An atmospheric pollution tracing and accurate prediction system based on multi-source observation and diffusion model

Abstract

The application discloses an atmospheric pollution tracing and accurate prediction system based on multi-source observation and diffusion models, comprising a data acquisition layer, a data processing layer, a model calculation layer and an application service layer. The application provides an innovative multi-source observation and diffusion model fusion scheme, provides all-round technical support for atmospheric environment management, and provides a scientific basis for pollution control decision-making.

Description

Technical Field

[0001] This invention relates to the field of atmospheric environment monitoring and pollution control technology, specifically to an atmospheric pollution source tracing and accurate forecasting system based on multi-source observation and diffusion models, which is applicable to various application scenarios such as urban air quality management, response to heavy pollution weather, environmental supervision of industrial parks, and regional joint prevention and control. Background Technology

[0002] With the acceleration of industrialization and urbanization, air pollution has become a major environmental problem restricting sustainable economic and social development and endangering public health. Air pollution prevention and control urgently needs to address key scientific issues such as pollution source identification, transmission mechanism analysis, and concentration evolution prediction; however, existing technologies still have many limitations.

[0003] 1. Insufficient Spatial Coverage of Pollution Monitoring: Traditional atmospheric environmental monitoring mainly relies on ground-based monitoring stations to obtain pollutant concentration data. However, these stations are sparsely distributed (typically 5-10 kilometers apart in cities), making it difficult to depict the refined spatial distribution characteristics of pollution. Monitoring stations can only provide location information and cannot reflect the pollution status in the areas between stations, resulting in a large number of pollution distribution blind spots, especially in sensitive areas such as industrial parks and urban-rural fringe areas where monitoring capabilities are weak.

[0004] 2. Lack of Vertical Dimensional Observation Capabilities: Ground-based monitoring stations can only acquire near-surface pollutant concentrations, lacking effective means to detect the vertical distribution of pollutants in the atmospheric boundary layer and free troposphere. The vertical transport and mixing of pollutants significantly affect ground concentrations; ignoring vertical structure information will lead to inaccurate analysis of pollution formation mechanisms and distorted judgments of transport paths.

[0005] 3. Limited accuracy of pollution source apportionment methods: Existing pollution source apportionment methods mostly employ single-receptor models, relying on human experience to determine the source composition profile, which is insufficient for distinguishing complex mixed pollution sources. Traditional methods struggle to quantify the contribution of secondary pollution, cannot identify unconventional emission events, and result in significant uncertainty and poor timeliness in source apportionment results, making it difficult to provide timely support for emergency management.

[0006] 4. Insufficient ability to track transport paths: Cross-regional transport of pollutants is a major cause of heavy pollution weather, but existing trajectory models mostly perform unidirectional backward calculations, which cannot accurately quantify the contribution rate of different source regions to receiver points. Lagrange particle diffusion models have low computational efficiency and are difficult to achieve large-scale high-resolution simulations; Eulerian grid models have numerical diffusion problems and their accuracy in characterizing pollution plumes is not high.

[0007] 5. Low accuracy of air quality forecasts: Although traditional numerical forecasting models include complete physicochemical processes, they are affected by factors such as emission inventory uncertainties, meteorological input errors, and limitations of parameterization schemes, resulting in significant forecast biases. While purely statistical forecasting methods can capture historical patterns, they lack constraints from physical mechanisms and have weak predictive capabilities for pollution processes under extreme weather conditions, making it difficult to accurately warn of heavy pollution events more than three days in advance.

[0008] 6. Lack of Multi-Source Data Fusion Mechanism: While ground-based monitoring, satellite remote sensing, and lidar are complementary observation methods, existing systems lack an effective data fusion framework. Various observation data differ significantly in spatiotemporal resolution, measurement accuracy, and coverage. Simply stitching together or substituting data cannot fully leverage the synergistic advantages of multi-source data, resulting in a waste of observational resources.

[0009] 7. Insufficient Scientific Basis for Precise Governance: Pollution control measures often rely on experience-based judgments and lack quantitative assessments of emission reduction effects. A "one-size-fits-all" approach results in high socio-economic costs, while differentiated control lacks scientific basis. Existing systems struggle to answer key questions such as "which industries, which time periods, and how much emission reduction is needed to meet standards," thus hindering the ability to achieve precise pollution control.

[0010] Therefore, there is an urgent need for an intelligent air pollution management system that integrates multi-source observation data and comprehensively utilizes advanced technologies such as data assimilation, pollution source analysis, transmission simulation, and deep learning. This system would enable refined monitoring, accurate source tracing, high-precision forecasting, and scientific governance of pollution, promoting a shift in air pollution prevention and control from "extensive management" to "precision control." This invention addresses the aforementioned technical challenges by proposing an innovative multi-source observation and diffusion model fusion scheme, providing comprehensive technical support for atmospheric environmental management. Summary of the Invention

[0011] To address the aforementioned issues, this invention provides an atmospheric pollution source tracing and accurate forecasting system based on multi-source observation and diffusion models, offering a scientific basis for pollution control decisions.

[0012] The specific plan is as follows:

[0013] An atmospheric pollution source tracing and accurate forecasting system based on multi-source observation and diffusion models includes:

[0014] The data acquisition layer deploys ground monitoring stations to acquire data from satellite remote sensing, lidar observations, meteorological data, and pollution source inventories.

[0015] The data processing layer uses data assimilation technology to integrate observation data from ground monitoring stations, satellite remote sensing, and lidar to reconstruct a high-resolution three-dimensional pollutant concentration field.

[0016] The model computation layer employs a combination of receptor model and machine learning to quantitatively analyze pollution sources; and utilizes a Lagrange-Euler mixed diffusion model to track pollution transport paths and quantify regional contributions.

[0017] The application service layer uses deep neural networks to correct numerical model forecast biases, achieving high-precision air quality forecasts.

[0018] Furthermore, in the data processing layer, multi-source observation data with significant differences in spatiotemporal resolution are collaboratively integrated into the chemical transport model to achieve high-precision three-dimensional reconstruction of the pollutant concentration field; ground monitoring stations provide high temporal resolution (hourly level) and high-precision point observations, but spatial coverage is sparse; satellite remote sensing provides column concentration or aerosol optical thickness information with wide coverage (hundreds of kilometers in width), but temporal resolution is low (daily level) and affected by cloud cover; lidar provides vertical profile detection (several meters in resolution), but the observation range is limited to a single point or along a line;

[0019] To fully leverage the advantages of various observation methods and overcome their limitations, this invention employs four-dimensional variational assimilation (4D-Var) technology to integrate observational data into the WRF-Chem chemical transport model; assuming the pollutant concentration field predicted by the model is... ,in Spatial location coordinates, Time; the set of observation data is denoted as . , representing observations from ground monitoring, satellite remote sensing, and lidar, respectively;

[0020] The data assimilation process is achieved by minimizing the following objective function:

[0021]

[0022] in, The background field (prior estimate) for the model is obtained from the free prediction of the chemical transport model WRF-Chem; The background error covariance matrix characterizes the spatial correlation structure of model uncertainty. This represents the total number of grid points in the model. As an observation operator, it maps the concentration field in the model space to the observation space. These correspond to ground stations, satellites, and lidar, respectively. The observation error covariance matrix reflects the accuracy characteristics of various observations; Indicates matrix transpose. Summing the three types of observation data;

[0023] The first term of the objective function The constraint is that the assimilation results do not deviate too far from the physical and chemical laws of the model, the second term The constrained assimilation results are consistent with the actual observations; the trade-off between the two is determined by the error covariance matrix. and Regulation; this optimization problem involves calculating the objective function with respect to the concentration field. The gradient is obtained iteratively using a quasi-Newton method (such as the L-BFGS algorithm): in each iteration, the gradient is calculated using the adjoint mode. Then, the concentration field is updated along the negative gradient direction until the gradient norm is less than the convergence threshold or the maximum number of iterations is reached. By solving this optimization problem, the optimal concentration field estimate that simultaneously satisfies the model dynamics constraints and observation constraints is obtained. (Analysis field), which has a spatial resolution of 1 km, a vertical resolution of 100 m and a temporal resolution of 1 hour, significantly improves the spatiotemporal precision of pollution monitoring compared with the original model forecast and sparse observation.

[0024] The assimilated three-dimensional concentration field not only filled the spatial blind spots between monitoring stations but also reconstructed the vertical distribution structure of pollutants, revealing the concentration gradient and mixing layer height variations within the boundary layer. This high-resolution concentration field... As fundamental data, it provides accurate receptor concentration information for subsequent pollution source analysis and initial concentration distribution for transmission path tracing, thus forming the data foundation for the entire source tracing analysis.

[0025] Furthermore, in the model calculation layer, pollution source apportionment is a crucial step in identifying pollution sources and clarifying key areas for remediation. Its analysis objects are the high-precision concentration field reconstructed from the aforementioned data assimilation and the chemical composition observation data from monitoring stations. This invention innovatively combines traditional receptor models with modern machine learning methods to achieve quantitative and refined analysis of pollution sources.

[0026] First, a positive definite matrix factorization (PMF) acceptor model was used to separate the source component spectra of pollutant chemical component data at the monitoring points; let the chemical component concentration matrix obtained at a certain monitoring point within a certain time period be... ,in The number of samples (time steps) Chemical components (such as organic carbon (OC), elemental carbon (EC), sulfates, nitrates, ammonium salts, crustal elements, etc., usually...) The PMF model decomposes the observed concentration matrix into the product of the source contribution matrix and the source component spectrum matrix.

[0027]

[0028] in, Contribution matrix to source, elements Indicates the w-th time step. The contribution of each pollution source to the mass concentration. The number of pollution source types (such as industrial emissions, vehicle exhaust, dust, biomass combustion, secondary emissions, etc., usually...) ); The source component spectral matrix has elements. Indicates the first The first of the pollution sources Mass fraction of the chemical components; This is the residual matrix, representing the portion that the model cannot explain.

[0029] The model parameters are solved by minimizing the weighted sum of squared residuals:

[0030]

[0031] in, For the q-th sample The observed concentrations of the components, The corresponding observation uncertainty is determined by measurement error and detection limit; this optimization problem is solved using the multivariate curve-resolved alternating least squares (MCR-ALS) method: first, the source component spectral matrix is ​​randomly initialized. Then, alternately fix Solve (via nonnegative least squares) and fixed Solve (Also using the nonnegative least squares method), iterate until the objective function is achieved. The algorithm converges to a local minimum; to avoid getting trapped in local optima, multiple random initializations and selections are used. The solution with the smallest value is taken as the final result; the source contribution matrix is ​​obtained by solving. Then, the contribution rate of each pollution source to PM2.5 is calculated as the percentage of the concentration contributed by that source to the total concentration;

[0032] However, traditional PMF methods have two shortcomings: first, they require manual judgment of source type and quantity, which is highly subjective; second, they cannot handle changes in source component spectra over time. To overcome these limitations, this invention introduces machine learning methods to correct and enhance the source resolution results.

[0033] Specifically, a gradient boosting decision tree (XGBoost) model is constructed, taking meteorological elements (wind speed, wind direction, temperature, humidity, boundary layer height, etc.), temporal characteristics (hours, days of the week, seasons), and geographical elements (distance from major emission sources, land use type) as inputs and the contribution of each source obtained from PMF analysis as the objective. This model can learn the complex nonlinear relationship between pollution source contributions and environmental factors, and identify the main controlling influencing factors. Through feature importance analysis and SHAP value calculation, the intensity of the effect of each influencing factor on different pollution sources can be quantified. For example, it was found that the contribution of motor vehicle emissions increases significantly during morning and evening peak hours, dust contribution is positively correlated with wind speed, and secondary pollution is aggravated under high temperature and high radiation conditions.

[0034] After the machine learning model is trained, the PMF (Pollutant Matrix Factor) analysis results are dynamically corrected to capture the time-varying characteristics of source component spectra. This extends source apportionment capabilities to time periods and locations where chemical component sampling was not conducted, significantly improving the spatiotemporal coverage and accuracy of source apportionment. The contribution rate information of each source obtained from source apportionment (the proportion of industry, transportation, dust, etc.) will be combined with subsequent transport path analysis to jointly determine the complete source chain of pollution. This method organically combines the chemical fingerprinting advantages of receptor models with the pattern discovery capabilities of machine learning, providing a reliable basis for accurately identifying pollution source types and quantitatively assessing the contribution of each source.

[0035] Furthermore, in the model's computational layer, the cross-regional transport of pollutants significantly impacts the air quality at the receiver sites. Accurately tracking transport paths and quantifying regional contributions are prerequisites for implementing joint prevention and control. Using the three-dimensional wind and concentration fields reconstructed from data assimilation as driving conditions, and combining pollutant emission inventory data, a hybrid method combining the Lagrange particle diffusion model and the Eulerian grid model is employed to balance computational accuracy and efficiency. The transport analysis results will supplement the local source contribution in the aforementioned source apportionment, revealing the extent of the impact of external transport on the receiver sites.

[0036] The Lagrange method simulates contaminant transport by releasing a large number of tracer particles and tracking their trajectories; from the receptor point At any moment Release backward A particle (usually) ), No. The positional evolution of each particle follows:

[0037]

[0038] in, For the first The three-dimensional position vector of each particle. For time, This represents the three-dimensional average wind field at the particle's location (obtained through interpolation from a meteorological model). To represent the turbulent fluctuation velocity, a random walk model is used to characterize the diffusion process at the sub-grid scale. This represents the rate of change of the particle's position over time; integrating this equation from... Tracing back to (generally (hours), to obtain the set of backward trajectories of particles;

[0039] Statistics in the spatiotemporal domain The residence time of particles within the region, combined with the emission flux in that region, is used to calculate the concentration contribution of the source region to the acceptor site. The specific calculation method is as follows: the spatial domain is divided into a regular grid, the number of particles in each grid at each time step is counted, and the result is normalized to obtain the residence time density function, which reflects the probability of pollutants being transported from that grid to the acceptor site; the source-acceptor relationship function is defined as follows:

[0040]

[0041] in, For the receptor point at time The concentration of pollutants, Let be the particle residence time density function (unit: time / volume), its numerical calculation is as follows: for each spatial coordinate... Statistics in time nearby The number of particles falling into the grid within the window Divide by the total number of particles And the mesh volume, to obtain ; for In time The emission flux (unit: mass / area / time) is read from the emission inventory database; For the entire simulation area, To trace the duration, the integral is discretized into a grid summation in the actual calculation: ,in The area is the grid area; by summing the values ​​of different spatial regions, the percentage contribution of each source region to the receptor point is obtained.

[0042] While the Lagrange method offers high accuracy in characterizing pollution plumes, its computational cost increases rapidly with the number of particles and the tracing time. To improve efficiency, this invention employs a coarse-grid simulation using the Eulerian mesh model (CMAQ) for far-field transport, obtaining regional-scale concentration distribution and transport flux. For the near-field (within a 100km radius of the receiver point), a high-resolution Lagrange simulation is used. The two models are seamlessly integrated at the boundary through concentration and flux matching, forming a hybrid simulation framework of "far-field Eulerian-near-field Lagrange". This method reduces computation time to one-fifth of the original while maintaining transport path tracing accuracy, meeting the real-time requirements of operational applications.

[0043] The transmission path analysis results can identify major pollution transport channels, high-risk meteorological conditions, and key source areas. Combined with the aforementioned source type analysis, a complete source tracing chain of "pollution source type - emission area - transmission path - receptor impact" is formed, providing a scientific basis for formulating regional joint prevention and control plans, such as identifying upwind cities and industries that need to coordinate emission reduction during periods of heavy pollution.

[0044] Furthermore, in the application service layer, although the numerical chemical transport model includes complete physicochemical processes, its forecasts suffer from systematic biases due to factors such as emission inventory uncertainties, boundary condition errors, and simplification of parameterization schemes. By using a high-precision historical concentration field generated through data assimilation as training labels, and combining model forecasts with actual observation data, a deep neural network is constructed to dynamically correct model forecast results for future periods. This combines the mechanistic interpretation capabilities of numerical simulation with the error capture capabilities of deep learning. The corrected high-precision forecast results can further guide the early deployment of control measures.

[0045] Air quality forecasting is considered a spatiotemporal series prediction problem; assuming the numerical model is at time... For the future The predicted concentration at time is The historical observation sequence is The weather forecast sequence is ,in Includes temperature, humidity, wind speed, wind direction, and boundary layer height;

[0046] A spatiotemporal graph neural network (ST-GNN) is constructed for forecast correction; monitoring stations are regarded as nodes in a graph, and the spatial proximity and wind field connectivity between stations are defined as edges, thus forming a spatial topology graph. ,in For a set of nodes, For edge sets; graph convolutional layers aggregate neighborhood node information:

[0047]

[0048] in, For node v (monitoring station) at the th Hidden feature representation of the layer, For the hidden representation of the next layer, Let v be the set of neighboring nodes (defined by distance or wind field connectivity). and These are the degrees (number of edges) of nodes v and u, respectively. For the first The learnable weight matrix of the layer, For non-linear activation functions (such as ReLU), This represents summing over all neighboring nodes; graph convolution operations aggregate normalized neighborhood information. (where is the normalization coefficient) enables the model to capture the spatial diffusion and transport processes of pollutants;

[0049] In the time dimension, a Long Short-Term Memory (LSTM) network is used to model the temporal dependence of concentration evolution. LSTM selectively remembers and forgets historical information through three gating mechanisms: input gate, forget gate, and output gate, avoiding the gradient vanishing problem. Specifically, the implementation method involves processing historical observation sequences... and model forecast sequence Input the data into the LSTM in chronological order, at each time step... The LSTM cells are updated according to the following rules:

[0050] (1) The Forgotten Gate determines which historical information to discard:

[0051] (2) The input gate determines which new information is stored:

[0052] (3) Update cell state:

[0053] (4) Output gate control output information:

[0054] (5) Calculate the hidden state output:

[0055] in, The input features for the current time step (including concentration values ​​and meteorological elements) are... It is in a hidden state. For cellular state (long-term memory), This is element-wise multiplication (Hadamard product). The sigmoid activation function (outputs between 0 and 1, controlling the information flow) The weight matrix and bias vector are learnable; these parameters are trained using the backpropagation algorithm, enabling the LSTM to learn to capture long-term dependency patterns in concentration time-series evolution; the hidden state at the last time step. The encoded temporal features, combined with the spatial features extracted by graph convolution, are fed into a fully connected decoder; the final corrected prediction is output by the decoder.

[0056]

[0057] in, This represents the forecast bias correction learned by the neural network. The model is trained end-to-end by minimizing the mean square error between the corrected forecast and the actual observation, learning the systematic bias of the numerical model from a large number of historical forecast-observation data pairs.

[0058] After training, the model can perform real-time corrections on numerical forecasts for the next 1-7 days. Experiments show that, compared with the original model, the forecasts corrected by deep learning reduce the root mean square error of PM2.5 concentration forecasts by 35-45%, increase the hit rate of heavy pollution events by 20-30 percentage points, and reduce the false alarm rate by 15-20 percentage points, significantly improving the accuracy and reliability of air quality forecasts and buying valuable decision-making time for pollution early warning and emergency response.

[0059] The beneficial effects of this invention are as follows:

[0060] 1. By using multi-source data assimilation technology, sparse ground station observations, low temporal resolution satellite remote sensing, and local lidar detection are fused into a three-dimensional concentration field with high spatiotemporal resolution (1-kilometer grid, hourly level, and 100-meter vertical resolution). This fills the monitoring space blind spots, reveals the three-dimensional structure of pollution, and improves spatial coverage by more than 80% compared with traditional single monitoring methods. Vertical information is obtained from scratch.

[0061] 2. The fusion of receptor model and machine learning not only ensures the physical interpretability of source resolution but also achieves full spatiotemporal coverage and real-time updates. Compared with the traditional PMF method, the uncertainty of source contribution estimation is reduced by 30%, the temporal resolution is improved from daily to hourly, and the spatial coverage is expanded from individual sites to all sites, providing a fine-grained scientific basis for precise control.

[0062] 3. The Lagrange-Eulerian hybrid model balances the accuracy and efficiency of transport simulation, accurately tracking the transport path of pollution plumes, quantitatively assessing the relative contributions of local emissions and external transport, and identifying major transport channels and key source areas. This capability provides a scientific basis for regional joint prevention and control and collaborative governance of cross-border pollution, avoiding isolated efforts by each region focusing solely on its own interests.

[0063] 4. Deep learning bias correction methods effectively compensate for the systematic errors of numerical models, reducing the root mean square error of PM2.5 concentration forecasts by 35-45%, increasing the hit rate of heavy pollution process warnings by 20-30 percentage points, and decreasing the false alarm rate by 15-20 percentage points. High-precision forecasts provide valuable lead time for government departments to initiate emergency responses and for the public to take protective measures, reducing the risk of pollution exposure.

[0064] 5. This invention constructs a complete technological system encompassing "monitoring-source tracing-forecasting-treatment," which not only answers questions such as "where does pollution come from" (source apportionment), "how did it originate" (transmission path), and "how will it develop in the future" (forecasting), but also guides "how to treat it" (precise emission reduction). Compared to single-point technological breakthroughs, this systematic solution has higher practical value and application prospects.

[0065] 6. Through the data assimilation framework, the advantages of heterogeneous observation data such as ground monitoring, satellite remote sensing, and lidar are complemented, and the investment in various observation methods is fully utilized, avoiding data silos and resource waste. The accuracy of the assimilated concentration field exceeds that of any single observation method, demonstrating a synergistic effect of "1+1>2".

[0066] 7. Hybrid diffusion model and parallel computing optimization keep the total computation time within 6 hours, meeting the business requirement of twice-daily forecast releases. Automated process design ensures stable and reliable system operation, reduces manual intervention, and lowers operation and maintenance costs.

[0067] 8. This invention transforms air pollution management from the traditional "experience-based judgment and extensive control" model to a new "data-driven, scientific decision-making, and precise pollution control" model, thereby improving the modernization level of environmental governance. The scientific products output by the system, such as source apportionment, transmission analysis, forecasting and early warning, and emission reduction recommendations, provide solid technical support for government decision-making and promote the scientific and refined approach to air pollution prevention and control.

[0068] 9. This invention is applicable to various scenarios such as urban air quality management, response to heavy pollution weather, industrial park supervision, regional joint prevention and control, pollution discharge permit supervision, environmental impact assessment, and coordinated carbon emission control, and has broad market prospects. The technical solution can be adapted to different geographical regions and pollution types, exhibiting good scalability and adaptability.

[0069] 10. This invention comprehensively utilizes internationally cutting-edge technologies such as data assimilation, machine learning, chemical transport simulation, and deep neural networks to form an innovative technological system with independent intellectual property rights. Key technologies such as multi-source data fusion, source apportionment enhancement, hybrid diffusion simulation, and deep learning correction have reached international advanced levels. Some technologies (such as the fusion of receptor models and machine learning) are the first of their kind in the world, enhancing my country's technological competitiveness in the field of atmospheric environmental science. Attached Figure Description

[0070] Figure 1 This is a flowchart of the present invention.

[0071] Figure 2 This is a diagram of the architecture of the present invention. Detailed Implementation

[0072] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.

[0073] As shown in the figure, the implementation process of an atmospheric pollution source tracing and accurate forecasting system based on multi-source observation and diffusion models in the air quality management of a typical city is described in detail. This city is located on the North China Plain, and in winter, influenced by heating emissions and unfavorable meteorological conditions, PM2.5 pollution is relatively severe, exhibiting significant regional transport characteristics. The specific plan is as follows:

[0074] Construct a four-layer architecture system comprising a data acquisition layer, a data processing layer, a model calculation layer, and an application service layer.

[0075] I. Data Acquisition Configuration:

[0076] Ground-based monitoring network: 35 national-level ambient air quality monitoring stations are deployed to monitor the hourly concentrations of six pollutants—PM2.5, PM10, SO2, NO2, CO, and O3—in real time, with data reporting every 5 minutes. Five superstations are equipped with chemical component analyzers to collect daily data on organic carbon (OC), elemental carbon (EC), and water-soluble inorganic ions (CIOs) in PM2.5. , , , The mass concentrations of 23 chemical components, including (e.g., etc.) and metallic elements (Fe, Al, Ca, K, Zn, etc.), were used for source apportionment.

[0077] Satellite remote sensing data: NO2 tropospheric column concentration products from Sentinel-5P satellite (spatial resolution 3.5×7km, daily transit time approximately 13:30) and aerosol optical depth (AOD) products from MODIS satellite (spatial resolution 1km, 2-4 observations per day), covering the city and surrounding 500km range.

[0078] Lidar observation: Two lidar units were deployed in the urban area, with a vertical detection height of 0-3000 meters, a resolution of 7.5 meters, and a time resolution of 30 seconds, to obtain the vertical profile of the extinction coefficient and retrieve the vertical concentration distribution of PM2.5 and PM10.

[0079] Meteorological data: Access to surface meteorological elements (temperature, humidity, air pressure, wind speed, wind direction, visibility, precipitation) and radiosonde data (temperature, humidity, vertical wind field profile) from the regional meteorological observation network, and receive global numerical weather prediction products from the European Centre for Medium-Range Weather Forecasts (ECMWF) to provide meteorological drivers for chemical transport models.

[0080] Pollution Source Inventory: Integrates emission source inventory data from environmental protection departments, including spatiotemporal distribution and emission intensity information of industrial point sources (chimney location, height, emission rate, and chemical composition of 213 key enterprises), mobile sources (road network distribution, traffic flow, vehicle type structure, and emission factors), and area sources (residential life, construction dust, and agricultural activities), with a grid resolution of 1km.

[0081] II. Implementation of Three-Dimensional Concentration Field Reconstruction Based on Multi-Source Data Assimilation

[0082] 1. Configuration of chemical transport model

[0083] The WRF-Chem model was used to simulate pollutant transport and diffusion. The model was configured with a three-layer nested grid: the outer layer covered East Asia (grid spacing 27 km), the middle layer covered North China (grid spacing 9 km), and the inner layer covered the target city and its surrounding areas (grid spacing 3 km). There were 35 layers vertically, with a simulation domain top height of 50 hPa. The model's physical parameterization schemes included: the YSU boundary layer scheme, the WSM6 microphysics scheme, and the RRTMG radiation scheme. The chemical mechanism employed was a CBMZ carbon bond mechanism coupled to the MOSAIC aerosol module. , , The formation, transformation, and sedimentation processes of aerosol components such as BC and OC.

[0084] The model uses ECMWF weather forecasts as the initial field and boundary conditions, is driven by local emission inventories, and outputs three-dimensional concentration fields of pollutants such as PM2.5, PM10, NO2, and O3, with a time step of 180 seconds and an output frequency of 1 hour.

[0085] 2. Implementation of Four-Dimensional Variational Data Assimilation

[0086] Multi-source observation data were integrated into the WRF-Chem model, and the 4D-Var assimilation algorithm was adopted. For a heavy pollution process in January 2024 (from 00:00 on January 15 to 24:00 on January 20, lasting 6 days), the assimilation window was set to 6 hours, and one assimilation cycle was performed within each window.

[0087] Observation operator design: For ground monitoring stations, observation operators For bilinear interpolation operators, interpolation is performed from the lowest-level grid of the model to the site location; for satellite NO2 column concentration, the observation operator... For the vertical integration operator, the NO2 concentrations of each layer in the model are weighted and summed to obtain the tropospheric column concentration; for the lidar extinction coefficient profile, the observation operator... The aerosol mass concentration of the model was converted into the extinction coefficient using Mie scattering theory.

[0088] Background error covariance matrix The spatial correlation was estimated using the NMC method (comparing the differences between 24-hour and 48-hour forecasts), with a horizontal scale of 50 km and a vertical scale of 500 m, and a standard deviation set to 30% of the background field. The observation error covariance matrix is ​​shown below. The concentrations were set to 10 μg / m³ for the ground station, 20% for the satellite column, and 15% for the lidar.

[0089] Construct the objective function according to formula (1):

[0090]

[0091] The L-BFGS quasi-Newton optimization algorithm was used to solve the problem, and it converged after 50 iterations to obtain the assimilation analysis field. .

[0092] Assimilation Result Evaluation: The biases of model forecasts and independent observations before and after assimilation were compared. Before assimilation, the model's average bias for surface PM2.5 concentration forecasts was +28 μg / m³ (overestimation), with a root mean square error (RMSE) of 45 μg / m³ and a correlation coefficient of 0.62. After assimilation, the average bias decreased to +5 μg / m³, the RMS error decreased to 23 μg / m³, and the correlation coefficient improved to 0.87. Assimilation effectively reduced the model's systematic bias and significantly improved the accuracy of the concentration field.

[0093] The reconstructed three-dimensional concentration field clearly shows the spatiotemporal evolution of this heavy pollution process: the pollution started in the industrial area in the southern part of the city on the night of January 15, with PM2.5 concentrations exceeding 150 μg / m³; during the day on the 16th, it was transported to the urban area under the influence of southeasterly winds, with a large concentration gradient within the boundary layer (0-500 meters), and the concentration rapidly decreased above 500 meters; from the 17th to the 18th, under the control of stable weather, the boundary layer height was compressed to 300 meters, the vertical mixing of pollutants was restricted, and the ground concentration accumulated to a peak of 210 μg / m³; in the afternoon of the 19th, cold air invaded from the northwest, the boundary layer rapidly rose to 1200 meters, the pollutants diffused vertically, and the ground concentration rapidly decreased.

[0094] This high-precision three-dimensional concentration field This data forms the foundation for subsequent analysis: its contained PM2.5 spatiotemporal distribution information provides receptor concentration data for source apportionment, its three-dimensional wind field information provides meteorological drivers for transport simulation, and it will be used as the initial field for model forecast calculations in future periods. Next, pollution source apportionment will be carried out based on this concentration field and chemical composition observation data.

[0095] III. Implementation of Pollution Source Appraisal

[0096] After obtaining a high-precision concentration field through data assimilation and reconstruction, this section uses chemical composition observation data from the superstation to perform quantitative analysis of pollution sources and identify the contribution ratio of different pollution source types to PM2.5.

[0097] 1. Source analysis of the PMF receptor model

[0098] Chemical component data collected from five superstations throughout the entire monitoring period (2023) were used to construct a concentration matrix. (One sample per station per day, 5 stations × 365 days = 1825 samples, 23 chemical components).

[0099] Establish the PMF decomposition model according to formula (2):

[0100]

[0101] Number of pollution sources Based on multiple trial calculations and the physical rationality of the source component spectrum, it was determined to be classified into 7 categories: industrial coal combustion emissions (rich in...) (OC, Zn), vehicle emissions (rich in EC, OC, Zn), OC), dust (rich in crustal elements such as Ca, Fe, and Al), biomass combustion (rich in... , OC), secondary generation (rich in) , , ), sea salt (rich in) , (This city is far from the sea and its contribution is very small), others.

[0102] Solve the problem of minimizing equation (3) to obtain the source contribution matrix. Source component spectral matrix The solution is obtained using the MCR-ALS alternating least squares method iteratively: first, random initialization... Matrix (each row normalized to the source component spectrum), then in the... In the next iteration, fix Solve using the nonnegative least squares method. (make sure ), then fix Solve (Also satisfying the non-negativity constraint and the sum of each row being 1), calculate the objective function:

[0103]

[0104] Iterate until The relative change is less than 0.01% or the maximum number of iterations (100) is reached. To avoid local optima, 20 different random initializations are run, and then... The solution with the minimum value. The annual average contribution rate of each source to PM2.5 was calculated as follows: industrial coal combustion 28%, motor vehicle emissions 22%, dust 15%, biomass combustion 8%, secondary formation 19%, and others 8%. The source contribution rates changed on heavily polluted days (PM2.5 > 150 μg / m³): the contribution of secondary formation increased to 32%, industrial coal combustion increased to 35%, and the contributions of motor vehicles and dust decreased relatively, indicating that heavy pollution was mainly caused by the accumulation of primary emissions and enhanced secondary transformation.

[0105] 2. Machine Learning Augmentation Source Analysis

[0106] To improve the spatiotemporal resolution and predictive capability of source apportionment, an XGBoost machine learning model was constructed. Input features included: meteorological elements (temperature, relative humidity, wind speed, wind direction, boundary layer height), temporal features (hour, day of the week, month, heating season), and spatial features (station number, distance to major industrial areas, distance to main roads), totaling 15 features. The objective variable was the hourly contribution concentration of the seven source classes obtained from PMF analysis.

[0107] The training set consists of data from January to October 2023 (approximately 7000 hours), the validation set consists of data from November (approximately 700 hours), and the test set consists of data from December (approximately 700 hours). Model hyperparameter settings: tree depth 6, learning rate 0.05, number of trees 500, minimum number of splits 20.

[0108] After training, the model performed as follows on the test set: the average R² of the predicted concentration contribution from each source was 0.76, and the root mean square error was 18% of the actual value. Feature importance analysis showed that boundary layer height had the greatest impact on the contribution of all sources (importance score 0.32), followed by wind speed (0.21), temperature (0.15), and time period (0.12), while spatial distance had a lower importance (0.08). SHAP value analysis revealed that a 100-meter decrease in boundary layer height increased the average ground-level contribution concentration of all sources by 15%; when the wind speed was below 2 m / s, the contribution of primary emission sources (industrial and motor vehicle sources) increased significantly; and when the temperature was above 25℃, the contribution of secondary generation was significantly enhanced.

[0109] Using a trained machine learning model, source apportionment extrapolation can be performed for time periods and sites where chemical components were not sampled (such as ordinary national monitoring stations). This expands the spatiotemporal coverage of source apportionment from 5 stations × 365 days to 35 stations × 8760 hours, providing fine-grained source contribution information for precise control. Source apportionment results indicate that secondary generation (32%) and industrial coal combustion (35%) are the main controlling source types during periods of heavy pollution. However, this only reflects the source type contribution of local receiver sites; it is still necessary to clarify whether these pollutants originate from local emissions or external transmission. Therefore, further transmission path tracing analysis is required.

[0110] IV. Implementation of Transmission Path Tracing and Regional Contribution Quantification

[0111] Having clarified the contributions of pollution source types, this section further tracks the spatial sources and transport paths of pollutants, quantifying the relative contributions of local emissions and regional transport. For the pollution situation where the PM2.5 concentration at a certain station in the city center reached its peak (210 μg / m³) at 20:00 on January 17th, a Lagrange-Euler hybrid model was used to track the transport path, with the three-dimensional wind field and emission inventory reconstructed from data as input.

[0112] 1. Backward trajectory tracking

[0113] Using the HYSPLIT Lagrange model, 50,000 tracer particles were released backward from the acceptor point (site coordinates, altitude 100 meters) at 20:00 on January 17th, for a retrospective period of 72 hours (up to 20:00 on January 14th). The particle position evolution was calculated according to formula (4):

[0114]

[0115] in, For the first The three-dimensional position vector of each particle. For time, This represents the three-dimensional average wind field at the particle's location (obtained by interpolation from the three-dimensional wind field output by the WRF meteorological model). The turbulent fluctuation velocity (generated by a random walk scheme, with the diffusion coefficient parameterized based on boundary layer stability) This represents the rate of change of a particle's position over time.

[0116] The residence time of particles in different regions was statistically analyzed, and the regional contribution was calculated according to formula (5) based on the emission inventory of each region. The specific calculation process is as follows: the entire simulation domain is divided into a regular grid of 1km×1km horizontally and 1km vertically. Spatial statistics are performed on the 72-hour backward trajectory of 50,000 particles. For each grid cell... and every hour Count the number of particles falling within the grid. Normalization yields the residence time density ,in The total number of particles, The grid volume is defined. Then, the PM2.5 emission flux for each grid at each time step is read from the emission inventory. (Unit: μg / m² / h), using discretized form:

[0117]

[0118] in For grid area, The time step is in hours. By summing the grid data for different spatial regions, the concentration contributions of the local area (within the administrative region of the recipient city), city A region, city B region, city C region, and other regions to the recipient point can be calculated separately.

[0119] The results showed that local emissions contributed 65 μg / m³ (31%), surrounding city A contributed 58 μg / m³ (28%), city B contributed 42 μg / m³ (20%), city C contributed 25 μg / m³ (12%), and other areas contributed 20 μg / m³ (9%). The main transport path was as follows: From January 15th to 16th, pollutants from cities A and B were transported northward under the influence of southerly winds, mixing with biomass combustion emissions when passing through agricultural areas; on the night of January 16th, the airflow shifted to southeasterly winds, and industrial emissions from city C were transported to the recipient city along the river valley; on January 17th, under stable weather conditions, local emissions accumulated, transport weakened, but the contribution of local sources increased.

[0120] This analysis clarifies the dominant role of regional transport during this heavy pollution event (accounting for 69%), identifies key cities (A, B, and C) requiring coordinated management, and provides a scientific basis for initiating regional emergency response. Combining the aforementioned source type analysis (industrial coal combustion 35%, secondary pollution 32%) with the spatial source analysis in this section (local 31%, external 69%), a complete pollution source tracing chain is formed: heavy pollution mainly originates from industrial emissions and secondary pollution from surrounding cities, transported via southerly winds, with local emissions playing a cumulative and aggravating role. This source tracing result lays the foundation for subsequent forecasting and control measures.

[0121] V. Implementation of Deep Learning Prediction Bias Correction

[0122] After completing the pollution source analysis, this section trains a deep learning model based on the historical high-precision concentration field data generated by data assimilation to correct the bias of numerical model forecasts, thereby improving the accuracy of air quality forecasts for future periods and buying time for the early deployment of control measures.

[0123] 1. Numerical model forecasts

[0124] Air quality forecasts for the next 72 hours are generated using the WRF-Chem model, with forecasts starting at 08:00 and 20:00 daily, outputting hourly PM2.5, PM10, O3, and NO2 concentration forecasts. The model configuration is the same as in the assimilation experiment, but no data assimilation is performed; the forecasts are driven solely by ECMWF weather forecasts and emission inventories.

[0125] Evaluation of the 2023 full-year forecast performance: The average deviation of the PM2.5 concentration forecast was +12 μg / m³, with a root mean square error of 38 μg / m³ and a correlation coefficient of 0.68. The forecast deviation was even greater for heavily polluted days (PM2.5 > 150 μg / m³), with an average underestimation of -25 μg / m³, indicating that the model's ability to simulate extreme pollution events is insufficient.

[0126] 2. Spatiotemporal graph neural network correction

[0127] An ST-GNN model was constructed to correct the numerical weather prediction. Thirty-five monitoring stations were considered as nodes in a graph, and edges were constructed based on the distance between stations (less than 50 km) and the connectivity of the prevailing wind direction to form a spatial graph. .

[0128] Input features include: PM2.5 concentrations predicted by the model (current time and the next 1-72 hours), historical observed concentrations (past 24 hours), and meteorological forecast elements (temperature, humidity, wind speed, wind direction, boundary layer height). Spatial features are extracted using a 3-layer graph convolutional network, according to formula (6):

[0129]

[0130] The hidden layer dimension is set to 128, and the activation function is ReLU. The time series is encoded using a two-layer LSTM, with a hidden state dimension of 256. The decoder outputs the bias correction for the next 72 hours. The corrected forecast is obtained according to formula (7):

[0131]

[0132] Training data: forecast-observation pairings from January to September 2023 (approximately 6500 pairs); validation data: October-November (approximately 1400 pairs); test data: December (approximately 700 pairs). The loss function was mean squared error, the optimizer was Adam, the learning rate was 0.001, the batch size was 32, and the validation set loss converged after 20 epochs of training.

[0133] Test set evaluation: The root mean square error of the corrected PM2.5 forecast decreased to 24 μg / m³ (a 37% reduction compared to the original model's 38 μg / m³), the average deviation decreased to +3 μg / m³, and the correlation coefficient improved to 0.84. The problem of underestimating heavy pollution days was significantly improved, with the average deviation decreasing from -25 μg / m³ to -8 μg / m³, the hit rate of heavy pollution events increasing from 55% to 78%, and the false alarm rate decreasing from 32% to 18%. The corrected high-precision forecast results, combined with source apportionment and transmission path analysis, can provide regulatory authorities with heavy pollution warnings and targeted emission reduction recommendations 2-3 days in advance.

[0134] VI. System Integration and Business Operation

[0135] The various functional modules are integrated into a unified operational system, forming a complete technology chain of "data acquisition → assimilation and reconstruction → source analysis → transmission tracing → forecasting and correction → control recommendations," enabling automated daily operation.

[0136] 08:00 - Data acquisition begins, acquiring ground monitoring, satellite remote sensing, and lidar observation data from the past 24 hours; 08:30 - 4D-Var data assimilation is performed to reconstruct the three-dimensional concentration field of the past 24 hours and update the initial field of the model; 09:00 - The WRF-Chem model is launched to forecast the next 72 hours; 11:00 - Forecast results are fed into the ST-GNN model for bias correction; 11:30 - Corrected air quality forecast products (grid map, station time series, pollution level) are released; 12:00 - Lagrange transport analysis is initiated for periods of heavy pollution to identify the main source areas; 13:00 - Based on source apportionment results, precise control recommendations (targeted emission reduction measures) are generated; 14:00 - Forecast products and control recommendations are pushed to the environmental protection department's decision-making platform and public APP.

[0137] Since its operational launch in January 2024, the system has significantly improved forecast accuracy, providing government departments with scientific pollution early warning and control decision support, reducing the number of heavily polluted days, and improving air quality.

[0138] In summary, this invention addresses the significant differences in spatiotemporal resolution, measurement accuracy, and coverage among observation methods such as ground monitoring, satellite remote sensing, and lidar. It designs a unified data assimilation framework, maps different types of observation data to the model space through observation operators, and rationally allocates the weights of various observations through the error covariance matrix, thereby achieving complementary advantages and information fusion of multi-source data.

[0139] This invention establishes an objective function that simultaneously satisfies both model physicochemical constraints and observational constraints, and employs an efficient quasi-Newton optimization algorithm to solve the high-dimensional nonlinear optimization problem, achieving a balance between computational efficiency and accuracy. A reasonable estimation of the background error covariance matrix is ​​crucial to ensuring the assimilation effect, requiring accurate characterization of the spatiotemporal correlation structure of model uncertainties.

[0140] This invention combines the chemical fingerprinting capability of the PMF receptor model with the nonlinear mapping capability of XGBoost machine learning, thus preserving the physical interpretability of the source component spectra while improving the spatiotemporal coverage and predictive ability of source resolution. Feature engineering and hyperparameter optimization are key to ensuring the generalization performance of the machine learning model.

[0141] This invention addresses the challenges of high accuracy but high computational cost of the Lagrange method and high efficiency but significant numerical diffusion of the Eulerian method. It proposes a hybrid scheme that combines far-field coarse-grid Eulerian simulation with near-field high-resolution Lagrange simulation. This scheme achieves continuity constraints on concentration and flux at the model interface, balancing computational accuracy and efficiency to meet the needs of operational applications.

[0142] This invention models the monitoring station network as a spatial graph, captures the spatial transport relationships of pollutants through graph convolution, and captures the temporal dependence of concentration evolution through LSTM, constructing an end-to-end deep learning correction model. Model training requires a large amount of historical forecast-observation pairing data; data quality control and feature standardization are crucial to ensuring training stability.

[0143] This invention automates the entire process of data acquisition, quality control, model calculation, result visualization, and product release. It employs a well-designed task scheduling and exception handling mechanism to ensure stable and reliable system operation. Optimization of computing resources and parallelization are key technologies for meeting real-time requirements.

[0144] The above embodiments are merely typical illustrative methods of the present invention, and the scope of protection of the present invention is not limited thereto. All equivalent substitutions and improvements made under the concept of the present invention should fall within the scope of protection. It should be emphasized that any modifications or minor adjustments made by those skilled in the art without departing from the basic principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. An atmospheric pollution source tracing and accurate forecasting system based on multi-source observation and diffusion models, characterized in that, include: The data acquisition layer deploys ground monitoring stations to acquire data from satellite remote sensing, lidar observations, meteorological data, and pollution source inventories. The data processing layer uses data assimilation technology to integrate observation data from ground monitoring stations, satellite remote sensing, and lidar to reconstruct a high-resolution three-dimensional pollutant concentration field. The model computation layer employs a combination of receptor model and machine learning to quantitatively analyze pollution sources; and utilizes a Lagrange-Euler mixed diffusion model to track pollution transport paths and quantify regional contributions. The application service layer uses deep neural networks to correct numerical model forecast biases, achieving high-precision air quality forecasts.

2. The atmospheric pollution source tracing and accurate forecasting system based on a multi-source observation and diffusion model according to claim 1, characterized in that, In the data processing layer, multi-source observation data with significant differences in spatiotemporal resolution are collaboratively integrated into the chemical transport model to achieve high-precision three-dimensional reconstruction of the pollutant concentration field; ground monitoring stations provide high temporal resolution and high-precision point observations; satellite remote sensing provides large-area coverage of column concentration or aerosol optical thickness information; lidar provides vertical profile detection. The observational data were integrated into the WRF-Chem chemical transport model using a four-dimensional variational assimilation technique; the pollutant concentration field predicted by the model was assumed to be... ,in Spatial location coordinates, Time; the set of observation data is denoted as . , representing observations from ground monitoring, satellite remote sensing, and lidar, respectively; The data assimilation process is achieved by minimizing the following objective function: in, The background field for the model was obtained from the free prediction of the chemical transport model WRF-Chem; The background error covariance matrix characterizes the spatial correlation structure of model uncertainty. This represents the total number of grid points in the model. As an observation operator, it maps the concentration field in the model space to the observation space. These correspond to ground stations, satellites, and lidar, respectively. The observation error covariance matrix reflects the accuracy characteristics of various observations; Indicates matrix transpose. Summing the three types of observation data; The first term of the objective function The constraint is that the assimilation results do not deviate too far from the physical and chemical laws of the model, the second term The constrained assimilation results are consistent with the actual observations; the trade-off between the two is determined by the error covariance matrix. and Regulation; this optimization problem involves calculating the objective function with respect to the concentration field. The gradient is obtained by iteratively solving using a quasi-Newton method: in each iteration, the gradient is calculated using the adjoint mode. Then, the concentration field is updated along the negative gradient direction until the gradient norm is less than the convergence threshold or the maximum number of iterations is reached. By solving this optimization problem, the optimal concentration field estimate that simultaneously satisfies the model dynamics constraints and observation constraints is obtained. .

3. The atmospheric pollution source tracing and accurate forecasting system based on a multi-source observation and diffusion model according to claim 2, characterized in that, In the model computation layer, the traditional receptor model is combined with modern machine learning methods to achieve quantitative and refined analysis of pollution sources; First, a positive definite matrix factorization receptor model was used to separate the source component spectra of pollutant chemical composition data at the monitoring points; Suppose that the chemical component concentration matrix obtained by a certain monitoring point within a certain time period is . ,in For the sample size, For chemical component fractions; the PMF model decomposes the observed concentration matrix into the product of the source contribution matrix and the source component spectral matrix: in, Contribution matrix to source, elements Indicates the w-th time step. The contribution of each pollution source to the mass concentration. Number of pollution source types; The source component spectral matrix has elements. Indicates the first The first of the pollution sources Mass fraction of the chemical components; This is the residual matrix, representing the portion that the model cannot explain. The model parameters are solved by minimizing the weighted sum of squared residuals: in, For the q-th sample The observed concentrations of the components, The corresponding observation uncertainty is determined by measurement error and detection limit; this optimization problem is solved using the multivariate curve-resolved alternating least squares method: first, the source component spectral matrix is ​​randomly initialized. Then, alternately fix Solve and fixed Solve Iterate until the objective function is achieved. The algorithm converges to a local minimum; to avoid getting trapped in local optima, multiple random initializations and selections are used. The solution with the smallest value is taken as the final result; the source contribution matrix is ​​obtained by solving. Then, the contribution rate of each pollution source to PM2.5 is calculated as the percentage of the concentration contributed by that source to the total concentration; Machine learning methods are introduced to correct and enhance the source analysis results; specifically, a gradient boosting decision tree model is constructed with meteorological elements, time features, and geographical elements as inputs and the source contributions obtained from PMF analysis as the target. After the machine learning model is trained, the PMF analysis results are dynamically corrected to capture the time-varying characteristics of the source component spectra and extend the source analysis capability to time periods and locations where chemical components were not sampled.

4. The atmospheric pollution source tracing and accurate forecasting system based on a multi-source observation and diffusion model according to claim 3, characterized in that, In the model calculation layer, the three-dimensional wind field and concentration field reconstructed by data assimilation are used as driving conditions. Combined with pollution source emission inventory data, a hybrid method combining the Lagrange particle diffusion model and the Eulerian grid model is adopted to reveal the degree of influence of external transport on the receiver point. The Lagrange method simulates contaminant transport by releasing a large number of tracer particles and tracking their trajectories; from the receptor point At any moment Release backward The particle, the first The positional evolution of each particle follows: in, For the first The three-dimensional position vector of each particle. For time, This represents the three-dimensional average wind field at the particle's location. To represent the turbulent fluctuation velocity, a random walk model is used to characterize the diffusion process at the sub-grid scale. This represents the rate of change of the particle's position over time; integrating this equation from... Tracing back to , thus obtaining the set of backward trajectories of the particles; Statistics in the spatiotemporal domain The residence time of particles within the region, combined with the emission flux in that region, is used to calculate the concentration contribution of the source region to the acceptor site. The specific calculation method is as follows: the spatial domain is divided into a regular grid, the number of particles in each grid at each time step is counted, and the result is normalized to obtain the residence time density function, which reflects the probability of pollutants being transported from that grid to the acceptor site; the source-acceptor relationship function is defined as follows: in, For the receptor point at time The concentration of pollutants, The particle residence time density function is calculated numerically as follows: for each spatial coordinate... Statistics in time nearby The number of particles falling into the grid within the window Divide by the total number of particles And the mesh volume, to obtain ; for In time The emission fluxes are read from the emission inventory database; For the entire simulation area, To trace the duration, the integral is discretized into a grid summation in the actual calculation. ,in The area is the grid area; by summing the values ​​of different spatial regions, the percentage contribution of each source region to the receptor point is obtained. For far-field transport, a coarse-grid simulation using an Eulerian grid model was employed to obtain the concentration distribution and transport flux at the regional scale, while for near-field transport, a high-resolution Lagrange simulation was used.

5. The atmospheric pollution source tracing and accurate forecasting system based on a multi-source observation and diffusion model according to claim 4, characterized in that, In the application service layer, a high-precision historical concentration field generated by data assimilation is used as a training label. Combined with model forecasts and actual observation data, a deep neural network is constructed to dynamically correct the model forecast results for future periods, combining the mechanism interpretation capability of numerical simulation with the error capture capability of deep learning. Air quality forecasting is considered a spatiotemporal series prediction problem; assuming the numerical model is at time... For the future The predicted concentration at time is The historical observation sequence is The weather forecast sequence is ,in Includes temperature, humidity, wind speed, wind direction, and boundary layer height; A spatiotemporal graph neural network is constructed for forecast correction; monitoring stations are regarded as nodes of a graph, and the spatial proximity and wind field connectivity between stations are defined as edges, thus forming a spatial topology graph. ,in For a set of nodes, For edge sets; graph convolutional layers aggregate neighborhood node information: in, For node v at the th Hidden feature representation of the layer, For the hidden representation of the next layer, Let v be the set of neighboring nodes. and Let v and u be the degrees of nodes v and u, respectively. For the first The learnable weight matrix of the layer, It is a non-linear activation function. This represents summing over all neighboring nodes; In the time dimension, a Long Short-Term Memory (LSTM) network is used to model the temporal dependence of concentration evolution; LSTM selectively remembers and forgets historical information through three gating mechanisms—input gate, forget gate, and output gate—to avoid the gradient vanishing problem; the specific implementation method is as follows: the historical observation sequence is processed... and model forecast sequence Input the data into the LSTM in chronological order, at each time step... The LSTM cells are updated according to the following rules: (1) The Forgotten Gate determines which historical information to discard: (2) The input gate determines what new information is stored: (3) Update cell state: (4) Output gate control output information: (5) Calculate the hidden state output: in, The input features are for the current time step. It is in a hidden state. In cellular state, For element-wise multiplication, It is the sigmoid activation function. The weight matrix and bias vector are learnable; these parameters are trained using the backpropagation algorithm, enabling the LSTM to learn to capture long-term dependency patterns in concentration time-series evolution; the hidden state at the last time step. The encoded temporal features, combined with the spatial features extracted by graph convolution, are fed into a fully connected decoder; the final corrected prediction is output by the decoder. in, This is the prediction bias correction amount learned by the neural network.

Images