterized by: The steps include: S1. Construct a road network topology, define the transportation network based on the roads and their connections, collect carbon emission characteristic data of road sections over the past t time steps and their corresponding road network topology data, and obtain emission history data for all road sections; S2. Input the emission history data of all road sections obtained in step S1 into a 1×1 convolutional layer for data expansion to obtain the emission history data with expanded features; S3. Construct a multi-scale spatial feature extraction module, including a district feature extraction layer, a road section feature extraction layer, and a gated fusion layer. Input the expanded emission history dataset obtained in step S2 into the district feature extraction layer and the road section feature extraction layer, respectively, to extract the district attention score and the road section attention score. The obtained district attention score and road section attention score are then input into the gated fusion layer for feature fusion to obtain emission feature data after spatial feature extraction. The specific implementation method of step S3 includes the following steps: S3.

1. Constructing the region feature extraction layer: First, construct the region query subspace Q area , regional key subspace K area and the region value subspace V area , the expression is: Q area =X'W Q area K area =X'W K area V area =X'W V area Among them, W Q area 、W K area 、W V area They are the regional query subspace Q area , regional key subspace K area and the region value subspace V area The learnable weight matrix, X' is the emission history data of all road sections after position embedding; Calculate the regional attention score between the i-th road segment The expression is: Among them, x i The matrix form represents the carbon emission characteristic data of the i-th road section in the past time; is the value corresponding to the i-th road segment in the regional query subspace, d k is the preset parameter, A i is the first scaling parameter optimized through training, B i is the second scaling parameter optimized through training; Then, a weighted mechanism is introduced for the regional attention score between the i-th road sections to obtain the adjusted regional attention score of the i-th road section. The expression is: Where ⊙ is the Hadamard product, T arae It is a learnable tensor, and softmax is a function that converts a real vector into a probability distribution; The ReLU function is introduced into the feature transformation layer for preliminary nonlinear transformation. The calculation expression is: Among them, U area is the feature transformation matrix, is the regional feature after preliminary transformation; Then, the multi-layer perceptron generation function is used to process the regional features after the preliminary transformation to obtain the processed regional features The expression is: Among them, G area is the multilayer perceptron generating function; Then, the regional attention score of the i-th road section and the processed regional features are fused to obtain the improved regional global features of the i-th road section, which is expressed as: in, represents the feature concatenation operation, P area is the projection matrix, is the regional global feature of the i-th road section after improvement; S3.

2. Construct the road segment feature extraction layer, including constructing the road segment query subspace Q of the i-th road segment road,i , road segment key subspace K road,i and the segment value subspace V road,i , the expression is: Q road,i =X'W Q road,i K road,i =X'W K road,i V road,i =X'W V road,i Among them, W Q road,i 、W K road,i 、W V road,i They are the road segment query subspace Q road,i , road segment key subspace K road,i and the segment value subspace V road,i The learnable weight matrix of The road segment query subspace Q road,i , road segment key subspace K road,i Perform dot product operation to calculate the road section attention score of the i-th road section The expression is: S3.

3. Construct the gated fusion layer; First, dimension adaptive weight calculation, road section feature dimension adaptive weight The calculation formula is: Among them, Dim(·) is a function used to obtain the dimension of the input feature vector; Adaptive weight of area feature dimension The calculation formula is: The expression for setting the gate signal is: Among them, g i represents the gating signal of the i-th road section, which is calculated by the sigmoid activation function and has a value range of [0,1], and is used to perform weighted selection processing on the input data; road (·), F area (·) are functions that perform linear transformations on the road segment features and the area features, respectively, and map the multidimensional feature vector into a one-dimensional vector through a specific weight matrix; Then, feature fusion is performed based on the gated signal to obtain the emission feature data X of the i-th road section after spatial feature extraction. S,i , the expression is: S4. Construct a temporal feature extraction module. Based on the Transformer architecture, design an attention score between time steps. Perform temporal feature extraction on the emission feature data obtained after spatial feature extraction in step S3 to obtain spatiotemporal extracted emission feature data. S5. Construct an emission prediction module, construct a method to analyze the impact of other road sections on the emission characteristic data obtained after spatiotemporal extraction in step S4, and obtain emission prediction results for urban road networks at a large-scale road section scale.

2. The method for dynamic emission assessment of urban road networks at a large-scale road section scale according to claim 1 is characterized in that: The road network topology constructed in step S1 is recorded as G = (V', E', J'), where V' represents the set of regional road segments, covering all road segments in the study area; E' is used to represent the connection relationship between road segments; J' is the adjacency matrix J'∈R constructed from the Euclidean distance between roads. N*N , N is the number of elements in the set V' of regional road sections, R represents the set of real numbers, indicating that the elements in the adjacency matrix are all real numbers, which is used in mathematics to accurately describe the properties of the matrix and the value range of the elements. The carbon emission characteristic data x collected for the i-th road section in the past t time steps i,t ∈R, 1≤i≤N, get the emission history data X of all road sections in ∈R T’×N .

3. The method for dynamic emission assessment of urban road networks at a large-scale road section scale according to claim 2 is characterized in that: The expression of the expansion process of step S2 is: X=Conv(X in ) Among them, Conv represents the convolution operation, X is the emission history data of all road sections after the expanded features, X∈R C×T’×N , C represents the number of feature channels; Then, the learnable spatial position embedding matrix is ​​used to learn the longitude and latitude of the road sections within the range. The emission history data of all road sections after position embedding is X'. The expression of the embedding process is: X'=F(X,Rarea) Where F(·) is a 1×1 convolutional layer, Rarea is the adjacency matrix represented by longitude and latitude coordinates, and Rarea∈R N×N .

4. The method for dynamic emission assessment of urban road networks at a large-scale road section scale according to claim 3 is characterized in that: The specific implementation method of step S4 includes the following steps: S4.

1. Designing attention scores between time steps based on the Transformer architecture Based on each x i,t The expression for data containing t time steps is: in, is represented as a learnable weight matrix for the temporal training query subspace, It is represented as a learnable weight matrix in the time-key subspace, W PPP represents the corresponding weight vector; Then the attention scores between time steps and the value subspace V between time steps are time Fusion, adding the residual connection layer normalization operation LN after feature extraction, to obtain the emission feature data of the i-th road section after spatiotemporal extraction The calculation expression is:

5. The method for dynamic emission assessment of urban road networks at a large-scale road section scale according to claim 4 is characterized in that: The specific implementation method of step S5 includes the following steps: S5.

1. Considering the impact of other road section features on the emission feature data of the i-th road section after spatiotemporal extraction, a road section feature fusion layer is constructed to fuse the emission feature data of the i-th road section after spatiotemporal extraction with its own features and the features of other road sections at the same time to obtain the fused feature vector F i,t The expression is: Among them, Z j≠i,t For other road segment features, Concat(·) represents feature concatenation operation, e i,j is the Euclidean distance from segment j to segment i, and j is the other segment relative to the i-th segment; S5.

2. Use the kernel function to transform the two-dimensional input feature vector [x i,t ,F i,t ] is mapped to a high-dimensional space and a high-dimensional feature matrix θ is constructed i,t ; Then, the high-dimensional feature matrix is ​​added to the regularization term to obtain the urban road network emission prediction result for large-scale road sections, which is expressed as: y i,t+1 =((θ T Wθ+λI) -1 i T W) T i i,t Among them, y i,t+1 represents the emission prediction value of the urban road network of the i-th road segment at time t+1, and θ is the total θ i,t From θ i,1 ,θ i,2 ...θ i,T’ In matrix form, W is the sample weight matrix and I is the identity matrix.

Images