A deep learning-based atmospheric pollution diffusion modeling method and system

Abstract

This invention discloses a deep learning-based method and system for modeling atmospheric pollution diffusion, comprising the following steps: Step 1: Constructing a two-dimensional spatial grid; Step 2: Constructing a three-dimensional voxel space model; Step 3: Generating pollution information units; Step 4: Constructing a three-dimensional pollution diffusion state field; Step 5: Performing spatiotemporal spectral decomposition of the topological manifold to form a set of pollution diffusion trajectories; Step 6: Inputting the three-dimensional pollution diffusion state field into the state branch of an improved DeepONet model, inputting the set of pollution diffusion trajectories into the trajectory branch, constructing a pollution diffusion operator matrix through an operator projection layer, performing spectral space decomposition, and introducing a spectral manifold folding mechanism to obtain the three-dimensional pollution diffusion state at the next time step; Step 7: Obtaining the updated pollution information units; Step 8: Generating the spatial distribution of pollution concentration. This invention achieves stable modeling of atmospheric pollution diffusion through spatiotemporal spectral decomposition of the topological manifold and an improved DeepONet model.

Description

Technical Field

[0001] This invention relates to the field of atmospheric pollution diffusion modeling technology, and in particular to an atmospheric pollution diffusion modeling method and system based on deep learning. Background Technology

[0002] With the accelerating pace of urbanization and the continuous increase in industrial and transportation emissions, air pollution has become an increasingly prominent problem. Accurate modeling and prediction of pollutant diffusion patterns in urban spaces has become an important research direction in environmental monitoring and pollution control. Existing air pollution diffusion modeling methods mainly rely on traditional methods such as Gaussian diffusion models, computational fluid dynamics models, and numerical simulation models for pollution diffusion calculations. However, these methods generally suffer from the following problems in practical applications: Traditional atmospheric pollution diffusion models are typically based on two-dimensional space, making it difficult to fully reflect the shielding and guiding effects of urban buildings on pollutant diffusion paths. Especially in high-density urban environments, the three-dimensional spatial structure of buildings significantly alters the propagation patterns of pollutants at different altitudes, making it difficult for models based on two-dimensional diffusion assumptions to accurately describe the actual diffusion process of pollutants in complex urban environments. Furthermore, traditional numerical simulation methods often require solving complex partial differential equations, placing high demands on computational resources and time. When the study area is large or the spatial resolution is high, the computational cost of the model increases significantly, making it difficult to meet the need for rapid prediction of large-scale pollution diffusion. In addition, the pollutant diffusion process exhibits significant spatiotemporal nonlinear characteristics. Pollutant emissions, changes in meteorological conditions, and urban spatial structure work together to create complex dynamic changes in pollutant diffusion trajectories. Traditional models have limited ability to characterize the evolution of pollution diffusion paths and major propagation modes, making it difficult to accurately extract the dominant propagation structure of pollution diffusion.

[0003] Therefore, how to provide a deep learning-based method and system for modeling atmospheric pollution diffusion is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] One objective of this invention is to propose a deep learning-based method and system for modeling atmospheric pollution diffusion. This invention combines topological manifold spatiotemporal spectral decomposition with an improved DeepONet model to model and predict the diffusion process of pollution information units in three-dimensional voxel space. It constructs a pollution diffusion operator matrix, performs spectral space decomposition on it, and introduces a spectral manifold folding mechanism to achieve stable identification and dynamic prediction of pollution diffusion paths. This enables high-precision modeling and stable prediction of three-dimensional pollution diffusion processes under the constraints of complex urban spatial structures.

[0005] A deep learning-based atmospheric pollution diffusion modeling method according to an embodiment of the present invention includes the following steps: Step 1: Obtain pollution source emission data and meteorological data for the target area over a continuous time period, and divide the target area into grids to form a two-dimensional spatial grid; Step 2: Obtain building height data for the target area, construct a three-dimensional spatial structure, and divide the three-dimensional spatial structure into voxels to form a three-dimensional voxel space model; Step 3: Generate pollution information units based on the pollution source emission data, and record the initial spatial location, pollutant type, pollution intensity and generation time for each pollution information unit, and map the pollution information units to the three-dimensional voxel space model; Step 4: Count the number of pollution information units and calculate the voxel pollution intensity in the three-dimensional voxel space model, and construct a three-dimensional pollution diffusion state field by combining meteorological data; Step 5: Perform topological manifold spatiotemporal resolution on the positional changes of pollution information units in the three-dimensional voxel space model in continuous time steps to form a set of pollution diffusion trajectories; Step 6: Input the three-dimensional pollution diffusion state field into the state branch of the improved DeepONet model, input the pollution diffusion trajectory set into the trajectory branch of the improved DeepONet model, construct the pollution diffusion operator matrix through the operator projection layer, perform spectral space decomposition on the pollution diffusion operator matrix, and introduce the spectral manifold folding mechanism to obtain the three-dimensional pollution diffusion state of the next time step. Step 7: Update the spatial position of the pollution information unit in the three-dimensional voxel space model according to the three-dimensional pollution diffusion state of the next time step to obtain the updated pollution information unit; Step 8: Map the updated pollution information units in the three-dimensional voxel space model to a two-dimensional spatial grid to generate the spatial distribution of pollution concentration in the target area.

[0006] Optionally, step one specifically includes: Acquire pollution source emission data for a target area over a continuous time period, including pollution source location, pollutant type, emission amount, and emission time; Acquire meteorological data over a continuous time period, including wind speed, wind direction, and temperature; The region boundary is determined based on the spatial extent of the target area, and the region boundary is divided according to the set spatial resolution to generate multiple grid cells; Each grid cell is assigned a unique grid number, and the spatial extent of the corresponding grid cell is recorded to form a two-dimensional spatial grid.

[0007] Optionally, step two specifically involves: Acquire building height data for the target area and record the spatial coordinates of each building; Based on the building height data, a building height volume is generated vertically at the spatial location coordinates to construct the three-dimensional spatial structure of the target area; The three-dimensional spatial structure is regularly divided according to the set voxel side length to generate multiple voxel units; Determine whether the spatial position of each voxel unit falls within the building height volume range, and mark the voxel units that fall within the building height volume range as building voxels, and mark the voxel units that do not fall within the building height volume range as spatial voxels; Record the three-dimensional coordinates and voxel type of each voxel unit to form a three-dimensional voxel space model; During the pollution diffusion calculation process, when a pollution information unit migrates to a target voxel unit, the voxel type of the target voxel unit is determined. If the target voxel unit is a building voxel, the migration of the pollution information unit is prevented; if the target voxel unit is a space voxel, the migration of the pollution information unit is allowed.

[0008] Optionally, step three specifically includes: Extract the location of the pollution source, the type of pollutant, the amount of emissions, and the time of emissions from the pollution source emission data; The emissions from pollution sources are divided into time segments according to a set time interval to obtain the emissions corresponding to each time step; The emissions at each time step are divided according to the set pollution intensity baseline value. The number of corresponding pollution information units is determined and multiple pollution information units are generated. A unique unit number is assigned to each pollution information unit and the pollutant type is recorded. The pollution intensity of a single pollution information unit is the ratio of the emissions at the corresponding time step to the number of pollution information units. The location of the pollution source is recorded as the initial spatial location of the pollution information unit, and the time of the corresponding time step is recorded as the generation time of the pollution information unit. Based on the initial spatial position of each pollution information unit, the corresponding voxel unit is determined in the three-dimensional voxel space model, and each pollution information unit is mapped to the corresponding voxel unit.

[0009] Optionally, step four specifically involves: In the three-dimensional voxel space model, each voxel unit is traversed, and the number of contamination information units in each voxel unit is counted. The contamination intensity of contamination information units within the same voxel unit is accumulated to obtain the voxel contamination intensity of the corresponding voxel unit. Acquire meteorological data corresponding to the current time step, and map the meteorological data to the spatial location of each voxel unit; By combining the voxel pollution intensity and the corresponding meteorological data according to the spatial arrangement of voxel units, a three-dimensional pollution diffusion state field is formed.

[0010] Optionally, step five specifically includes: The voxel coordinates of each pollution information unit in the three-dimensional voxel space model are recorded in continuous time steps. The voxel coordinates include the row index, column index and height index of the voxel in the three-dimensional voxel space model, and the position sequence of the pollution information units is formed in time order. A time connection edge is established based on the voxel coordinate change of the pollution information unit between adjacent time steps. When the voxel coordinate change of the pollution information unit in adjacent time steps does not exceed the set voxel adjacency distance, a time connection edge is established between the corresponding pollution information units. Calculate the spatial distance between the voxel units where each pollution information unit is located within the same time step. When the spatial distance is less than the set spatial adjacency distance, establish a spatial connection edge between the corresponding pollution information units. A spatiotemporal adjacency matrix of pollution information units is constructed based on the time connection edge and the spatial connection edge, wherein the pollution information unit number is used as the matrix row and column index. When there is a time connection edge or a spatial connection edge between two pollution information units, the corresponding matrix element is assigned a value of 1, otherwise it is assigned a value of 0. The corresponding Laplacian matrix is ​​calculated based on the spatiotemporal adjacency matrix, and the Laplacian matrix is ​​decomposed into eigenvalues ​​to obtain the corresponding eigenvalues ​​and eigenvectors. The eigenvectors are then sorted according to the size of the eigenvalues. The first few feature vectors after sorting are selected as the main diffusion spectrum vectors, and the main diffusion direction of the pollution information unit in the three-dimensional voxel space is determined according to the main diffusion spectrum vectors. Based on the main diffusion direction, the voxel coordinate sequence of the corresponding pollution information unit is connected in chronological order to generate a pollution diffusion trajectory, forming a pollution diffusion trajectory set.

[0011] Optionally, step six specifically includes: The improved DeepONet model includes state branching, trajectory branching, operator projection layer, and spectral decomposition operator layer, and embeds a spectral manifold folding mechanism in the spectral decomposition operator layer; The three-dimensional pollution diffusion state field is arranged in the order of row index, column index and height index of voxels in the three-dimensional voxel space model to form a voxel sequence, and then input into the state branch. The state branch includes a three-layer fully connected network to obtain the pollution diffusion state feature vector. The set of pollution diffusion trajectories is input into the trajectory branch, which includes a three-layer fully connected network, to obtain the pollution diffusion trajectory feature vector. The pollution diffusion state feature vector and the pollution diffusion trajectory feature vector are input into the operator projection layer and spliced ​​to form a joint feature vector. The i-th feature element and the j-th feature element in the joint feature vector are used as the matrix row and column indices. The absolute value of the difference between the feature element values ​​is calculated and written into the matrix element in the i-th row and j-th column to obtain the diffusion energy matrix. Summing the elements of each row of the diffusion energy matrix yields the energy sum of the corresponding row. Then, each element in the row is divided by the corresponding energy sum to obtain the operator projection matrix. A pollution diffusion operator matrix is ​​generated by linearly mapping the joint eigenvectors using the operator projection matrix. The pollution diffusion operator matrix is ​​input into the spectral decomposition operator layer for spectral space decomposition. The degree matrix is ​​calculated based on the pollution diffusion operator matrix, and a normalized diffusion Laplace operator is constructed using the degree matrix and the pollution diffusion operator matrix. The normalized diffusion Laplace operator is decomposed into eigenvalues ​​to obtain an eigenvalue sequence and an eigenvector matrix. Several eigenvectors are selected in ascending order of eigenvalues ​​to form a spectral space feature matrix. The spectral decomposition operator layer performs a spectral manifold folding mechanism on the spectral space feature matrix, calculates the Euclidean distance between any two feature vectors in the spectral space feature matrix, divides two feature vectors whose Euclidean distance is less than a set distance threshold into the same spectral neighborhood, performs a weighted summation of the feature vectors in each spectral neighborhood using the reciprocal of the corresponding eigenvalue as the weight, generates the corresponding folded feature vector, and replaces the feature vectors in the original spectral neighborhood with the folded feature vectors corresponding to each spectral neighborhood to obtain the folded spectral space feature matrix; The folded spectral space feature matrix is ​​multiplied with its transpose to generate an updated pollution diffusion operator matrix. The updated pollution diffusion operator matrix is ​​then used to calculate the voxel pollution intensity of each voxel unit in the three-dimensional voxel space model at the next time step, thus obtaining the three-dimensional pollution diffusion state at the next time step.

[0012] Optionally, step seven specifically includes: Based on the three-dimensional contamination diffusion state at the next time step, read the voxel contamination intensity of each voxel unit in the three-dimensional voxel space model. For each pollution information unit, obtain the voxel coordinates of the current voxel unit and determine the set of voxel units adjacent to the current voxel unit; Compare the voxel contamination intensity of each voxel unit in adjacent voxel unit sets, and determine the voxel unit with the highest voxel contamination intensity as the target voxel unit. When a contamination information unit migrates to a target voxel unit, the migration judgment is made based on the voxel type of the target voxel unit. When the target voxel unit allows the contamination information unit to migrate in, the spatial position of the contamination information unit is updated from the voxel coordinates corresponding to the current voxel unit to the voxel coordinates corresponding to the target voxel unit, and the updated voxel coordinates are recorded as the spatial position of the contamination information unit at the current time step, thus obtaining the updated contamination information unit.

[0013] Optionally, step eight specifically includes: The corresponding voxel unit is determined based on the voxel coordinates of the updated pollution information unit in the three-dimensional voxel space model, and the number of pollution information units in each voxel unit is counted. The contamination intensity of contamination information units within the same voxel unit is accumulated to obtain the voxel contamination intensity of the corresponding voxel unit. The voxel cells with the same row index and column index in the three-dimensional voxel space model are traversed according to the height index, and the voxel contamination intensity of the corresponding voxel cells is accumulated to obtain the aggregate contamination intensity at the corresponding row index and column index position. The aggregated pollution intensity is mapped to the corresponding two-dimensional spatial grid cell, and the spatial distribution of pollution concentration in the target area is generated according to the row and column arrangement of the two-dimensional spatial grid.

[0014] An atmospheric pollution diffusion modeling system based on deep learning according to an embodiment of the present invention includes the following modules: The data acquisition module is used to acquire pollution source emission data and meteorological data of the target area over a continuous time period, and to divide the target area into a grid to form a two-dimensional spatial grid. The 3D space construction module is used to acquire building height data of the target area, construct a 3D space structure, and divide the 3D space structure into voxels to form a 3D voxel space model. The pollution information unit generation module is used to generate pollution information units based on the pollution source emission data, record the initial spatial location, pollutant type, pollution intensity and generation time for each pollution information unit, and map the pollution information unit to a three-dimensional voxel space model. The diffusion state construction module is used to count the number of pollution information units and calculate the voxel pollution intensity in the three-dimensional voxel space model, and to construct a three-dimensional pollution diffusion state field in combination with meteorological data. The diffusion trajectory extraction module is used to perform topological manifold spatiotemporal spectral analysis on the positional changes of pollution information units in the three-dimensional voxel space model in continuous time steps to form a set of pollution diffusion trajectories. The diffusion prediction module is used to input the three-dimensional pollution diffusion state field into the state branch of the improved DeepONet model, input the pollution diffusion trajectory set into the trajectory branch of the improved DeepONet model, construct the pollution diffusion operator matrix through the operator projection layer, perform spectral space decomposition on the pollution diffusion operator matrix and introduce a spectral manifold folding mechanism to obtain the three-dimensional pollution diffusion state of the next time step. The pollution unit update module is used to update the spatial position of the pollution information unit in the three-dimensional voxel space model according to the three-dimensional pollution diffusion state of the next time step, so as to obtain the updated pollution information unit. The pollution concentration generation module is used to map the distribution of updated pollution information units in the three-dimensional voxel space model to a two-dimensional spatial grid, thereby generating the spatial distribution of pollution concentration in the target area.

[0015] The beneficial effects of this invention are: This invention addresses the challenge of traditional atmospheric pollution diffusion models failing to characterize the evolution of pollution diffusion paths and the influence of complex three-dimensional spatial structures by jointly modeling the pollution diffusion process using topological manifold spatiotemporal spectral analysis and an improved DeepONet model. It employs topological manifold spatiotemporal spectral analysis to analyze the positional changes of pollution information units within a three-dimensional voxel space model at continuous time steps. By constructing a spatiotemporal adjacency matrix and extracting the main diffusion spectrum vector through Laplace matrix eigenvalue decomposition, a set of pollution diffusion trajectories is formed, thereby identifying the main propagation direction of pollution diffusion. Furthermore, in the pollution diffusion prediction stage, the three-dimensional pollution diffusion state field is input into the state branch of the improved DeepONet model, and the set of pollution diffusion trajectories is... The invention combines input trajectory branches, constructs a pollution diffusion operator matrix through an operator projection layer, and simultaneously performs spectral space decomposition in the spectral decomposition operator layer and introduces a spectral manifold folding mechanism to perform neighborhood folding processing on spectral space features to eliminate high-frequency noise and retain the main diffusion modes, thereby obtaining the three-dimensional pollution diffusion state of the next time step. Furthermore, the invention constructs a three-dimensional voxel space model and generates pollution information units to achieve discrete representation of pollution source emissions in three-dimensional space. Under the constraints of the voxel space structure, the positions of the pollution information units are updated, and finally, the updated pollution information unit distribution is mapped to a two-dimensional spatial grid to generate a spatial distribution of pollution concentration. This enables stable modeling and high-precision prediction of pollution diffusion processes in complex urban environments. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of a deep learning-based atmospheric pollution diffusion modeling method proposed in this invention. Figure 2This is a schematic diagram of the improved DeepONet model structure, which is a deep learning-based atmospheric pollution diffusion modeling method proposed in this invention. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0018] refer to Figure 1 and Figure 2 A deep learning-based method for modeling atmospheric pollution diffusion includes the following steps: Step 1: Obtain pollution source emission data and meteorological data for the target area over a continuous time period, and divide the target area into grids to form a two-dimensional spatial grid; Step 2: Obtain building height data for the target area, construct a three-dimensional spatial structure, and divide the three-dimensional spatial structure into voxels to form a three-dimensional voxel space model; Step 3: Generate pollution information units based on pollution source emission data, and record the initial spatial location, pollutant type, pollution intensity and generation time for each pollution information unit, and map the pollution information units to a three-dimensional voxel space model; Step 4: Count the number of pollution information units and calculate the voxel pollution intensity in the three-dimensional voxel space model, and construct a three-dimensional pollution diffusion state field by combining meteorological data; Step 5: Perform topological manifold spatiotemporal spectral analysis on the positional changes of pollution information units in the three-dimensional voxel space model in continuous time steps to form a set of pollution diffusion trajectories; Step 6: Input the three-dimensional pollution diffusion state field into the state branch of the improved DeepONet model, input the pollution diffusion trajectory set into the trajectory branch of the improved DeepONet model, construct the pollution diffusion operator matrix through the operator projection layer, perform spectral space decomposition on the pollution diffusion operator matrix, and introduce the spectral manifold folding mechanism to obtain the three-dimensional pollution diffusion state of the next time step. Step 7: Update the spatial position of the pollution information unit in the three-dimensional voxel space model according to the three-dimensional pollution diffusion state of the next time step to obtain the updated pollution information unit; Step 8: Map the updated pollution information units in the three-dimensional voxel space model to a two-dimensional spatial grid to generate the spatial distribution of pollution concentration in the target area.

[0019] In this embodiment, step one specifically includes: Acquire pollution source emission data for the target area over a continuous time period. The pollution source emission data includes the location of the pollution source, the type of pollutant, the emission amount, and the emission time. Acquire meteorological data over a continuous period of time, including wind speed, wind direction, and temperature; The region boundary is determined based on the spatial extent of the target area, and the region boundary is divided according to the set spatial resolution to generate multiple grid cells; Each grid cell is assigned a unique grid number, and the spatial extent of the corresponding grid cell is recorded to form a two-dimensional spatial grid. In this invention, the target area is selected as a study area at the city or regional scale. Pollution source emission data is obtained through a fixed emission source list or emission list database, and the longitude, latitude, pollutant type, emission amount, and emission time of each pollution source are recorded. Meteorological data is obtained through meteorological monitoring stations and uniformly converted to the same time resolution as the pollution source emission data. The meteorological data includes wind speed, wind direction, and temperature, where the wind speed unit is m / s and the wind direction is represented from 0° to 360°. The regional boundary is determined according to the longitude and latitude range of the target area and is regularly divided. The spatial resolution is set to 500m×500m, and the area is divided into multiple regular grid units. Each grid unit is assigned a unique grid number and its longitude and latitude spatial range is recorded, thereby forming a two-dimensional spatial grid structure covering the target area.

[0020] In this embodiment, step two specifically involves: Acquire building height data for the target area and record the spatial coordinates of each building; Based on the building height data, a building height volume is generated vertically at the spatial location coordinates to construct the three-dimensional spatial structure of the target area; The three-dimensional spatial structure is regularly divided according to the set voxel side length to generate multiple voxel units; Determine whether the spatial position of each voxel unit falls within the building height volume range, and mark the voxel units that fall within the building height volume range as building voxels, and mark the voxel units that do not fall within the building height volume range as spatial voxels; Record the three-dimensional coordinates and voxel type of each voxel unit to form a three-dimensional voxel space model; During the pollution diffusion calculation process, when a pollution information unit migrates to a target voxel unit, the voxel type of the target voxel unit is determined. If the target voxel unit is a building voxel, the migration of the pollution information unit is prevented. If the target voxel unit is a spatial voxel, the migration of the pollution information unit is allowed. In this invention, the building height data of the target area comes from a three-dimensional database of urban buildings. The building height data includes the building height value and the corresponding spatial coordinates. Based on the building height data, a building height volume is generated vertically at the corresponding spatial coordinates to construct the three-dimensional spatial structure of the target area. To achieve spatial discretization in pollution diffusion calculation, the three-dimensional spatial structure is divided according to a set voxel side length, where the voxel side length is set to 5 meters, forming voxel units with dimensions of 5m×5m×5m in three-dimensional space. Then, it is determined whether the spatial position of each voxel unit falls within the range of the building height volume. When the spatial coordinates of the center point of the voxel unit are located inside the building height volume, the voxel unit is marked as a building voxel. When located outside the building's height, voxel units are marked as spatial voxels, and the three-dimensional coordinates and corresponding voxel types of each voxel unit are recorded to form a three-dimensional voxel spatial model. During the pollution diffusion calculation process, when a pollution information unit migrates to a target voxel unit based on the diffusion calculation results, the voxel type of the target voxel unit is first determined. If the target voxel unit is a building voxel, the pollution information unit is prevented from entering, and the migration direction is recalculated. If the target voxel unit is a spatial voxel, the pollution information unit is allowed to enter and the diffusion calculation continues. Through the above technical solution, the blocking effect of buildings on the diffusion path of pollutants can be realistically reflected in three-dimensional space, so that the pollution diffusion process is no longer limited to two-dimensional planar diffusion, thereby improving the spatial accuracy and physical realism of the pollution diffusion simulation results.

[0021] In this embodiment, step three specifically includes: Extract the location of the pollution source, the type of pollutant, the amount of emissions, and the time of emissions from the pollution source emission data; The emissions from pollution sources are divided into time segments according to a set time interval to obtain the emissions corresponding to each time step; The emissions at each time step are divided according to the set pollution intensity baseline value. The number of corresponding pollution information units is determined and multiple pollution information units are generated. A unique unit number is assigned to each pollution information unit and the pollutant type is recorded. The pollution intensity of a single pollution information unit is the ratio of the emissions at the corresponding time step to the number of pollution information units. The location of the pollution source is recorded as the initial spatial location of the pollution information unit, and the time of the corresponding time step is recorded as the generation time of the pollution information unit. Based on the initial spatial position of each pollution information unit, the corresponding voxel unit is determined in the three-dimensional voxel space model, and each pollution information unit is mapped to the corresponding voxel unit. In this invention, pollution source emission data is divided into time segments according to a fixed time interval, set to 60 seconds. Simultaneously, a pollution intensity base value of 1g is set, and each 1g emission corresponds to the generation of a pollution information unit. When the emission amount at a certain time step is Q, a quantity of Q / 1g of pollution information units is generated. Each pollution information unit is assigned a unique unit number and inherits the corresponding pollutant type. The latitude and longitude coordinates of the pollution source are used as the initial spatial position of the pollution information unit, and the timestamp of that time step is recorded as the generation time of the pollution information unit. Based on the initial spatial position, the corresponding voxel unit is determined in the three-dimensional voxel space model and the mapping is completed. By discretizing continuous emissions into pollution information units with uniform intensity, the pollution diffusion process can be calculated in the form of discrete particles in the three-dimensional voxel space, improving the correspondence accuracy between pollution source emissions and spatial diffusion, thereby enhancing the stability and computational efficiency of pollution diffusion modeling.

[0022] In this embodiment, step four specifically includes: In the three-dimensional voxel space model, traverse each voxel unit and count the number of contamination information units in each voxel unit. The contamination intensity of contamination information units within the same voxel unit is accumulated to obtain the voxel contamination intensity of the corresponding voxel unit. Acquire meteorological data corresponding to the current time step, and map the meteorological data to the spatial location of each voxel unit; The voxel pollution intensity and the corresponding meteorological data are combined according to the spatial arrangement of voxel units to form a three-dimensional pollution diffusion state field. This step involves statistically analyzing pollution information units within a 3D voxel space model and cumulatively calculating the pollution intensity within each voxel unit to obtain the voxel pollution intensity corresponding to each voxel unit. Simultaneously, the voxel pollution intensity is spatially mapped to meteorological data at the corresponding time step and combined to form a 3D pollution diffusion state field. This processing method allows the pollution diffusion state to be expressed in 3D space, reflecting the distribution of pollutants at different altitudes. Furthermore, spatial aggregation of pollution information units using voxel units reduces the discrete fluctuations caused by individual pollution information units, improving the stability of the pollution diffusion state data. Finally, constructing a 3D pollution diffusion state field by spatially mapping meteorological data to voxel pollution intensity provides a unified input structure for the deep learning model, containing both pollution distribution and meteorological condition information, thereby improving the accuracy and stability of pollution diffusion process modeling.

[0023] In this embodiment, step five specifically includes: The voxel coordinates of each pollution information unit in the three-dimensional voxel space model are recorded in continuous time steps. The voxel coordinates include the row index, column index and height index of the voxel in the three-dimensional voxel space model, and the position sequence of the pollution information units is formed in time order. A time connection edge is established based on the voxel coordinate change of the pollution information unit between adjacent time steps. When the voxel coordinate change of the pollution information unit in adjacent time steps does not exceed the set voxel adjacency distance, a time connection edge is established between the corresponding pollution information units. Calculate the spatial distance between the voxel units where each pollution information unit is located within the same time step. When the spatial distance is less than the set spatial adjacency distance, establish a spatial connection edge between the corresponding pollution information units. The spatiotemporal adjacency matrix of pollution information units is constructed based on the temporal and spatial connection edges, where the pollution information unit number is used as the matrix row and column index. When there is a temporal or spatial connection edge between two pollution information units, the corresponding matrix element is assigned a value of 1, otherwise it is assigned a value of 0. Calculate the corresponding Laplacian matrix based on the spatiotemporal adjacency matrix, perform eigenvalue decomposition on the Laplacian matrix to obtain the corresponding eigenvalues ​​and eigenvectors, and sort the eigenvectors according to the size of the eigenvalues; The first few eigenvectors after sorting are selected as the main diffusion spectrum vectors, and the main diffusion direction of the pollution information unit in the three-dimensional voxel space is determined based on the main diffusion spectrum vectors. By connecting the voxel coordinate sequences of the corresponding pollution information units in chronological order according to the main diffusion direction, a pollution diffusion trajectory is generated, forming a pollution diffusion trajectory set. In step five, the positional changes of pollution information units in the three-dimensional voxel space model are processed by topological manifold spatiotemporal spectral analysis. First, the voxel coordinates of pollution information units in continuous time steps are used as basic data, where the voxel coordinates are composed of row index, column index, and height index. Temporal connection edges are established for pollution information units whose voxel coordinate changes do not exceed one voxel unit between adjacent time steps. Within the same time step, when the three-dimensional Euclidean distance between the voxels containing two pollution information units is less than 10m, spatial connection edges are established. The spatiotemporal adjacency matrix of pollution information units is constructed based on the temporal and spatial connection edges, and the corresponding Laplacian matrix is ​​calculated. The Laplacian matrix is ​​decomposed into eigenvalues, and the top three eigenvectors with the smallest eigenvalues ​​are extracted as the main diffusion spectrum vectors. The main diffusion direction of pollution information units in the three-dimensional voxel space is determined by the main diffusion spectrum vectors, and the corresponding voxel coordinate sequences are connected in time order along the main diffusion direction to form the pollution diffusion trajectory.

[0024] In this context, the topological manifold refers to a spatiotemporally continuous structure composed of the spatial positions of pollution information units and their interconnections within consecutive time steps. This structure exhibits geometric properties consistent with three-dimensional Euclidean space within a local range and forms a continuous low-dimensional surface structure through spatiotemporal adjacency relationships. Within this structure, the migration trajectories of pollution information units in three-dimensional voxel space are organized into a continuous spatial distribution, thus forming a pollution diffusion manifold. By performing spectral decomposition on this pollution diffusion manifold, the dominant propagation mode of pollution diffusion can be extracted while maintaining the spatiotemporal structural relationships. This allows the complex three-dimensional pollution diffusion process to manifest as a stable low-dimensional structure in the spectral space, improving the stability and accuracy of pollution diffusion trajectory identification and reducing the impact of local random diffusion on the modeling results.

[0025] In this embodiment, step six specifically includes: The improved DeepONet model includes state branching, trajectory branching, operator projection layer, and spectral decomposition operator layer, and embeds a spectral manifold folding mechanism in the spectral decomposition operator layer; The three-dimensional pollution diffusion state field is arranged in the order of row index, column index and height index of voxels in the three-dimensional voxel space model to form a voxel sequence, and input into the state branch. The state branch includes three fully connected network layers with 512, 256 and 128 neurons in each layer, respectively, to obtain the pollution diffusion state feature vector. The set of pollution diffusion trajectories is input into the trajectory branch, which consists of a three-layer fully connected network with 256, 128 and 64 neurons in each layer, respectively, to obtain the pollution diffusion trajectory feature vector. The pollution diffusion state feature vector and the pollution diffusion trajectory feature vector are input into the operator projection layer and spliced ​​to form a joint feature vector. The i-th feature element and the j-th feature element in the joint feature vector are used as the matrix row and column indices. The absolute value of the difference between the feature element values ​​is calculated and written into the matrix element in the i-th row and j-th column to obtain the diffusion energy matrix. Summing the elements of each row of the diffusion energy matrix yields the energy sum of the corresponding row. Then, each element in the row is divided by the corresponding energy sum to obtain the operator projection matrix. A pollution diffusion operator matrix is ​​generated by linearly mapping the joint eigenvectors using the operator projection matrix. The pollution diffusion operator matrix is ​​input into the spectral decomposition operator layer for spectral space decomposition. The degree matrix is ​​calculated based on the pollution diffusion operator matrix, and a normalized diffusion Laplace operator is constructed using the degree matrix and the pollution diffusion operator matrix. The normalized diffusion Laplace operator is decomposed into eigenvalues ​​to obtain the eigenvalue sequence and eigenvector matrix. Several eigenvectors are selected in ascending order of eigenvalues ​​to form the spectral space feature matrix. Specifically, each row of the pollution diffusion operator matrix is ​​considered as a diffusion connection set. All elements in the same row are summed, and the sum is written into the corresponding diagonal position to form a diagonal matrix. The remaining off-diagonal elements are set to zero, thus obtaining the degree matrix. This degree matrix reflects the overall diffusion connection strength between each feature node and other nodes. Then, a normalized diffusion Laplace operator is constructed using the degree matrix and the pollution diffusion operator matrix. The construction method is as follows: first, the reciprocal of the square root of the diagonal elements of the degree matrix is ​​calculated to form a normalized matrix of the degree matrix. Then, this normalized matrix is ​​multiplied on the left and right by the pollution diffusion operator matrix, and the product result is subtracted from the identity matrix, thus obtaining the normalized diffusion Laplace operator. This normalized diffusion Laplace operator can... The scale effect caused by the difference in the number of connections between different nodes is eliminated, making the diffusion structure more stable. Then, the normalized diffusion Laplace operator is decomposed into eigenvalues ​​to obtain the corresponding eigenvalue sequence and eigenvector matrix, which are sorted in ascending order of eigenvalues. Since the smaller eigenvalues ​​of the normalized Laplace operator correspond to the low-frequency components in the diffusion diagram structure, which can characterize the overall propagation trend of pollution diffusion, while the larger eigenvalues ​​usually correspond to local disturbances or high-frequency noise, selecting eigenvectors with smaller eigenvalues ​​can more stably describe the main propagation mode of pollution diffusion. In this embodiment, the first 6 eigenvectors are selected to form the spectral space eigenvector matrix, thereby reducing the spectral space dimension while maintaining the main diffusion structure information and improving the stability of diffusion evolution calculation. The spectral decomposition operator layer performs a spectral manifold folding mechanism on the spectral space feature matrix. It calculates the Euclidean distance between any two eigenvectors in the spectral space feature matrix, and divides two eigenvectors whose Euclidean distance is less than a set distance threshold into the same spectral neighborhood. The eigenvectors in each spectral neighborhood are weighted and summed using the reciprocal of their corresponding eigenvalues ​​as weights to generate the corresponding folded eigenvectors. The folded eigenvectors of each spectral neighborhood replace the eigenvectors in the original spectral neighborhood, resulting in the folded spectral space feature matrix. In this embodiment, the distance threshold is set to 0.3. When the Euclidean distance between two eigenvectors is less than 0.3, the two eigenvectors are divided into the same spectral neighborhood. The folded spectral space feature matrix is ​​multiplied with its transpose to generate an updated pollution diffusion operator matrix. The updated pollution diffusion operator matrix is ​​then used to calculate the voxel pollution intensity of each voxel unit in the three-dimensional voxel space model at the next time step, thus obtaining the three-dimensional pollution diffusion state at the next time step. The improved DeepONet model still inherits the operator learning framework of the existing DeepONet model in terms of overall structure. That is, it extracts features through input information and builds operator mapping relationships in the intermediate layer to realize the mapping relationship between the input function space and the output function space. Compared with the traditional DeepONet model, this invention also uses a multi-layer fully connected network as the feature extraction structure. It obtains high-dimensional feature representation by performing nonlinear mapping on the input data, thereby realizing the ability to model complex pollution diffusion processes.

[0026] Building upon this foundation, this invention structurally improves the traditional DeepONet model for pollution diffusion modeling scenarios. First, the original single-input structure of the DeepONet model is expanded into a dual-input structure with a state branch and a trajectory branch. The state branch extracts the spatial distribution features of the three-dimensional pollution diffusion state field, while the trajectory branch extracts the spatiotemporal propagation features of the pollution diffusion trajectory set. This allows the model to simultaneously characterize both the spatial distribution and diffusion path information of pollution diffusion. Second, an operator projection layer is added after the two branch networks. By constructing and normalizing the diffusion energy matrix, an operator projection matrix is ​​obtained, enabling the joint feature vector to be mapped to the pollution diffusion operator matrix, thus establishing a correspondence between pollution diffusion features and diffusion dynamics. Furthermore, a spectral decomposition operator layer is introduced, embedding a spectral manifold folding mechanism within this layer. The pollution diffusion operator is spectrally decomposed using a normalized diffusion Laplacian operator, and neighborhood folding is applied to the spectral spatial features to eliminate high-frequency noise features while preserving the main diffusion modes.

[0027] Through the above improvements, this invention can explicitly introduce the topological structure information of pollution diffusion during the operator learning process, enabling the model to not only learn the numerical variation law of pollution diffusion, but also capture the main propagation structure of pollution diffusion, thereby improving the stability and accuracy of pollution diffusion prediction results and reducing noise interference in the process of complex three-dimensional diffusion field modeling.

[0028] In this embodiment, step seven specifically includes: Based on the three-dimensional contamination diffusion state at the next time step, read the voxel contamination intensity of each voxel unit in the three-dimensional voxel space model. For each pollution information unit, obtain the voxel coordinates of the current voxel unit and determine the set of voxel units adjacent to the current voxel unit; Compare the voxel contamination intensity of each voxel unit in adjacent voxel unit sets, and determine the voxel unit with the highest voxel contamination intensity as the target voxel unit. When a contamination information unit migrates to a target voxel unit, the migration judgment is made based on the voxel type of the target voxel unit. When the target voxel unit allows the contamination information unit to migrate in, the spatial position of the contamination information unit is updated from the voxel coordinates corresponding to the current voxel unit to the voxel coordinates corresponding to the target voxel unit, and the updated voxel coordinates are recorded as the spatial position of the contamination information unit at the current time step, thus obtaining the updated contamination information unit. After obtaining the three-dimensional pollution diffusion state at the next time step, this invention updates the position of pollution information units based on the voxel pollution intensity of each voxel unit in the three-dimensional voxel space model. First, the voxel pollution intensity of each voxel unit is read, and a set of neighboring voxel units is determined, centered on the voxel unit where the pollution information unit currently resides. Then, the voxel pollution intensity of each voxel unit is compared within the neighboring voxel unit set, and the voxel unit with the highest pollution intensity is selected as the candidate migration position for the pollution information unit in the next time step. Simultaneously, the migration process is constrained by the pre-labeled voxel types in the three-dimensional voxel space model. When the target voxel unit is a spatial voxel that allows the pollution information unit to migrate into, the voxel coordinates of the pollution information unit are updated to the voxel coordinates corresponding to the target voxel unit. When the target voxel unit is a building voxel, the pollution information unit remains in its original voxel unit. Through this method, the pollution information unit gradually migrates in the voxel space, enabling the pollution diffusion process to be updated under the constraints of the three-dimensional spatial structure.

[0029] In this embodiment, step eight specifically includes: The corresponding voxel unit is determined based on the voxel coordinates of the updated pollution information unit in the three-dimensional voxel space model, and the number of pollution information units in each voxel unit is counted. The contamination intensity of contamination information units within the same voxel unit is accumulated to obtain the voxel contamination intensity of the corresponding voxel unit. The voxel cells with the same row index and column index in the three-dimensional voxel space model are traversed according to the height index, and the voxel contamination intensity of the corresponding voxel cells is accumulated to obtain the aggregate contamination intensity at the corresponding row index and column index position. The aggregated pollution intensity is mapped to the corresponding two-dimensional spatial grid cell, and the spatial distribution of pollution concentration in the target area is generated according to the row and column arrangement of the two-dimensional spatial grid. In this invention, by statistically analyzing the distribution of the updated pollution information units in the three-dimensional voxel space model, the voxel pollution intensity of each pollution information unit is first calculated. Then, voxel units with the same row index and column index are accumulated along the height direction to obtain the aggregated pollution intensity at the corresponding two-dimensional position. This aggregated pollution intensity is then mapped to the two-dimensional spatial grid, thereby realizing the conversion of the three-dimensional pollution diffusion result to the two-dimensional pollution concentration distribution. This allows the pollution diffusion modeling result to intuitively reflect the surface pollution concentration distribution of the target area.

[0030] A deep learning-based atmospheric pollution diffusion modeling system includes the following modules: The data acquisition module is used to acquire pollution source emission data and meteorological data of the target area over a continuous time period, and to divide the target area into a grid to form a two-dimensional spatial grid. The 3D space construction module is used to acquire building height data of the target area, construct a 3D space structure, and divide the 3D space structure into voxels to form a 3D voxel space model. The pollution information unit generation module is used to generate pollution information units based on pollution source emission data. It records the initial spatial location, pollutant type, pollution intensity and generation time for each pollution information unit, and maps the pollution information unit to a three-dimensional voxel space model. The diffusion state construction module is used to count the number of pollution information units and calculate the voxel pollution intensity in the three-dimensional voxel space model, and to construct a three-dimensional pollution diffusion state field in combination with meteorological data. The diffusion trajectory extraction module is used to perform topological manifold spatiotemporal spectral analysis on the positional changes of pollution information units in the three-dimensional voxel space model in continuous time steps to form a set of pollution diffusion trajectories. The diffusion prediction module is used to input the three-dimensional pollution diffusion state field into the state branch of the improved DeepONet model, input the pollution diffusion trajectory set into the trajectory branch of the improved DeepONet model, construct the pollution diffusion operator matrix through the operator projection layer, perform spectral space decomposition on the pollution diffusion operator matrix and introduce a spectral manifold folding mechanism to obtain the three-dimensional pollution diffusion state of the next time step. The pollution unit update module is used to update the spatial position of the pollution information unit in the three-dimensional voxel space model according to the three-dimensional pollution diffusion state of the next time step, so as to obtain the updated pollution information unit. The pollution concentration generation module is used to map the distribution of updated pollution information units in the three-dimensional voxel space model to a two-dimensional spatial grid, thereby generating the spatial distribution of pollution concentration in the target area.

[0031] Example 1: To verify the feasibility of this invention in practice, it was applied to an atmospheric pollution diffusion prediction scenario at the city scale, modeling and analyzing the pollutant diffusion process in a complex built environment. In this application scenario, the study area is approximately 95 km². 2The area encompasses various typical urban functional zones, including residential areas, commercial areas, and major traffic arteries. Building heights vary significantly, ranging from 8m to 110m, with a large number of high-rise buildings and a high building density. In real-world environments, buildings significantly obstruct and guide airflow, causing pollutant diffusion paths to change markedly between different height levels. Traditional two-dimensional pollution diffusion models struggle to accurately describe the pollution diffusion process under such complex spatial structures, leading to substantial errors in pollution concentration prediction. Furthermore, pollution emissions exhibit significant temporal variations, while wind speed, direction, and temperature conditions continuously change, resulting in pronounced spatiotemporal nonlinearities in the pollution diffusion process. Therefore, the method described in this invention is employed for pollution diffusion modeling and prediction in this scenario.

[0032] In practical applications, firstly, pollution source emission data and meteorological data for the target area over a continuous time period are acquired, and then a two-dimensional spatial grid is formed based on the area boundaries. Subsequently, building height data for the target area is acquired and a three-dimensional spatial structure is constructed. This three-dimensional spatial structure is then divided into voxels with a side length of 5m × 5m × 5m, forming a three-dimensional voxel space model. In this embodiment, the three-dimensional voxel space model contains approximately 7.6 × 10⁻⁶ voxels. 6 Individual voxel units are generated, with building voxels accounting for approximately 30.9% of the total number of voxels and spatial voxels accounting for approximately 69.1%. Subsequently, pollution information units are generated based on pollution source emission data, and the initial spatial location, pollutant type, pollution intensity, and generation time of each pollution information unit are recorded. The pollution information units are then mapped to a three-dimensional voxel space model. The number of pollution information units within a voxel is counted and the voxel pollution intensity is calculated in continuous time steps. At the same time, a three-dimensional pollution diffusion state field is constructed by combining meteorological data such as wind speed, wind direction, and temperature.

[0033] In the pollution diffusion trajectory extraction stage, topological manifold spatiotemporal spectral processing is performed on the positional changes of pollution information units in the 3D voxel space model to construct the spatiotemporal adjacency matrix of the pollution information units and calculate the corresponding Laplacian matrix. The main diffusion spectrum vector is extracted through eigenvalue decomposition, thus forming a set of pollution diffusion trajectories. Subsequently, the 3D pollution diffusion state field is input into the state branch of the improved DeepONet model, and the set of pollution diffusion trajectories is input into the trajectory branch. The pollution diffusion operator matrix is ​​constructed through the operator projection layer. At the same time, spectral space decomposition is performed in the spectral decomposition operator layer, and a spectral manifold folding mechanism is introduced to obtain the 3D pollution diffusion state of the next time step. The spatial position of the pollution information units in the 3D voxel space model is updated according to the prediction results, and the migration of pollution information units is completed under the voxel type constraint. Finally, the updated pollution information unit distribution is mapped to a 2D spatial grid to generate the spatial distribution of pollution concentration in the target area.

[0034] To verify the performance of the method of the present invention in pollution concentration prediction, multiple air quality monitoring points were set up in the study area, and the measured pollution concentrations at the monitoring points were recorded. At the same time, the traditional Gaussian diffusion model, the LSTM deep neural network model, and the method of the present invention were used to predict the pollution concentrations.

[0035] Table 1. Comparison of Pollution Concentration Prediction Results under Complex Urban Spatial Structures The statistical analysis of the pollution concentration prediction results of different models in Table 1 above can intuitively reflect the differences in accuracy of each model in pollution diffusion prediction; the Gaussian diffusion model has errors of 21 μg / m³ at monitoring points 1 to 6. 3 18μg / m 3 18μg / m 3 21μg / m 3 17μg / m 3 and 18μg / m 3 The total error is 113 μg / m 3 The average error was 18.83 μg / m 3 The error of the LSTM deep neural network model at the corresponding monitoring points was 12 μg / m. 3 11μg / m 3 11μg / m 3 13μg / m 3 11μg / m 3 and 12μg / m 3 The total error is 70 μg / m 3 The average error was 11.67 μg / m 3 The method of this invention has an error of 3 μg / m at monitoring points 1 to 6. 3 4μg / m 3 3μg / m 3 3μg / m 3 2μg / m 3 and 3μg / m 3 The total error is 18 μg / m 3 The average error is 3.00 μg / m 3 The comparison shows that the prediction results of the method of this invention are closest to the measured pollution concentrations at all monitoring points, with an average error reduction of approximately 84.06% compared to the Gaussian diffusion model and approximately 74.29% compared to the LSTM deep neural network model. These results demonstrate that the joint modeling approach of topological manifold spatiotemporal spectral analysis and the improved DeepONet model can significantly improve the accuracy of pollution concentration prediction in complex urban spatial environments.

[0036] As can be seen from this embodiment, in the process of atmospheric pollution diffusion modeling, this invention constructs a three-dimensional voxel space model to refine the expression of urban building spatial structure, enabling the pollution diffusion process to be described in a three-dimensional spatial environment. This overcomes the problem that traditional two-dimensional diffusion models cannot reflect the influence of buildings on pollution diffusion paths. Simultaneously, by generating pollution information units and discretizing them in the three-dimensional voxel space model, this invention establishes a unified data expression structure between pollution source emissions and spatial diffusion processes, improving the stability of the pollution diffusion calculation process. In the pollution diffusion path identification process, this invention introduces a topological manifold spatiotemporal spectral decomposition method to perform spatiotemporal structural analysis on the positional changes of pollution information units in continuous time steps, thereby extracting the dominant propagation direction of pollution diffusion and improving the accuracy of pollution diffusion trajectory identification. In the pollution diffusion prediction stage, by improving the DeepONet model to construct a pollution diffusion operator matrix, and performing spectral space decomposition at the spectral decomposition operator layer and introducing a spectral manifold folding mechanism, the model can reduce the noise impact of local random diffusion while maintaining the main propagation mode of pollution diffusion, thereby improving the stability of the pollution diffusion prediction results. This verifies that this invention can achieve high-precision modeling and stable prediction of pollution diffusion processes in complex urban spatial environments, demonstrating good practical application value.

[0037] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A deep learning-based method for modeling atmospheric pollution diffusion, characterized in that, Includes the following steps: Step 1: Obtain pollution source emission data and meteorological data for the target area over a continuous time period, and divide the target area into grids to form a two-dimensional spatial grid; Step 2: Obtain building height data for the target area, construct a three-dimensional spatial structure, and divide the three-dimensional spatial structure into voxels to form a three-dimensional voxel space model; Step 3: Generate pollution information units based on the pollution source emission data, and record the initial spatial location, pollutant type, pollution intensity and generation time for each pollution information unit, and map the pollution information units to the three-dimensional voxel space model; Step 4: Count the number of pollution information units and calculate the voxel pollution intensity in the three-dimensional voxel space model, and construct a three-dimensional pollution diffusion state field by combining meteorological data; Step 5: Perform topological manifold spatiotemporal resolution on the positional changes of pollution information units in the three-dimensional voxel space model in continuous time steps to form a set of pollution diffusion trajectories; Step 6: Input the three-dimensional pollution diffusion state field into the state branch of the improved DeepONet model, input the pollution diffusion trajectory set into the trajectory branch of the improved DeepONet model, construct the pollution diffusion operator matrix through the operator projection layer, perform spectral space decomposition on the pollution diffusion operator matrix, and introduce the spectral manifold folding mechanism to obtain the three-dimensional pollution diffusion state of the next time step. Step 7: Update the spatial position of the pollution information unit in the three-dimensional voxel space model according to the three-dimensional pollution diffusion state of the next time step to obtain the updated pollution information unit; Step 8: Map the updated pollution information units in the three-dimensional voxel space model to a two-dimensional spatial grid to generate the spatial distribution of pollution concentration in the target area.

2. The deep learning-based atmospheric pollution diffusion modeling method according to claim 1, characterized in that, Step one specifically involves: Acquire pollution source emission data for a target area over a continuous time period, including pollution source location, pollutant type, emission amount, and emission time; Acquire meteorological data over a continuous time period, including wind speed, wind direction, and temperature; The region boundary is determined based on the spatial extent of the target area, and the region boundary is divided according to the set spatial resolution to generate multiple grid cells; Each grid cell is assigned a unique grid number, and the spatial extent of the corresponding grid cell is recorded to form a two-dimensional spatial grid.

3. The deep learning-based atmospheric pollution diffusion modeling method according to claim 1, characterized in that, Step two specifically involves: Acquire building height data for the target area and record the spatial coordinates of each building; Based on the building height data, a building height volume is generated vertically at the spatial location coordinates to construct the three-dimensional spatial structure of the target area; The three-dimensional spatial structure is regularly divided according to the set voxel side length to generate multiple voxel units; Determine whether the spatial position of each voxel unit falls within the building height volume range, and mark the voxel units that fall within the building height volume range as building voxels, and mark the voxel units that do not fall within the building height volume range as spatial voxels; Record the three-dimensional coordinates and voxel type of each voxel unit to form a three-dimensional voxel space model; During the pollution diffusion calculation process, when a pollution information unit migrates to a target voxel unit, the voxel type of the target voxel unit is determined. If the target voxel unit is a building voxel, the migration of the pollution information unit is prevented; if the target voxel unit is a space voxel, the migration of the pollution information unit is allowed.

4. The deep learning-based atmospheric pollution diffusion modeling method according to claim 1, characterized in that, Step three specifically involves: Extract the location of the pollution source, the type of pollutant, the amount of emissions, and the time of emissions from the pollution source emission data; The emissions from pollution sources are divided into time segments according to a set time interval to obtain the emissions corresponding to each time step; The emissions at each time step are divided according to the set pollution intensity baseline value. The number of corresponding pollution information units is determined and multiple pollution information units are generated. A unique unit number is assigned to each pollution information unit and the pollutant type is recorded. The pollution intensity of a single pollution information unit is the ratio of the emissions at the corresponding time step to the number of pollution information units. The location of the pollution source is recorded as the initial spatial location of the pollution information unit, and the time of the corresponding time step is recorded as the generation time of the pollution information unit. Based on the initial spatial position of each pollution information unit, the corresponding voxel unit is determined in the three-dimensional voxel space model, and each pollution information unit is mapped to the corresponding voxel unit.

5. The deep learning-based atmospheric pollution diffusion modeling method according to claim 1, characterized in that, Step four specifically involves: In the three-dimensional voxel space model, each voxel unit is traversed, and the number of contamination information units in each voxel unit is counted. The contamination intensity of contamination information units within the same voxel unit is accumulated to obtain the voxel contamination intensity of the corresponding voxel unit. Acquire meteorological data corresponding to the current time step, and map the meteorological data to the spatial location of each voxel unit; By combining the voxel pollution intensity and the corresponding meteorological data according to the spatial arrangement of voxel units, a three-dimensional pollution diffusion state field is formed.

6. The deep learning-based atmospheric pollution diffusion modeling method according to claim 1, characterized in that, Step five specifically involves: The voxel coordinates of each pollution information unit in the three-dimensional voxel space model are recorded in continuous time steps. The voxel coordinates include the row index, column index and height index of the voxel in the three-dimensional voxel space model, and the position sequence of the pollution information units is formed in time order. A time connection edge is established based on the voxel coordinate change of the pollution information unit between adjacent time steps. When the voxel coordinate change of the pollution information unit in adjacent time steps does not exceed the set voxel adjacency distance, a time connection edge is established between the corresponding pollution information units. Calculate the spatial distance between the voxel units where each pollution information unit is located within the same time step. When the spatial distance is less than the set spatial adjacency distance, establish a spatial connection edge between the corresponding pollution information units. A spatiotemporal adjacency matrix of pollution information units is constructed based on the time connection edge and the spatial connection edge, wherein the pollution information unit number is used as the matrix row and column index. When there is a time connection edge or a spatial connection edge between two pollution information units, the corresponding matrix element is assigned a value of 1, otherwise it is assigned a value of 0. The corresponding Laplacian matrix is ​​calculated based on the spatiotemporal adjacency matrix, and the Laplacian matrix is ​​decomposed into eigenvalues ​​to obtain the corresponding eigenvalues ​​and eigenvectors. The eigenvectors are then sorted according to the size of the eigenvalues. The first few feature vectors after sorting are selected as the main diffusion spectrum vectors, and the main diffusion direction of the pollution information unit in the three-dimensional voxel space is determined according to the main diffusion spectrum vectors. Based on the main diffusion direction, the voxel coordinate sequence of the corresponding pollution information unit is connected in chronological order to generate a pollution diffusion trajectory, forming a pollution diffusion trajectory set.

7. The deep learning-based atmospheric pollution diffusion modeling method according to claim 1, characterized in that, Step six specifically involves: The improved DeepONet model includes state branching, trajectory branching, operator projection layer, and spectral decomposition operator layer, and embeds a spectral manifold folding mechanism in the spectral decomposition operator layer; The three-dimensional pollution diffusion state field is arranged in the order of row index, column index and height index of voxels in the three-dimensional voxel space model to form a voxel sequence, and then input into the state branch. The state branch includes a three-layer fully connected network to obtain the pollution diffusion state feature vector. The set of pollution diffusion trajectories is input into the trajectory branch, which includes a three-layer fully connected network, to obtain the pollution diffusion trajectory feature vector. The pollution diffusion state feature vector and the pollution diffusion trajectory feature vector are input into the operator projection layer and spliced ​​to form a joint feature vector. The i-th feature element and the j-th feature element in the joint feature vector are used as the matrix row and column indices. The absolute value of the difference between the feature element values ​​is calculated and written into the matrix element in the i-th row and j-th column to obtain the diffusion energy matrix. Summing the elements of each row of the diffusion energy matrix yields the energy sum of the corresponding row. Then, each element in the row is divided by the corresponding energy sum to obtain the operator projection matrix. A pollution diffusion operator matrix is ​​generated by linearly mapping the joint eigenvectors using the operator projection matrix. The pollution diffusion operator matrix is ​​input into the spectral decomposition operator layer for spectral space decomposition. The degree matrix is ​​calculated based on the pollution diffusion operator matrix, and a normalized diffusion Laplace operator is constructed using the degree matrix and the pollution diffusion operator matrix. The normalized diffusion Laplace operator is decomposed into eigenvalues ​​to obtain an eigenvalue sequence and an eigenvector matrix. Several eigenvectors are selected in ascending order of eigenvalues ​​to form a spectral space feature matrix. The spectral decomposition operator layer performs a spectral manifold folding mechanism on the spectral space feature matrix, calculates the Euclidean distance between any two feature vectors in the spectral space feature matrix, divides two feature vectors whose Euclidean distance is less than a set distance threshold into the same spectral neighborhood, performs a weighted summation of the feature vectors in each spectral neighborhood using the reciprocal of the corresponding eigenvalue as the weight, generates the corresponding folded feature vector, and replaces the feature vectors in the original spectral neighborhood with the folded feature vectors corresponding to each spectral neighborhood to obtain the folded spectral space feature matrix; The folded spectral space feature matrix is ​​multiplied with its transpose to generate an updated pollution diffusion operator matrix. The updated pollution diffusion operator matrix is ​​then used to calculate the voxel pollution intensity of each voxel unit in the three-dimensional voxel space model at the next time step, thus obtaining the three-dimensional pollution diffusion state at the next time step.

8. The deep learning-based atmospheric pollution diffusion modeling method according to claim 1, characterized in that, Step seven specifically involves: Based on the three-dimensional contamination diffusion state at the next time step, read the voxel contamination intensity of each voxel unit in the three-dimensional voxel space model. For each pollution information unit, obtain the voxel coordinates of the current voxel unit and determine the set of voxel units adjacent to the current voxel unit; Compare the voxel contamination intensity of each voxel unit in adjacent voxel unit sets, and determine the voxel unit with the highest voxel contamination intensity as the target voxel unit. When a contamination information unit migrates to a target voxel unit, the migration judgment is made based on the voxel type of the target voxel unit. When the target voxel unit allows the contamination information unit to migrate in, the spatial position of the contamination information unit is updated from the voxel coordinates corresponding to the current voxel unit to the voxel coordinates corresponding to the target voxel unit, and the updated voxel coordinates are recorded as the spatial position of the contamination information unit at the current time step, thus obtaining the updated contamination information unit.

9. The deep learning-based atmospheric pollution diffusion modeling method according to claim 1, characterized in that, Step eight specifically involves: The corresponding voxel unit is determined based on the voxel coordinates of the updated pollution information unit in the three-dimensional voxel space model, and the number of pollution information units in each voxel unit is counted. The contamination intensity of contamination information units within the same voxel unit is accumulated to obtain the voxel contamination intensity of the corresponding voxel unit. The voxel cells with the same row index and column index in the three-dimensional voxel space model are traversed according to the height index, and the voxel contamination intensity of the corresponding voxel cells is accumulated to obtain the aggregate contamination intensity at the corresponding row index and column index position. The aggregated pollution intensity is mapped to the corresponding two-dimensional spatial grid cell, and the spatial distribution of pollution concentration in the target area is generated according to the row and column arrangement of the two-dimensional spatial grid.

10. A deep learning-based atmospheric pollution diffusion modeling system, executing the deep learning-based atmospheric pollution diffusion modeling method according to any one of claims 1 to 9, characterized in that, Includes the following modules: The data acquisition module is used to acquire pollution source emission data and meteorological data of the target area over a continuous time period, and to divide the target area into a grid to form a two-dimensional spatial grid. The 3D space construction module is used to acquire building height data of the target area, construct a 3D space structure, and divide the 3D space structure into voxels to form a 3D voxel space model. The pollution information unit generation module is used to generate pollution information units based on the pollution source emission data, record the initial spatial location, pollutant type, pollution intensity and generation time for each pollution information unit, and map the pollution information unit to a three-dimensional voxel space model. The diffusion state construction module is used to count the number of pollution information units and calculate the voxel pollution intensity in the three-dimensional voxel space model, and to construct a three-dimensional pollution diffusion state field in combination with meteorological data. The diffusion trajectory extraction module is used to perform topological manifold spatiotemporal spectral analysis on the positional changes of pollution information units in the three-dimensional voxel space model in continuous time steps to form a set of pollution diffusion trajectories. The diffusion prediction module is used to input the three-dimensional pollution diffusion state field into the state branch of the improved DeepONet model, input the pollution diffusion trajectory set into the trajectory branch of the improved DeepONet model, construct the pollution diffusion operator matrix through the operator projection layer, perform spectral space decomposition on the pollution diffusion operator matrix and introduce a spectral manifold folding mechanism to obtain the three-dimensional pollution diffusion state of the next time step. The pollution unit update module is used to update the spatial position of the pollution information unit in the three-dimensional voxel space model according to the three-dimensional pollution diffusion state of the next time step, so as to obtain the updated pollution information unit. The pollution concentration generation module is used to map the distribution of updated pollution information units in the three-dimensional voxel space model to a two-dimensional spatial grid, thereby generating the spatial distribution of pollution concentration in the target area.

Images