Backward trajectory analysis method and apparatus for atmospheric pollutant transport pathway identification

Abstract

The application provides a backward trajectory analysis method and device for atmospheric pollutant transport path identification, and belongs to the field of pollution source tracking. The method comprises the following steps: setting trajectory simulation parameters, generating a plurality of backward trajectories through a trajectory simulation model; constructing a trajectory dataset comprising each coded backward trajectory; automatically or manually selecting a clustering strategy from a plurality of preset trajectory clustering algorithms based on the dimension, time sequence characteristics, community structure complexity and nonlinear path characteristics of each coded backward trajectory in the trajectory dataset; applying the selected clustering strategy to the trajectory dataset for classification, thereby generating a plurality of trajectory categories; and performing classification processing on each trajectory category, and outputting the processing result after the classification processing. The application realizes flexible classification of trajectory data with different structures, supports abnormal trajectory identification, trajectory dimension reduction mapping and unified format output.

Description

Technical Field

[0001] This invention belongs to the field of pollution source tracing, specifically relating to a backward trajectory analysis method and apparatus for identifying the transport paths of air pollutants. Background Technology

[0002] Backward trajectory analysis is a commonly used and crucial meteorological data analysis method in environmental science research and practical monitoring, including air pollution source tracing, pollutant transport identification, and air quality management. This method simulates the movement trajectory of air masses or pollutants over time using meteorological dynamic models to determine the possible source area or transport path of pollutants at a given ground location. Its core lies in using meteorological data (such as wind field, temperature, and humidity) to deduce the movement trajectory of air masses in a specific time and space, thus providing a scientific basis for identifying pollution sources and transport paths.

[0003] Backward trajectory analysis has been widely used in various environmental science scenarios. However, existing backward trajectory analysis techniques face several challenges in practical applications:

[0004] 1. Trajectory data complexity: Data scenarios with complex trajectory structures, inconsistent lengths, or outliers are difficult to process effectively, especially when the number of trajectories is large, which limits the efficiency and accuracy of analysis.

[0005] 2. Limitations of Clustering Methods: Existing clustering methods are inefficient, easily affected by initial conditions, and struggle to adapt to the temporal characteristics, high dimensionality, or nonlinear path features of trajectory data. Furthermore, the number of clusters often needs to be manually specified, lacking an adaptive judgment mechanism based on the data's inherent characteristics.

[0006] 3. Output compatibility issues: Different clustering methods or tools have significantly different output structures, making it difficult to directly interface with Geographic Information System (GIS) platforms or conduct unified statistical analysis, thus limiting the practicality and comparability of the results.

[0007] To address the aforementioned problems, there is an urgent need for a backward trajectory analysis method and apparatus that supports multi-strategy trajectory clustering, can automatically adapt to trajectory data characteristics (such as high dimensionality, time series characteristics, community structure complexity, and nonlinear path characteristics), and outputs a unified classification structure. This invention introduces an integrated method based on autoencoders and recurrent neural networks, along with a graph-based community detection method, combined with an automated clustering strategy selection mechanism. The aim is to improve the intelligence level of pollution source tracing and the accuracy and efficiency of regional transmission research. Summary of the Invention

[0008] In view of this, the present invention provides a backward trajectory analysis method and apparatus for identifying atmospheric pollutant transport paths, which realizes flexible classification of trajectory data with different structures and supports abnormal trajectory identification, trajectory dimensionality reduction mapping and unified format output.

[0009] To achieve the above objectives, the present invention is implemented as follows:

[0010] This invention discloses a backward trajectory analysis method for identifying atmospheric pollutant transport paths, comprising the following steps:

[0011] Step 1: Set the trajectory simulation parameters; the trajectory simulation parameters include the simulation start position, start time, simulation height, tracing duration, and the time interval between trajectory points of the backward trajectory; generate multiple backward trajectories through the trajectory simulation model;

[0012] Step 2: Number and structure-encode each backward trajectory and its trajectory points to form the coded backward trajectory corresponding to each backward trajectory, and construct a trajectory dataset including each coded backward trajectory;

[0013] Step 3: Based on the dimension, time series characteristics, community structure complexity, and nonlinear path characteristics of each coded backward trajectory in the trajectory dataset, automatically or manually select a clustering strategy from a variety of preset trajectory clustering algorithms;

[0014] Step 4: Apply the selected clustering strategy to the trajectory dataset to classify it and generate multiple trajectory categories; the clustering strategy includes an ensemble method based on autoencoders and recurrent neural networks, as well as a graph-based clustering method;

[0015] Step 5: Classify each trajectory category and output the classification results.

[0016] Specifically, in step 1, the step of generating multiple backward trajectories through a trajectory simulation model is a HYSPLIT model, using GDAS meteorological data; the simulation altitude is set to 100 meters, the tracing time is set to 48 hours, and the time interval between trajectory points of the backward trajectories is set to 1 hour.

[0017] Specifically, when the dimension of each encoded backward trajectory in the trajectory dataset exceeds the first preset threshold, the trajectory dataset is determined to have high dimensionality.

[0018] The time dependence of each encoded backward trajectory is analyzed. When the time dependence of each encoded backward trajectory in the trajectory dataset exceeds the second preset threshold, it is determined that the trajectory dataset has strong time series characteristics. Furthermore, when the trajectory dataset has both high dimensionality and strong time series characteristics, the selected clustering strategy is an integrated method based on autoencoders and recurrent neural networks.

[0019] Specifically, a graph structure is constructed based on the trajectory points of the backward trajectories corresponding to each encoded backward trajectory, and community structures are detected. When the complexity of the existing community structure exceeds a third preset threshold, it is determined that the trajectory dataset has a complex community structure, and the selected clustering strategy is a graph-based clustering method. In this method, the trajectory points of the backward trajectory are used as nodes in the graph structure, and edges are established between nodes that have spatial relationships or similarity between backward trajectories. The similarity of the backward trajectories corresponding to each encoded backward trajectory is measured. When the similarity measurement results indicate that the trajectory dataset has non-linear path characteristics, the selected clustering strategy is a graph-based clustering method.

[0020] Specifically, the integrated method based on autoencoders and recurrent neural networks includes:

[0021] An autoencoder is used to reduce the dimensionality and extract features from each encoded backward trajectory, generating a low-dimensional latent representation for each encoded backward trajectory. A recurrent neural network is used to process the low-dimensional latent representation to obtain the temporal dynamic features corresponding to each encoded backward trajectory, generating a feature representation for each encoded backward trajectory. The feature representation is clustered to generate multiple trajectory categories. The autoencoder includes an encoder and a decoder, wherein the encoder maps each encoded backward trajectory to the latent space, and the decoder reconstructs the low-dimensional latent representation back to the input space and is trained by minimizing the reconstruction error.

[0022] Specifically, the recurrent neural network is a long short-term memory network, configured to process the sequence of the low-dimensional latent representations and generate feature representations.

[0023] Specifically, the graph-based clustering method includes: using a community detection algorithm to identify community structures in the graph structure; and grouping coded backward trajectories belonging to the same community into the same trajectory category.

[0024] Specifically, the community detection algorithm is the Louvain algorithm, which divides the community structure by maximizing the modularity;

[0025] The edge weights in the graph structure are based on a similarity metric between encoded backward trajectories, wherein the similarity metric includes dynamic time warping distance.

[0026] Specifically, the processing results after classification in step 5 include: trajectory category labels, representative trajectory paths, number of trajectory members, monthly distribution ratio, and trajectory classification layer files applicable to geographic information systems; the representative trajectory path is the average trajectory calculated based on each backward trajectory corresponding to the trajectory category, and the latitude and longitude coordinates of the trajectory points in the average trajectory are the arithmetic mean of the latitude and longitude coordinates of the corresponding trajectory points in each backward trajectory corresponding to the trajectory category.

[0027] This invention also discloses a backward trajectory analysis device for identifying atmospheric pollutant transport paths, specifically comprising:

[0028] The backward trajectory generation module is used to set trajectory simulation parameters. These parameters include the simulation start position, start time, simulation height, tracing duration, and the time interval between trajectory points in the backward trajectory. Multiple backward trajectories are generated through the trajectory simulation model.

[0029] The trajectory dataset construction module is used to number and structure each backward trajectory and its trajectory points, forming a coded backward trajectory corresponding to each backward trajectory, and constructing a trajectory dataset including each coded backward trajectory;

[0030] The clustering strategy selection module is used to automatically or manually select a clustering strategy from a variety of preset trajectory clustering algorithms based on the dimension, time series characteristics, community structure complexity and nonlinear path characteristics of each encoded backward trajectory in the trajectory dataset.

[0031] The trajectory category generation module is used to classify the trajectory dataset by applying a selected clustering strategy to generate multiple trajectory categories; the clustering strategy includes an ensemble method based on autoencoders and recurrent neural networks, as well as a graph-based clustering method;

[0032] The results output module is used to classify each trajectory category and output the processed results.

[0033] Beneficial effects:

[0034] The technical solution of this invention, based on the integrated method of autoencoder and recurrent neural network, can effectively process high-dimensional and time-series complex trajectory data. The autoencoder performs dimensionality reduction and feature extraction to generate low-dimensional latent representations, and then the recurrent neural network is used to capture time dynamic features. Finally, the trajectory categories are generated through clustering technology, thereby improving the accuracy of trajectory classification and adaptability to complex data structures, and providing more accurate data support for the identification of atmospheric pollutant transport paths.

[0035] Through the technical solution of this invention, the autoencoder maps trajectory data to a latent space through the encoder and reconstructs it through the decoder. By adopting a training method that minimizes the reconstruction error, the key features of the trajectory data can be effectively preserved, the computational complexity can be reduced, and the dimensionality reduction data can still reflect the spatial and temporal characteristics of the original trajectory. This improves the processing efficiency and classification quality of backward trajectory analysis in high-dimensional data scenarios.

[0036] The technical solution of this invention uses a Long Short-Term Memory (LSTM) network as a recurrent neural network to process the potential representation sequence output by the autoencoder. This effectively captures the temporal dependence of trajectory data and generates a representation with temporal characteristics, thereby improving the classification accuracy of dynamically changing trajectories. It is particularly suitable for analyzing the transmission patterns of pollutants over time, enhancing the reliability and depth of pollution source tracing.

[0037] By analyzing the dimensionality and time dependence of trajectory datasets, the high-dimensionality and time-series characteristics are automatically evaluated. When a preset threshold is met, an integrated method based on autoencoders and recurrent neural networks is selected. This enables adaptive processing of high-dimensional and time-series complex trajectory data, thereby improving the intelligence level of clustering strategy selection, reducing human intervention, and enhancing analysis efficiency and result stability.

[0038] By constructing a similarity graph structure for trajectory data and detecting community structures to assess the complexity of community structures, and by calculating similarity metrics between trajectories to assess nonlinear path features, the method of graph-based community detection can be preferentially selected in complex community structures or nonlinear path scenarios, thereby improving the classification ability of complex transmission networks and nonlinear trajectories and enhancing the robustness and accuracy of contaminated transmission path identification.

[0039] Through the technical solution of this invention, the trajectory classification results are output in a standard structure, including trajectory category labels, representative paths, number of trajectories, time distribution ratio, and GIS layer compatible fields, which support subsequent geographic information system loading, statistical analysis, and graphic visualization, thereby improving the practicality of the method in environmental management scenarios such as pollution inversion and regional joint prevention and control. Attached Figure Description

[0040] Figure 1 This is a schematic diagram of a backward trajectory analysis method for identifying atmospheric pollutant transport paths.

[0041] Figure 2 This is a schematic diagram of a backward trajectory analysis device used for identifying the transport paths of air pollutants. Detailed Implementation

[0042] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0043] This invention provides a backward trajectory analysis method for identifying atmospheric pollutant transport paths, such as... Figure 1 As shown, it includes the following steps:

[0044] Step 1: Set the trajectory simulation parameters; the trajectory simulation parameters include the simulation start position, start time, simulation height, tracing duration, and the time interval between trajectory points of the backward trajectory; generate multiple backward trajectories through the trajectory simulation model.

[0045] In this invention, the trajectory simulation is performed using the HYSPLIT (HYbrid Single-Particle Lagrangian Integrated Trajectory) model, which was developed by the National Oceanic and Atmospheric Administration (NOAA) and is widely used in air quality retrieval and transport path tracking.

[0046] (1.1) The parameters to be set include:

[0047] Simulated starting location: such as the latitude and longitude of the city center of XX;

[0048] Start time: such as 08:00 UTC daily, or set the interval to simulate once every 6 hours;

[0049] Simulation altitude: set to 100 meters to simulate the behavior of near-ground pollutants;

[0050] Tracing duration: 48 hours; the tracing trajectory reflects the possible range of the source.

[0051] The time interval between trajectory points in the backward trajectory is set to 1 hour, generating 49 time nodes;

[0052] Meteorological data source: Meteorological data is used from GDAS (Global Data Assimilation System).

[0053] (1.2) Trajectory generation method:

[0054] Users can generate backward trajectories over a period of time by calling the above parameters through a web interface or by deploying the HYSPLIT model locally. The output backward trajectories are in text or Shapefile format, and each backward trajectory contains multiple trajectory points with information such as longitude, latitude, altitude, time, and wind speed.

[0055] Step 2: Number and structure-encode each backward trajectory and its trajectory points to form the coded backward trajectory corresponding to each backward trajectory, and construct a trajectory dataset including each coded backward trajectory.

[0056] To ensure the computability and traceability of subsequent analysis of backward trajectories, this invention assigns unified numbering and structured encoding to all backward trajectories, forming a standard trajectory dataset.

[0057] (2.1) Track numbering method:

[0058] Each trajectory can be numbered according to the following formula (categorized by year, month, and ten-day period):

[0059] LN = M × 10⁴ + D × 10² + N;

[0060] in:

[0061] LN represents the trajectory number;

[0062] M represents the month (e.g., January is 01);

[0063] D indicates the first, middle, and last ten days of the month (January, February, and March).

[0064] N represents the trajectory number within that ten-day period.

[0065] (2.2) Trajectory point numbering:

[0066] Each trajectory consists of num trajectory points, numbered PN=1~num, used to construct a two-dimensional trajectory vector.

[0067] (2.3) Trajectory dataset structure:

[0068] Construct a trajectory dataset table, with each row recording one trajectory node. The fields included in the trajectory dataset table are shown in Table 1 below.

[0069] Table 1. Fields of the Trajectory Dataset Table

[0070] Fields describe Backward trajectory number (LN) Unique track identifier Node Number (PN) Time Node Index Longitude (λ) Longitude of nodes Latitude (φ) Node latitude Timestamp UTC time Meteorological variables (optional) Wind speed, temperature, etc.

[0071] In this invention, all backward trajectories are sorted in ascending order by number and can be exported as .txt or .csv files for clustering input; they can also be converted into spatial vector layer (Shapefile) for visualization pre-detection.

[0072] Step 3: Based on the dimension, time series characteristics, community structure complexity, and nonlinear path characteristics of each coded backward trajectory in the trajectory dataset, automatically or manually select a clustering strategy from a variety of preset trajectory clustering algorithms;

[0073] Design the following algorithm based on the trajectory dataset generated in step 2. Each encoded backward trajectory It contains spatial coordinates and time series information, where t=1,…,n; n is the total number of backward trajectories.

[0074] Specifically, when the dimension of each encoded backward trajectory in the trajectory dataset exceeds the first preset threshold, the trajectory dataset is determined to have high dimensionality.

[0075] The formula for calculating dimension is:

[0076] ;

[0077] If dimension Exceeding the first preset threshold (For example, 98-dimensional), then the encoded backward trajectory is considered to have high dimensionality.

[0078] When the time dependency of each encoded backward trajectory in the trajectory dataset exceeds a second preset threshold, the trajectory dataset is determined to have strong time series characteristics. The formula for calculating the time dependency is as follows:

[0079] ;

[0080] in, Due to time lag, The specific value ranges from 1 to 48 (in hours). For example: if A value of 1 indicates that data points in the trajectory are compared at intervals of 1 hour; if If it is 12, it means comparing data points in the trajectory that are 12 hours apart; if If the value is 48, it indicates that the data of the starting point and the ending point in the trajectory are compared. This is the mean of the encoded backward trajectory. If time-dependent... (For example, 0.5), that is, when the time dependence of the encoded backward trajectory exceeds the second preset threshold, the encoded backward trajectory is considered to have strong time series characteristics.

[0081] In this invention, a graph structure is constructed, and the modularity is calculated to assess the complexity of the community structure. The formula for calculating the modularity is:

[0082]

[0083] in, Let represent the total edge weights of the graph structure, and let i and j be the nodes of the graph structure, respectively. Let be the edge weight of the edge between graph nodes i and j. , Let i and j be the node degrees of graph nodes i and j, respectively. For indicator functions, These represent the communities to which graph node i and graph node j belong, respectively. If the two nodes belong to the same community, then the indicator function... Returns 1 otherwise, returns 0. If the module degree... (For example, 0.4), that is, if the complexity of the community structure exceeds the third preset threshold, then the trajectory dataset is considered to have a complex community structure. This is the third preset threshold.

[0084] A similarity measure is performed on the backward trajectories corresponding to each encoded backward trajectory. When the similarity measure results indicate that the trajectory dataset has non-linear path characteristics, the selected clustering strategy is a graph-based clustering method, where:

[0085] Calculate the encoded backward trajectory Dynamic Time Warping (DTW) Distance :

[0086] ;

[0087] in, =1,…,n; =1,…,n; Encoding the backward trajectory At time step Location, Encoding the backward trajectory At time step Location, For time alignment paths, which are a sequence of indices, it is defined how each point of one time series is matched with one or more points of another time series to minimize the cumulative distance between them. Indicates the time alignment path Guided by the trajectory The Each time step corresponds to Which time step in the time step? T represents the total time step, and the trajectory node interval is set to 1 hour; if... ,Right now If the value exceeds the fourth preset threshold, the trajectory dataset is considered to have nonlinear characteristics.

[0088] Clustering strategy selection is based on the above characteristics, and a clustering strategy is automatically selected:

[0089] (3.1) Decision-making rules:

[0090] When the trajectory dataset has both high dimensionality and strong time series characteristics, the selected clustering strategy is an integrated method based on autoencoders and recurrent neural networks.

[0091] When the complexity of the community structure exceeds the third preset threshold, the trajectory dataset is determined to have a complex community structure, and the selected clustering strategy is a graph-based clustering method; when the similarity measurement results indicate that the trajectory dataset has non-linear path characteristics, the selected clustering strategy is a graph-based clustering method.

[0092] (3.2) Validation and optimization: Validate the classification results of the selected clustering strategy, for example by comparing the consistency or stability of trajectory categories of different algorithms, and optimize the final selection.

[0093] (3.3) Applicable Scenarios and Selection Mechanism

[0094] Manual selection: Users can specify algorithms based on their analysis needs (e.g., emphasizing visualization), such as choosing an ensemble method based on autoencoders and recurrent neural networks to generate low-dimensional visualization results.

[0095] Automatic selection: The optimal algorithm is automatically selected based on the characteristics of the trajectory dataset (dimensionality, time series complexity, distribution characteristics). For example, autoencoders and LSTMs are preferred for high-dimensional data, and graph community detection methods are preferred for complex networks.

[0096] Visualization support: When it is necessary to map high-dimensional trajectory data to a low-dimensional space for visualization analysis, the ensemble method based on autoencoders and recurrent neural networks is preferred to generate low-dimensional feature representations.

[0097] Step 4: Apply clustering strategies to the trajectory dataset to classify it and generate multiple trajectory categories; the clustering strategies include an ensemble method based on autoencoders and recurrent neural networks, as well as a graph-based clustering method.

[0098] In this step, the integrated method based on autoencoders and recurrent neural networks specifically includes:

[0099] This method combines an autoencoder (AE) and a long short-term memory network (LSTM), making it suitable for processing high-dimensional and time-series complex trajectory data.

[0100] An autoencoder (AE) consists of an encoder and a decoder. Trajectory data is first flattened into fixed-dimensional vectors (e.g., serializing the latitude and longitude coordinates of each trajectory into a 98-dimensional vector), and then mapped by the encoder to a low-dimensional latent space to generate a low-dimensional latent representation. The decoder will Reconstruct back into the input space and generate reconstructed data. The autoencoder is trained by minimizing the reconstruction error.

[0101] Encoder formula:

[0102] ;

[0103] Decoder formula:

[0104] ;

[0105] Loss function:

[0106] ;

[0107] in, for The low-dimensional potential representation, for Reconstructed data, and These are the weights for the encoder and decoder, respectively. and For bias, and It is an activation function (such as ReLU or Sigmoid).

[0108] Long Short-Term Memory (LSTM) networks process the latent representation sequence output by the autoencoder, capture temporal dependencies, and generate feature representations.

[0109] The K-means algorithm is applied to the feature representation generated by the LSTM, iteratively assigning trajectories to the nearest cluster centers and updating the centers until the clusters converge, thus generating trajectory categories. This strategy is suitable for high-dimensional trajectory data (e.g., 98-dimensional vectors) generated in step 2 or scenarios with significant time-series characteristics (e.g., dynamic transport patterns of pollutants over time). The generated low-dimensional feature representation can be mapped to a two-dimensional space, facilitating visualization and analysis.

[0110] In this invention, the graph-based clustering method includes: using a community detection algorithm to identify community structures in a graph structure; and grouping backward trajectories belonging to the same community into the same trajectory category.

[0111] The formula for calculating the edge weights in the graph structure is as follows:

[0112] ;

[0113] in, Let i be the edge weight between nodes i and j in the graph structure. Let be the distance between the backward trajectories belonging to nodes i and j in the graph structure. This is the scaling parameter.

[0114] The community detection algorithm is the Louvain algorithm, which divides the community structure by maximizing the modularity.

[0115] This method can identify outlier trajectories generated by meteorological anomalies in step 1, label them as noise, and enhance the robustness of clustering. This strategy is applicable to trajectory data exhibiting complex community structures or nonlinear paths in step 2, such as the identification of cross-regional pollution transmission channels.

[0116] Step 5: Classify each trajectory category and output the classification results.

[0117] The processing results after classification in step 5 include: trajectory category labels, representative trajectory paths, number of trajectory members, monthly distribution ratio, and trajectory classification layer files applicable to geographic information systems; the representative trajectory path is the average trajectory calculated based on each backward trajectory corresponding to the trajectory category, and the latitude and longitude coordinates of the trajectory points in the average trajectory are the arithmetic mean of the latitude and longitude coordinates of the corresponding trajectory points in each backward trajectory corresponding to the trajectory category.

[0118] This invention also discloses a backward trajectory analysis device for identifying atmospheric pollutant transport paths, such as... Figure 2 As shown, it specifically includes:

[0119] The backward trajectory generation module is used to set trajectory simulation parameters. These parameters include the simulation start position, start time, simulation height, tracing duration, and the time interval between trajectory points in the backward trajectory. Multiple backward trajectories are generated through the trajectory simulation model.

[0120] The trajectory dataset construction module is used to number and structure each backward trajectory and its trajectory points, forming a coded backward trajectory corresponding to each backward trajectory, and constructing a trajectory dataset including each coded backward trajectory;

[0121] The clustering strategy selection module is used to automatically or manually select a clustering strategy from a variety of preset trajectory clustering algorithms based on the dimension, time series characteristics, community structure complexity and nonlinear path characteristics of each encoded backward trajectory in the trajectory dataset.

[0122] The trajectory category generation module is used to classify the trajectory dataset by applying a selected clustering strategy to generate multiple trajectory categories; the clustering strategy includes an ensemble method based on autoencoders and recurrent neural networks, as well as a graph-based clustering method;

[0123] The results output module is used to classify each trajectory category and output the processed results.

[0124] It will be apparent to those skilled in the art that the embodiments of the present invention are not limited to the details of the exemplary embodiments described above, and that the embodiments of the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the embodiments of the present invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the embodiments of the present invention is defined by the appended claims rather than the foregoing description. Therefore, all variations falling within the meaning and scope of equivalents of the claims are intended to be encompassed within the embodiments of the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims. Furthermore, it is clear that the word "comprising" does not exclude other units or steps, and the singular does not exclude the plural. Multiple units, modules, or devices stated in a system, apparatus, or terminal may also be implemented by the same unit, module, or device through software or hardware. The terms "first," "second," etc., are used to indicate names and do not indicate any particular order.

[0125] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the embodiments of the present invention and are not intended to limit them. Although the embodiments of the present invention have been described in detail with reference to the above preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions to the technical solutions of the embodiments of the present invention should not depart from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A backward trajectory analysis method for atmospheric pollutant transport pathway identification, characterized in that, The method comprises the following steps: Step 1, setting trajectory simulation parameters; the trajectory simulation parameters comprise simulation starting position, starting time, simulation height, trace-back time length, and time interval between trajectory points of the backward trajectory; a plurality of backward trajectories are generated by a trajectory simulation model; Step 2, numbering and structurally coding each backward trajectory and its trajectory points to form an encoded backward trajectory corresponding to each backward trajectory, and constructing a trajectory data set comprising the encoded backward trajectories; Step 3, automatically or manually selecting a clustering strategy from a plurality of preset trajectory clustering algorithms based on the dimensionality, time sequence characteristics, community structure complexity, and nonlinear path characteristics of each encoded backward trajectory in the trajectory data set; Step 4, applying the selected clustering strategy to the trajectory data set for classification to generate a plurality of trajectory categories; the clustering strategy comprises an integrated method based on an autoencoder and a recurrent neural network, and a graph-based clustering method; Step 5, performing classification processing on each trajectory category and outputting a processing result after the classification processing.

2. The back-trajectory analysis method for atmospheric pollutant transport pathway identification of claim 1, wherein, In the step of generating a plurality of backward trajectories by a trajectory simulation model in step 1, the trajectory simulation model is a HYSPLIT model, GDAS meteorological data is used, the simulation height is set to 100 meters, the trace-back time length is set to 48 hours, and the time interval between trajectory points of the backward trajectory is set to 1 hour.

3. The back-trajectory analysis method for atmospheric pollutant transport pathway identification according to any one of claims 1-2, characterized in that, In step 3, when the dimensionality of each encoded backward trajectory in the trajectory data set exceeds a first preset threshold, it is determined that the trajectory data set has high dimensionality; The time dependence of each encoded backward trajectory is analyzed, and when the time dependence of each encoded backward trajectory in the trajectory data set exceeds a second preset threshold, it is determined that the trajectory data set has strong time sequence characteristics; And when the trajectory data set has high dimensionality and strong time sequence characteristics at the same time, the selected clustering strategy is the integrated method based on the autoencoder and the recurrent neural network.

4. The back-trajectory analysis method for atmospheric pollutant transport pathway identification according to any one of claims 1-2, characterized in that, In step 3, a graph structure is constructed based on the trajectory points of the backward trajectory corresponding to each encoded backward trajectory and a community structure is detected, and when the complexity of the community structure exceeds a third preset threshold, it is determined that the trajectory data set has a complex community structure, and the selected clustering strategy is the graph-based clustering method, wherein the trajectory points of the backward trajectory are nodes of the graph structure, and edges are established between nodes having spatial relationships or having similarity between the backward trajectories; similarity measurement is performed on the backward trajectory corresponding to each encoded backward trajectory, and when the similarity measurement result indicates that the trajectory data set has nonlinear path characteristics, the selected clustering strategy is the graph-based clustering method.

5. The back-trajectory analysis method for atmospheric pollutant transport pathway identification according to any one of claims 1-2, wherein, In step 4, the integrated method based on the autoencoder and the recurrent neural network comprises: The automatic encoder is used for dimension reduction and feature extraction of each encoded backward trajectory, and a low-dimensional latent representation corresponding to each encoded backward trajectory is generated; a recurrent neural network is used for processing the low-dimensional latent representation, and a time dynamic feature corresponding to each encoded backward trajectory is obtained, and a feature representation corresponding to each encoded backward trajectory is generated; the feature representation is clustered to generate a plurality of trajectory categories; the automatic encoder comprises an encoder and a decoder, wherein the encoder maps each encoded backward trajectory to a latent space, the decoder reconstructs the low-dimensional latent representation back to an input space, and is trained by minimizing the reconstruction error.

6. The back trajectory analysis method for atmospheric pollutant transport pathway identification of claim 5, wherein, The recurrent neural network is a long short-term memory network, which is configured to process a sequence of the low-dimensional latent representation and generate a feature representation.

7. The back-trajectory analysis method for atmospheric pollutant transport pathway identification according to any one of claims 1-2, wherein, In step 4, the graph-based clustering method comprises: using a community detection algorithm to identify community structure in the graph structure; and grouping encoded backward trajectories belonging to the same community into the same trajectory category.

8. The back trajectory analysis method for atmospheric pollutant transport pathway identification of claim 7, wherein, The community detection algorithm is a Louvain algorithm, which divides the community structure by maximizing the modularity; The edge weight in the graph structure is based on a similarity measure between the encoded backward trajectories, wherein the similarity measure comprises a dynamic time warping distance.

9. The back trajectory analysis method for atmospheric pollutant transport pathway identification according to claim 1, wherein, The processing result after the classification processing in step 5 comprises: a trajectory category label, a representative trajectory path, a trajectory member quantity, a monthly distribution ratio, and a trajectory classification layer file suitable for a geographic information system; the representative trajectory path is an average trajectory calculated based on each backward trajectory corresponding to the trajectory category, and the latitude and longitude coordinates of the trajectory points in the average trajectory are the arithmetic mean of the latitude and longitude coordinates of the corresponding trajectory points in each backward trajectory corresponding to the trajectory category.

10. A backward trajectory analysis device for atmospheric pollutant transport pathway identification, characterized by, Specifically comprising: A backward trajectory generation module is configured to set trajectory simulation parameters, wherein the trajectory simulation parameters comprise a simulation starting position, a starting time, a simulation height, a traceback duration, and a time interval between trajectory points of the backward trajectory; and a plurality of backward trajectories are generated by a trajectory simulation model; A trajectory dataset construction module is configured to number and structure code each backward trajectory and its trajectory points to form an encoded backward trajectory corresponding to each backward trajectory, and construct a trajectory dataset comprising each encoded backward trajectory; A clustering strategy selection module is configured to automatically or manually select a clustering strategy from a plurality of preset trajectory clustering algorithms based on the dimension, time sequence characteristics, community structure complexity, and nonlinear path features of each encoded backward trajectory in the trajectory dataset; A trajectory category generation module is configured to apply the selected clustering strategy to the trajectory dataset to generate a plurality of trajectory categories; the clustering strategy comprises an integrated method based on an automatic encoder and a recurrent neural network, and a graph-based clustering method; A result output module is configured to perform classification processing on each trajectory category and output a processing result after the classification processing.

Images