A method for quickly depicting heavy-duty diesel exhaust pollution fields in a city area based on dynamic point distribution

Abstract

The present application relates to the field of urban road traffic pollution assessment, and particularly relates to a city heavy-duty vehicle emission pollution field rapid characterization method based on dynamic point distribution, comprising the following steps: S1. constructing a joint input data of road emission space characteristics and meteorological conditions; S2. extracting multi-scale characteristics of the emission space based on a residual coding structure; S3. adaptively modulating the multi-scale space characteristics according to the meteorological conditions; S4. using a deep learning model to establish a mapping relationship between the high-dimensional space characteristics after the conditional modulation and an optimal receptor point distribution strategy; and the like. Compared with the prior art, the present application significantly reduces the number of pollution concentration receptor points under the premise of ensuring the spatial resolution of the pollution concentration field, improves the overall efficiency of road traffic pollution emission simulation and pollution hotspot identification, and provides an efficient and reliable technical means for urban traffic environmental impact assessment, pollution prevention and control decision-making and fine traffic management.

Description

Technical Field

[0001] This invention relates to the field of urban road traffic pollution assessment, specifically to a method for rapidly characterizing urban heavy diesel emission pollution fields based on dynamic point layout. Background Technology

[0002] With the acceleration of urbanization and the continuous growth of motor vehicle ownership, road traffic emissions have become one of the major sources of urban air pollution. Heavy-duty diesel trucks, in particular, due to their large size, long mileage, and high emission intensity, contribute significantly to urban NOx emissions. X Heavy-duty diesel trucks account for a significant proportion of emissions and are a major factor contributing to urban air pollution and localized pollution hotspots. Therefore, a detailed characterization of the emission features of heavy-duty diesel trucks and the rapid identification of pollution hotspots they cause are of great significance for urban traffic environment assessment and pollution prevention and control decisions.

[0003] In pollution diffusion simulation, the spatial layout of pollution concentration receptor points has a decisive impact on the accuracy of pollution field characterization and computational efficiency. In existing technologies, receptor placement typically employs a regular grid layout or an empirically based fixed-interval layout. To ensure no pollution hotspots are missed, a large number of receptor points are often deployed across the city, resulting in enormous computational load and long computation time, making it difficult to meet the needs of multi-scenario and dynamic analysis. Furthermore, fixed-point placement lacks adaptability to spatial heterogeneity of emissions and changes in meteorological conditions, easily leading to wasted receptor resources or insufficient placement in key areas, affecting the accuracy of pollution hotspot identification.

[0004] Therefore, there is an urgent need for a rapid identification method for urban traffic pollution hotspots that can comprehensively consider the spatial characteristics of road emissions and changes in meteorological conditions, dynamically optimize receptor deployment strategies, and significantly reduce computational costs while ensuring the accuracy of pollution field characterization, so as to overcome the shortcomings of existing technologies in terms of efficiency, adaptability and practicality. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a method for rapidly characterizing the pollution field of heavy-duty diesel emissions in urban areas based on dynamic point layout. This method can quickly and accurately obtain the pollution field caused by heavy-duty diesel truck emissions in urban areas based on the spatial characteristics of road emissions and meteorological conditions, and realize the rapid identification of pollution hotspots.

[0006] The technical solution adopted by this invention to solve its technical problem is:

[0007] A rapid characterization method for urban heavy diesel emission pollution fields based on dynamic point layout includes the following steps:

[0008] S1. Construct joint input data of road emission spatial characteristics and meteorological conditions;

[0009] S2. Extracting multi-scale features of emission space based on residual coding structure;

[0010] S3. Adaptive modulation of multi-scale spatial characteristics based on meteorological conditions;

[0011] S4. Establish the mapping relationship between the conditionally modulated high-dimensional spatial features and the optimal receptor placement strategy;

[0012] S5. Based on the mapping relationship described in step S4, generate the optimal receptor placement strategy for the city under a specific combination of emission characteristics and meteorological conditions;

[0013] S6. Based on the optimal receptor placement strategy, the pollution diffusion model is invoked and combined with spatial interpolation methods to quickly obtain the pollution field causing emissions from heavy-duty diesel trucks in the city, thereby achieving rapid identification of pollution hotspots.

[0014] Step S1 uses urban road-level emission data and urban meteorological station meteorological data as inputs to construct joint input data of road emission spatial characteristics and meteorological condition characteristics.

[0015] The spatial characteristics of road emissions are represented using grid data with a fixed spatial resolution, including emission intensity characteristics, directional characteristics, and spatial heterogeneity characteristics. See below:

[0016] Emission intensity characteristics are used to characterize the pollutant emission levels of road segments within a grid cell, specifically the weighted or cumulative result of emissions from each road segment falling within that grid cell;

[0017] Directional characteristics are quantified using Shannon entropy to characterize the distribution characteristics of pollutant emissions in different spatial directions. By statistically analyzing the distribution of emissions in multiple directional intervals and calculating its dispersion, the concentration or dispersion of emission directional distribution can be reflected.

[0018] Spatial heterogeneity is measured by a weighted Gini coefficient, which characterizes the degree of unevenness in the spatial distribution of pollutant emissions within and around a grid cell. It is obtained by weighting the differences in emissions at different spatial locations to depict the spatial clustering or dispersion of emissions.

[0019] Meteorological condition characteristics are obtained through feature screening of candidate meteorological variables. The screening process is conducted independently for stable atmospheric conditions and unstable atmospheric conditions, and the screening steps include at least the following:

[0020] Multicollinearity analysis was performed on candidate meteorological variables, and meteorological variables with collinearity exceeding a preset threshold were eliminated through iterative methods until the remaining meteorological variables met the collinearity constraint conditions.

[0021] Based on the pollutant concentration output, the influence of each meteorological variable on the change of pollutant concentration is assessed, and the meteorological variables that have a dominant influence on the pollution diffusion process are selected and retained to form a combination of meteorological features for model input.

[0022] In one embodiment of the present invention, the meteorological feature combination retained under stable meteorological conditions is: wind direction, wind speed, and Moning-Obukhov length; the meteorological feature combination retained under unstable meteorological conditions is: wind direction, sensible heat flux, wind speed, Moning-Obukhov length, and convective mixing height.

[0023] Step S2, based on the road emission spatial feature input constructed in S1, processes it through a convolutional front-end layer and then uses a residual coding structure to extract multi-scale features from the road emission spatial features. The residual coding structure consists of several concatenated residual blocks. Each residual block achieves hierarchical feature extraction through stacked convolutional layers and shortcut connections. In each residual block, the original spatial mapping function... H (·) By learning the residual function F (·) is used to approximate the expression, that is:

[0024] (7)

[0025] in The input features are shown in Equation (8), and the output form of the residual block is as follows:

[0026] (8)

[0027] in, This represents the convolution operation. W s Optional projective convolution kernel, used to match feature dimensions as needed. The corresponding residual function in each residual block. F (·) consists of a series of convolutional transformations and subsequent nonlinear activation functions, and its form is shown in equation (9):

[0028] (9)

[0029] in, W and b These represent the weights and biases of the convolution filter, respectively. δ (·) represents the ReLU activation function.

[0030] By reformulating the feature learning process into a residual learning form—that is, superimposing the spatial emission features extracted by the current unit with the features extracted by the previous unit—the residual coding structure can effectively preserve key information in deep structures and mitigate performance degradation and gradient vanishing problems that may occur during network deepening. By stacking residual blocks with different receptive fields and downsampling operations, the residual coding structure can capture emission features at multiple spatial scales. Shallow features are primarily used to characterize local emission patterns, while deep features are used to represent larger-scale spatial structural information.

[0031] The final high-dimensional space feature representation is shown in Equation (10).

[0032] (10)

[0033] Where Z represents the set of multi-scale spatial features extracted by the residual coding structure at each layer. In the residual coding structure, the first... i Feature maps output by each residual block and These represent the height and width of the spatial resolution of the residual block's output feature map, respectively. The residual block outputs the number of feature channels. L Φ represents the total number of residual blocks in the residual coding structure. i (·) indicates the first in the residual coding structure i The nonlinear mapping function corresponding to each residual block (including residual convolutional blocks and optional downsampling operations).

[0034] Step S3 is based on the meteorological condition feature input constructed in S1. In each residual module of the residual coding network, a feature-wise linear modulation (FiLM) mechanism is introduced. By mapping the meteorological condition feature vector C, channel-level scaling parameters and offset parameters corresponding to the number of feature channels are generated. Using the channel-level scaling parameters and offset parameters, the intermediate features output by the residual module are linearly modulated channel by channel, so that the emission space features can be adaptively weighted under different meteorological conditions, as shown in Equation (11).

[0035] (11)

[0036] in, γ and β These represent the linear scaling factor and offset factor for each feature channel, respectively; ⊙ represents element-wise multiplication and broadcasting in the spatial dimension; scaling factor γ With offset coefficient β It is calculated by a lightweight multilayer perceptron network based on meteorological condition vectors.

[0037] Step S4: After completing the feature adaptive modulation based on meteorological conditions in Step S3, the residual coding network outputs a high-level spatial latent feature representation, which serves as a shared feature representation for downstream decision-making tasks. This representation is then input to the multi-task decoding module for jointly predicting receptor placement results and interpolation control results. (See below.)

[0038] (12)

[0039] in, This represents the set of predictions output by the model, including receptor deployment predictions. Interpolation control prediction results HGI represents high-level spatial hidden feature representation; W R The learnable weight matrix corresponding to the receptor deployment task is used to map high-dimensional space features to receptor deployment decision results; W M The learnable weight matrix corresponding to the interpolation exclusion task serves to transform shared space features into interpolation control decisions. The "·" indicates that a sigmoid activation function follows the linear transformation. and These represent the height and width of the spatial resolution of the model's output feature map, respectively.

[0040] The multi-task decoding module includes an upsampling layer, a convolutional thinning layer, an adaptive pooling layer, and a multi-task output header connected in sequence.

[0041] The system comprises the following layers: an upsampling layer for spatial scale recovery of low-resolution feature maps from a residual coding network at a preset scaling factor to obtain decoded features with higher spatial resolution; a convolutional thinning layer for local texture enhancement and boundary refinement of the upsampled decoded features (preferably with 64 output channels) and suppression of artifacts introduced by upsampling through convolution operations; and an adaptive pooling layer for scale adaptive reshaping and statistical convergence of the convolutionally thinned features, enhancing robustness to different input scales while maintaining spatial layout information (preferably with 32 output channels). A multi-task output head connected to the adaptive pooling layer generates prediction results for receptor deployment and interpolation control tasks based on shared decoded features. Both tasks have one output channel. The receptor deployment task identifies grid locations where receptors need to be deployed to obtain pollutant concentrations; the interpolation control task identifies deployed receptor locations that do not participate in subsequent spatial interpolation calculations. The outputs of the receptor deployment and interpolation control tasks together determine the effective receptor set for contamination field reconstruction.

[0042] For the model consisting of the residual coding structure, feature-level linear modulation and multi-task decoding module, the basic unit of the model (i.e. the spatial range of processing) is set, data is collected and training and test sets are constructed.

[0043] The training label annotation method is as follows:

[0044] Receptor deployment task label: By traversing and searching different combinations of "emission spatial characteristics – meteorological condition characteristics", the optimal deployment strategy for this combination is selected based on the scheme with the fewest required receptor points, while satisfying the error constraint. The error constraint is defined as follows: the absolute percentage error of the concentration values ​​of each grid estimated based on the optimal deployment strategy and combined with the spatial interpolation method is less than a preset threshold.

[0045] Interpolation control task label: Define the interpolation control position as the following candidate receptor position: When adding a receptor point only at this position, although the concentration estimation of the corresponding grid can meet the error constraint, it will lead to a decrease in overall error accuracy (i.e., the error of other grids increases or the overall error index deteriorates). Label this candidate receptor position to characterize the trade-off between the addition of receptors on satisfying local constraints and global accuracy.

[0046] The model was trained using a training dataset, and the loss function adopted a multi-task joint loss form. Binary cross-entropy loss was applied to both the receptor placement output and the interpolation control output, with an L2 regularization term superimposed. The interpolation control loss was only calculated at locations determined to have receptors already deployed, ensuring consistency between the two task outputs and suppressing overfitting. Finally, the mapping relationship between road emission spatial characteristics, meteorological conditions, and the optimal receptor placement strategy was determined.

[0047] Step S5, based on the mapping relationship between emission spatial characteristics, meteorological condition characteristics and optimal receptor placement strategy constructed in step S4 (i.e. based on the trained model), generates the corresponding optimal receptor placement strategy for the city (output result) for a given combination of road emission characteristics and meteorological conditions (input parameter combination).

[0048] Step S6, based on the optimal receptor placement strategy obtained in step S5, calls the air pollution diffusion model to calculate the pollutant concentration at the receptor point, and reconstructs the continuous pollution concentration field by combining spatial interpolation method. The receptor points of the interpolation control task do not participate in the spatial interpolation calculation, thereby quickly obtaining the pollution field caused by heavy-duty diesel truck emissions in the city and realizing the rapid identification of pollution hotspots.

[0049] The present invention proposes a rapid characterization method for urban heavy diesel emission pollution fields based on dynamic point layout, which has the following beneficial effects:

[0050] To achieve rapid identification of urban pollution hotspots and improve computational efficiency: This invention introduces a dynamic receptor point placement strategy, which avoids the problem of needing to set up a large number of receptor points at the urban scale by the traditional fixed grid or high-density uniform point placement method. While ensuring the accuracy of the pollution concentration field, it significantly reduces the number of receptor points, significantly reduces the computational scale and computation time of pollution diffusion simulation, and is suitable for rapid analysis needs at multiple scenarios, hourly or even higher time resolutions.

[0051] Improving the targeting and resource utilization efficiency of receptor deployment: Based on the joint modeling of road emission spatial characteristics and meteorological conditions, this invention can prioritize the deployment of receptors in key areas where pollution gradients change drastically and pollution hotspots are prone to form, avoiding redundant deployment in low-contribution or low-change areas, thereby achieving optimized allocation of receptor resources and improving the overall efficiency and targeting of pollution field reconstruction.

[0052] This invention provides efficient technical support for traffic environment management and pollution prevention and control decisions: It can quickly characterize and identify pollution fields caused by heavy-duty diesel truck emissions at the city scale in a short period of time, providing timely and reliable technical support for the formulation of traffic restrictions, route optimization, time-based control and refined traffic management measures. Attached Figure Description

[0053] Figure 1 This is a schematic diagram of the overall process of a method for rapidly characterizing urban heavy diesel emission pollution fields based on dynamic point layout proposed in this invention.

[0054] Figure 2 This is a schematic diagram of the deep learning model architecture based on conditional modulation and its key parameter configuration used in the embodiments of the present invention.

[0055] Figure 3 This is a schematic diagram illustrating the spatial distribution effects of the dynamic receptor placement strategy and the dense receptor placement strategy in this invention.

[0056] Figure 4 The NO of heavy-duty diesel trucks at the hourly scale urban road level obtained based on the method of this invention. X Schematic diagram of pollution concentration field caused by emissions. Detailed Implementation

[0057] To provide a clearer understanding of the technical features, objectives, and effects of the present invention, specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0058] like Figure 1 As shown, the present invention provides a method for rapidly characterizing urban heavy diesel emission pollution fields based on dynamic point layout, comprising the following steps:

[0059] S1. Construct joint input data of road emission spatial characteristics and meteorological conditions;

[0060] S2. Extracting multi-scale features of emission space based on residual coding structure;

[0061] S3. Adaptive modulation of multi-scale spatial characteristics based on meteorological conditions;

[0062] S4. Establish the mapping relationship between the conditionally modulated high-dimensional spatial features and the optimal receptor placement strategy;

[0063] S5. Based on the mapping relationship, generate the optimal receptor placement strategy for the city under the combination of specific emission characteristics and meteorological conditions;

[0064] S6. Based on the optimal receptor placement strategy, the pollution diffusion model is invoked and combined with spatial interpolation methods to quickly obtain the pollution field causing emissions from heavy-duty diesel trucks in the city, thereby achieving rapid identification of pollution hotspots.

[0065] This invention embodiment uses the road-grade NO of heavy-duty diesel trucks in Shanghai for a continuous week as an example. X Using emission distribution data as the research object, this invention proposes a method for rapidly characterizing urban heavy-duty diesel emission pollution fields based on dynamic data point allocation. A deep learning model based on conditional modulation is constructed and trained to establish the NOx emission data of heavy-duty diesel trucks on urban roads. X The mapping relationship between emission spatial characteristics, meteorological conditions, and optimal receptor placement strategy is established to achieve NOx emission reduction for heavy-duty diesel trucks at the urban scale, given a combination of emission characteristics and meteorological conditions. X The rapid construction of pollution fields caused by emissions and the efficient identification of pollution hotspots are implemented through the following specific steps:

[0066] S1. Constructing joint input data for road emission spatial characteristics and meteorological conditions: Road emission spatial characteristics are represented using grid data with a fixed spatial resolution (200 m × 200 m in this embodiment), and include at least the following feature types: emission intensity characteristics, used to characterize the pollutant emission level of road segments within the grid cell, specifically the weighted or cumulative result of emissions from each road segment falling within the grid cell; directional characteristics, quantified using Shannon entropy, used to characterize the distribution characteristics of pollutant emissions in different spatial directions, by statistically analyzing the distribution of emissions in multiple directional intervals and calculating its dispersion to reflect the concentration or dispersion of emission directional distribution; spatial heterogeneity characteristics, measured by the weighted Gini coefficient, used to characterize the degree of spatial unevenness of pollutant emissions within the grid cell and its neighborhood, obtained by weighting the differences in emissions at different spatial locations to characterize the spatial aggregation or dispersion of emissions. The results are shown in Table 1.

[0067] Table 1. Gridded emission input data (partial list)

[0068]

[0069] The meteorological condition characteristics are obtained by feature screening of candidate meteorological variables. The screening process is conducted independently for stable and unstable atmospheric conditions. The screening steps include at least: performing multicollinearity analysis on the candidate meteorological variables and iteratively eliminating meteorological variables whose collinearity exceeds a preset threshold until the remaining meteorological variables meet the collinearity constraint; based on the pollutant concentration output results, assessing the degree of influence of each meteorological variable on the change of pollutant concentration, and preferentially retaining meteorological variables that have a dominant influence on the pollution diffusion process to form a combination of meteorological features for model input. (See below.)

[0070] The meteorological feature combination retained under stable meteorological conditions is: wind direction, wind speed, and Monin-Obukhov length; the meteorological feature combination retained under unstable meteorological conditions is: wind direction, sensible heat flux, wind speed, Monin-Obukhov length, and convective mixing height.

[0071] S2. Extracting multi-scale features of emission space based on residual coding structure:

[0072] Constructing deep learning models based on conditional modulation, such as Figure 2 .

[0073] Using S1 as the input of road emission spatial features, a residual coding structure is employed to extract multi-scale features from these features. The encoder consists of a convolutional front-end layer and a series of residual blocks. Each residual block achieves hierarchical feature extraction through stacked convolutional layers and shortcut connections. Within each residual block, the original spatial mapping function... H (·) By learning the residual function F (·) is used to approximate the expression, that is:

[0074] (13)

[0075] The output form of the residual block is shown in equation (14):

[0076] (14)

[0077] residual function F (·) consists of a series of convolutional transformations and subsequent nonlinear activation functions, and its form is shown in equation (15):

[0078] (15)

[0079] By reformulating the feature learning process into a residual learning form, that is, superimposing the spatial emission features extracted by the current unit with the features extracted by the previous unit, the encoder can effectively retain key information in the deep structure and alleviate the performance degradation and gradient vanishing problems that may occur during network deepening. By stacking residual blocks with different receptive fields and downsampling operations, the encoder can capture emission features at multiple spatial scales, where shallow features are mainly used to characterize local emission patterns, while deep features are used to characterize spatial structural information at larger scales. The final high-dimensional spatial feature representation can be expressed as shown in Equation (16). The framework is as follows: Figure 2 As shown.

[0080] (16)

[0081] S3. Adaptive modulation of multi-scale spatial features based on meteorological conditions: Using the meteorological features constructed in S1 as input, a feature-level linear modulation mechanism is introduced into each residual module of the residual coding network. By mapping the condition vector C, channel-level scaling parameters and offset parameters corresponding to the number of feature channels are generated. Using the channel-level scaling parameters and offset parameters, the intermediate features output by the residual module are linearly modulated channel by channel, so that the emission spatial features are adaptively weighted under different meteorological conditions, as shown in Equation (17). The framework is as follows: Figure 2 As shown.

[0082] (17)

[0083] S4. Establish the mapping relationship between the conditionally modulated high-dimensional spatial features and the optimal receptor placement strategy: After completing the feature adaptive modulation based on meteorological conditions in step S3, the encoder-decoder backbone network outputs the high-level spatial latent feature representation. This high-level spatial latent feature representation serves as the shared feature representation for downstream decision-making tasks and is input to the multi-task decoding module for jointly predicting receptor placement results and interpolation control results, i.e.:

[0084] (18)

[0085] Finally, based on the preset model architecture, the basic unit of the model was set (in this embodiment, the basic research unit is 2 km * 2 km, that is, the distribution area is 2 km * 2 km; considering the influence of emission sources on receptors near the grid boundary, the emission characteristic area is a 2.8 km * 2.8 km area centered on the 2 km * 2 km distribution area). The error constraint was set to 25%. After constructing the training dataset, the parameters were calibrated to determine the mapping relationship between the spatial characteristics of road emissions, meteorological conditions, and the corresponding optimal distribution strategy. The model training parameter results are as follows: Figure 2 As shown.

[0086] S5. Based on the mapping relationship, generate the optimal receptor placement strategy for the city under a specific combination of emission characteristics and meteorological conditions: Based on the mapping relationship between emission spatial characteristics, meteorological conditions, and receptor placement strategy established in step S4, Shanghai is divided into multiple 2km*2km research units. For a given combination of road emission characteristics and meteorological conditions, the optimal receptor placement strategy for the city corresponding to each research unit is automatically generated, as shown in the example below. Figure 3 As shown, the top row of dense point distribution maps (traditional method) represents the strategy of evenly distributing points in each grid region within the basic unit. The dynamic point distribution map represents the optimal point distribution strategy proposed in this invention (where the red dots and pentagrams are the optimal combination of receptor points, and the receptor points corresponding to the pentagrams are interpolation control points), which can minimize the number of receptor points while satisfying error constraints. In the middle row, the left image shows the regional contamination field characterized by the traditional method, and the right image shows the regional contamination field obtained by the optimal point distribution strategy proposed in this invention and the spatial interpolation method. The bottom row of images shows the absolute error of the regional contamination field characterized by the two methods. It can be found that the absolute error of the concentration at each grid point is controlled within 5 μg / m³. 3 Within.

[0087] S6. Based on the optimal receptor placement strategy, the pollution diffusion model is invoked and spatial interpolation methods are combined to quickly obtain the pollution field causating heavy-duty diesel truck emissions in the urban area, achieving rapid identification of pollution hotspots: the air pollution diffusion model is invoked to calculate the pollutant concentration at the receptor points, and a continuous pollution concentration field is reconstructed using spatial interpolation methods. All research units are combined to quickly obtain the pollution field causating heavy-duty diesel truck emissions in the urban area, achieving rapid identification of pollution hotspots. An example of hourly pollution field results is shown below. Figure 4 As shown, the distribution of pollution hotspots caused by emissions from heavy-duty diesel trucks in the city exhibits a clear corridor characteristic.

[0088] The above description is merely a description of preferred embodiments of this application and is not intended to limit the scope of this application in any way. Any changes or modifications made by those skilled in the art based on the above-disclosed technical content should be considered as equivalent and valid embodiments and fall within the scope of protection of the technical solution of this application.

Claims

1. A method for rapidly characterizing urban heavy diesel emission pollution fields based on dynamic point layout, characterized in that, Includes the following steps: S1. Construct joint input data of road emission spatial characteristics and meteorological conditions; S2. Extracting multi-scale features of emission space based on residual coding structure; By stacking residual blocks with different receptive fields and downsampling operations, the residual coding structure captures emission features at multiple spatial scales, where shallow features are used to characterize local emission patterns, while deep features are used to characterize spatial structural information at larger scales. The final high-dimensional feature representation is shown in equation (10): (10) Where Z represents the set of multi-scale spatial features extracted by the residual coding structure at each layer. This represents the feature map output by the i-th residual block in the residual coding structure. and Let the height and width of the spatial resolution of the output feature map of the i-th residual block be represented respectively. Output the number of feature channels of the i-th residual block, where L represents the total number of residual blocks in the residual coding structure, and Φi(·) represents the nonlinear mapping function corresponding to the i-th residual block in the residual coding structure; S3. Adaptive modulation of multi-scale spatial characteristics based on meteorological conditions; S4. Establish the mapping relationship between the conditionally modulated high-dimensional spatial features and the optimal receptor placement strategy; S5. Based on the mapping relationship described in step S4, generate the optimal receptor placement strategy for the city under a specific combination of emission characteristics and meteorological conditions; S6. Based on the optimal receptor placement strategy described in step S5, the pollution diffusion model is invoked and combined with spatial interpolation methods to quickly obtain the pollution field causing emissions from heavy-duty diesel trucks in the city, thereby achieving rapid identification of pollution hotspots. Step S3, based on the meteorological condition feature input constructed in S1, introduces a feature-level linear modulation (FiLM) mechanism into each residual module of the residual coding network. By mapping the meteorological condition feature vector C, channel-level scaling parameters and offset parameters corresponding to the number of feature channels are generated. Using the channel-level scaling parameters and offset parameters, the intermediate features output by the residual module are linearly modulated channel by channel, so that the emission space features can be adaptively weighted under different meteorological conditions, as shown in Equation (11): (11) in, γ and β These represent the linear scaling factor and offset factor for each feature channel, respectively; ⊙ represents element-wise multiplication and broadcasting in the spatial dimension; scaling factor γ With offset coefficient β It is calculated by a lightweight multilayer perceptron network based on meteorological condition vectors.

2. The method for rapidly characterizing urban heavy diesel emission pollution fields based on dynamic point layout according to claim 1, characterized in that, In step S1, urban road-level emission data and urban meteorological station meteorological data are used as inputs to construct joint input data of road emission spatial characteristics and meteorological condition characteristics; The spatial characteristics of road emissions are represented using grid data with a fixed spatial resolution, including emission intensity characteristics, directional characteristics, and spatial heterogeneity characteristics, as follows: Emission intensity characteristics are used to characterize the pollutant emission levels of road segments within a grid cell, specifically the weighted or cumulative result of emissions from each road segment falling within that grid cell; Directional characteristics are quantified using Shannon entropy to characterize the distribution characteristics of pollutant emissions in different spatial directions. By statistically analyzing the distribution of emissions in multiple directional intervals and calculating its dispersion, the concentration or dispersion of emission directional distribution can be reflected. Spatial heterogeneity is measured by the weighted Gini coefficient, which is used to characterize the degree of unevenness in the spatial distribution of pollutant emissions within and around the grid cell. It is obtained by weighting the differences in emissions at different spatial locations to characterize the spatial clustering or dispersion of emissions. Meteorological condition characteristics are obtained through feature screening of candidate meteorological variables. The screening process is conducted independently for stable atmospheric conditions and unstable atmospheric conditions, and the screening steps include at least the following: Multicollinearity analysis was performed on candidate meteorological variables, and meteorological variables with collinearity exceeding a preset threshold were eliminated through iterative methods until the remaining meteorological variables met the collinearity constraint conditions. Based on the pollutant concentration output, the influence of each meteorological variable on the change of pollutant concentration is assessed, and the meteorological variables that have a dominant influence on the pollution diffusion process are retained to form a combination of meteorological features for model input.

3. The method for rapidly characterizing urban heavy diesel emission pollution fields based on dynamic point layout according to claim 2, characterized in that, The meteorological feature combination retained under stable meteorological conditions is: wind direction, wind speed, and Monin-Obukhov length; the meteorological feature combination retained under unstable meteorological conditions is: wind direction, sensible heat flux, wind speed, Monin-Obukhov length, and convective mixing height.

4. The method for rapidly characterizing urban heavy diesel emission pollution fields based on dynamic point layout according to claim 1, characterized in that, Step S2 is based on the road emission spatial feature input constructed in S1. After processing by the convolutional front-end layer, the road emission spatial features are extracted at multiple scales using the residual coding structure. The residual coding structure consists of several concatenated residual blocks; each residual block achieves hierarchical feature extraction through stacked convolutional layers and shortcut connections; within each residual block, the original spatial mapping function... H (·) By learning the residual function F (·) is used to approximate the expression, that is: (7) in The input features are shown in Equation (8), and the output form of the residual block is as follows: (8) in, This represents the convolution operation. W s Optional projection convolution kernel, used to match feature dimensions; The residual function corresponding to each residual block F (·) consists of a series of convolutional transformations and subsequent nonlinear activation functions, and its form is shown in equation (9): (9) in, W and b These represent the weights and biases of the convolution filter, respectively. δ (·) represents the ReLU activation function.

5. The method for rapidly characterizing urban heavy diesel emission pollution fields based on dynamic point layout according to claim 1, characterized in that, After completing step S3, which involves adaptive modulation of features based on meteorological conditions, the residual coding network outputs a high-level spatial latent feature representation as a shared feature representation for downstream decision-making tasks. This representation is then input to the multi-task decoding module for jointly predicting receptor placement results and interpolation control results; as follows: (12) in, This represents the set of predictions output by the model, including receptor deployment predictions. Interpolation control prediction results HGI represents high-level spatial hidden feature representation; W R The learnable weight matrix corresponding to the receptor deployment task is used to map high-dimensional space features to receptor deployment decision results; W M The learnable weight matrix corresponding to the interpolation exclusion task serves to transform shared space features into interpolation control decisions; "·" indicates that a sigmoid activation function follows the linear transformation. and These represent the height and width of the spatial resolution of the model's output feature map, respectively. The multi-task decoding module includes an upsampling layer, a convolutional thinning layer, an adaptive pooling layer, and a multi-task output header connected in sequence. The upsampling layer is used to restore the spatial scale of the low-resolution feature map from the residual coding network at a preset ratio in order to obtain decoded features with higher spatial resolution. The convolutional thinning layer is used to enhance the local texture and refine the boundaries of the upsampled decoded features. The number of output channels of the convolutional thinning layer is set to 64, and the artifacts introduced by upsampling are suppressed through convolution operations. The adaptive pooling layer is used to perform scale-adaptive reshaping and statistical convergence on the features after convolution refinement, so that the output features can maintain spatial layout information while enhancing robustness to different input scales. The number of output channels of the adaptive pooling layer is set to 32. The multi-task output head is connected to the adaptive pooling layer and is used to generate prediction results for the receptor deployment task and the interpolation control task respectively based on shared decoding features. The number of output channels for both the receptor deployment task and the interpolation control task is 1. The receptor deployment task is used to identify the grid positions where receptors need to be deployed to obtain pollutant concentrations. The interpolation control task is used to identify the positions of receptors that have been deployed but do not participate in subsequent spatial interpolation calculations. The outputs of the receptor deployment task and the interpolation control task together determine the effective receptor set for contamination field reconstruction. For the model consisting of the above residual coding structure, feature-level linear modulation and multi-task decoding module, the basic unit of the model is defined as the spatial range of processing, data is collected and training and test sets are constructed. The model is trained using the training dataset. The loss function adopts a multi-task joint loss form. By setting binary cross-entropy loss for the receptor placement output and interpolation control output respectively and superimposing an L2 regularization term, the interpolation control loss is only calculated at the positions determined to have been placed receptors, so as to ensure the consistency of the outputs of the two tasks and suppress overfitting. Finally, the mapping relationship between the spatial characteristics of road emissions, meteorological conditions, and the optimal receptor placement strategy was determined.

6. The method for rapidly characterizing urban heavy diesel emission pollution fields based on dynamic point layout according to claim 5, characterized in that, The training label annotation method is as follows: Receptor deployment task label: By traversing and searching different combinations of "emission spatial characteristics - meteorological conditions characteristics", the optimal deployment strategy corresponding to the combination is the one with the fewest required receptor points, under the premise of satisfying the error constraint. The error constraint is defined as: the absolute percentage error of the concentration values ​​of each grid estimated based on the optimal deployment strategy and combined with the spatial interpolation method is less than a preset threshold. Interpolation control task label: Define the interpolation control position as the following receptor candidate position: When only a receptor point is added at this position, although the concentration estimation of the corresponding grid can meet the error constraint, it will lead to a decrease in the overall error accuracy; label the receptor candidate position to characterize the trade-off between the addition of receptors on the satisfaction of local constraints and global accuracy.

7. The method for rapidly characterizing urban heavy diesel emission pollution fields based on dynamic point layout according to claim 1, characterized in that, In step S5, based on the mapping relationship between emission spatial characteristics, meteorological condition characteristics and optimal receptor placement strategy constructed in step S4, a corresponding optimal receptor placement strategy for the city is generated for a given combination of road emission characteristics and meteorological conditions.

8. The method for rapidly characterizing urban heavy diesel emission pollution fields based on dynamic point layout according to claim 1, characterized in that, In step S6, based on the optimal receptor placement strategy obtained in step S5, the air pollution diffusion model is called to calculate the pollutant concentration at the receptor point, and the continuous pollution concentration field is reconstructed by combining the spatial interpolation method. The receptor points of the interpolation control task do not participate in the spatial interpolation calculation, thereby quickly obtaining the pollution field caused by heavy-duty diesel truck emissions in the city and realizing the rapid identification of pollution hotspots.

Images