Artificial Intelligence-Based Aerosol Prediction Method and System

Abstract

The present invention relates to the technical field of meteorological data analysis, and provides an aerosol prediction method and system based on artificial intelligence for accurately identifying abnormal aerosol aggregation patterns in extreme meteorological events. Among them, the method includes: obtaining an original multi-modal meteorological observation data set, which includes a surface aerosol concentration sequence, three-dimensional atmospheric state variables, and boundary layer dynamic parameters; performing spatio-temporal heterogeneity correction on the original multi-modal meteorological observation data set to generate a spatio-temporally aligned target multi-modal meteorological observation data set; inputting the target multi-modal meteorological observation data set into a spatio-temporal feature fusion network to extract multi-scale spatial distribution features and cross-modal temporal correlation features, and generating a spatio-temporal feature matrix based on the multi-scale spatial distribution features and cross-modal temporal correlation features; performing multi-stage recursive optimization on the spatio-temporal feature matrix through a hierarchical time-domain aggregation algorithm to generate an aerosol concentration prediction sequence for the target area.

Description

Technical Field

[0001] The present invention belongs to the technical field of meteorological data analysis, and particularly relates to an aerosol prediction method and system based on artificial intelligence. Background Art

[0002] Aerosol concentration prediction is one of the core technologies for atmospheric environment monitoring and pollution control. Traditional methods mainly rely on ground station observation data or a single meteorological model for statistical analysis. However, in the actual application process of traditional methods, there are often problems such as data heterogeneity constraints, model architecture limitations, insufficient time-domain modeling, and failure in predicting extreme events.

[0003] The above defects lead to bottlenecks in existing aerosol prediction technologies, such as low spatio-temporal resolution, insufficient coupling modeling of multi-physical processes, and poor stability of continuous prediction, making it difficult to meet the urgent needs of high-precision environmental early warning and refined governance.

[0004] Therefore, how to accurately identify abnormal aerosol aggregation patterns in extreme meteorological events is a technical problem that urgently needs to be solved at present. Summary of the Invention

[0005] The present invention provides an aerosol prediction method and system based on artificial intelligence to accurately identify abnormal aerosol aggregation patterns in extreme meteorological events.

[0006] In a first aspect, an embodiment of the present invention provides an aerosol prediction method based on artificial intelligence, which is applied to an aerosol prediction system. The method includes: obtaining an original multi-modal meteorological observation data set, where the original multi-modal meteorological observation data set includes a surface aerosol concentration sequence, three-dimensional atmospheric state variables, and boundary layer dynamic parameters; performing spatio-temporal heterogeneity correction on the original multi-modal meteorological observation data set to generate a spatio-temporally aligned target multi-modal meteorological observation data set; inputting the target multi-modal meteorological observation data set into a spatio-temporal feature fusion network to extract multi-scale spatial distribution features and cross-modal temporal correlation features, and generating a spatio-temporal feature matrix based on the multi-scale spatial distribution features and the cross-modal temporal correlation features; performing multi-stage recursive optimization on the spatio-temporal feature matrix through a hierarchical time-domain aggregation algorithm to generate an aerosol concentration prediction sequence for a target area.

[0007] In a second aspect, an embodiment of the present invention provides an aerosol prediction system, which includes a processor and a memory. Among them, the memory stores a computer program, and when the computer program is executed by the processor, the processor is caused to execute the steps of the above method.

[0008] In a third aspect, an embodiment of the present invention provides a computer-readable storage medium, which includes a computer program. When the computer program runs on an aerosol prediction system, the computer program is used to cause the aerosol prediction system to execute the steps of the above method.

[0009] Through a multi-dimensional data fusion and adaptive modeling mechanism, the embodiments of the present invention significantly improve the accuracy and reliability of aerosol concentration prediction.

[0010] First, by integrating surface aerosol dynamics, three-dimensional atmospheric state, and boundary layer dynamic parameters, a multi-modal collaborative observation system covering the vertical atmosphere and surface environment is constructed, effectively capturing the cross-dimensional coupling effect among meteorological elements. Among them, an innovative spatio-temporal heterogeneity correction method overcomes the differences in spatio-temporal resolution, monitoring dimension, and dimension system of multi-source heterogeneous data, forming a highly consistent standardized meteorological data base.

[0011] On this basis, the spatio-temporal feature fusion network realizes the multi-scale spatial feature analysis from micro-turbulence to macro-circulation by dynamically perceiving the non-linear correlation between atmospheric parameters at different altitudes and the surface environment, enhancing the modeling ability of the aerosol distribution law under complex meteorological conditions.

[0012] In addition, the hierarchical time-domain aggregation framework simultaneously takes into account the instantaneous response characteristics and long-term evolution trend of the meteorological system through a progressive optimization strategy, significantly improving the time coherence of continuous prediction while ensuring real-time prediction efficiency.

[0013] In summary, the embodiments of the present invention can accurately identify abnormal aerosol aggregation patterns in extreme meteorological events, providing high-timeliness dynamic decision-making support for atmospheric pollution source analysis and regional environmental governance strategy formulation. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is a schematic flowchart of a method for aerosol prediction based on artificial intelligence provided by an embodiment of the present invention.

[0015] Figure 2 It is a schematic structural diagram of an aerosol prediction system provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0016] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments recorded in this document of the present invention without creative efforts shall fall within the scope of the technical solutions of the present invention.

[0017] See Figure 1 , which is an artificial intelligence-based aerosol prediction method provided in an embodiment of the present invention. This method can be applied to an aerosol prediction system, and the specific process is as follows in S110 - S140:

[0018] S110: Obtain the original multi-modal meteorological observation dataset, which includes the surface aerosol concentration sequence, three-dimensional atmospheric state variables, and boundary layer dynamic parameters.

[0019] In the embodiment of the present invention, the aerosol prediction system first obtains the original multi-modal meteorological observation dataset from a multi-source observation platform.

[0020] Exemplarily, the surface aerosol concentration sequence is from the ground environmental monitoring station network and the MODIS sensor of the polar orbiting satellite. For example, the PM2.5 and PM10 mass concentration data uploaded by multiple air quality monitoring stations distributed in a certain plain area every few minutes, and the 550nm aerosol optical depth inversion product obtained by the geostationary satellite every hour.

[0021] The three-dimensional atmospheric state variables are generated by fusing meteorological reanalysis data and sounding observations, including the 0.25° spatial resolution three-dimensional wind speed field, temperature field, and humidity field grid data provided by a certain medium-term weather forecast center, and the boundary layer vertical profile data collected by multiple wind profilers deployed in the target area at regular intervals.

[0022] The boundary layer dynamic parameters involve characteristic quantities such as the atmospheric mixed layer height and turbulent kinetic energy flux, and are obtained by assimilating the network observations of lidar and the WRF model. For example, the 1km vertical resolution boundary layer structure parameters output by several Doppler lidars deployed in a certain urban agglomeration at a set time step.

[0023] The above-mentioned multi-source heterogeneous data is standardized in format through the data acquisition module of the aerosol prediction system to form the original multi-modal meteorological observation dataset containing time stamps, geographical coordinates, and numerical quality identifiers.

[0024] S120: Perform spatio-temporal heterogeneity correction on the original multi-modal meteorological observation dataset to generate a spatio-temporally aligned target multi-modal meteorological observation dataset.

[0025] In the embodiment of the present invention, the aerosol prediction system conducts correction processing for the spatio-temporal heterogeneity existing in the original dataset.

[0026] For example, during a typical pollution process in winter in a certain plain, the time resolution of the ground station data is 5 minutes while that of the satellite data is hourly. The aerosol prediction system uses an adaptive time alignment algorithm to perform cubic spline interpolation on the MODIS aerosol optical depth to generate a 5-minute interval dataset synchronized with the ground monitoring.

[0027] In addition, in terms of spatial dimension, in view of the scale difference between the 0.25° grid of meteorological reanalysis data and the 1 km observation grid of lidar, the aerosol prediction system uses an improved Kriging spatial interpolation method to upscale the wind profile radar data to a spatial resolution matching the reanalysis field.

[0028] In response to the problem of missing monitoring of boundary layer parameters on the east side of a certain mountain range, the aerosol prediction system activates the data filling mechanism for the earth's boundary region. Based on the dynamic uplift effect of mountain terrain simulated by the WRF model and combined with the daily variation law of the mixing layer height observed by lidar in the adjacent area, a spatially continuous boundary layer parameter field is generated.

[0029] It can be understood that the dataset after spatio-temporal correction forms a regular grid such as 500 m × 500 m under the UTM-50N coordinate system, and the time reference adopts Coordinated Universal Time to ensure strict alignment of different modality data in the spatio-temporal dimension.

[0030] S130: Input the target multi-modal meteorological observation dataset into the spatio-temporal feature fusion network to extract multi-scale spatial distribution features and cross-modal temporal correlation features, and generate a spatio-temporal feature matrix based on the multi-scale spatial distribution features and the cross-modal temporal correlation features.

[0031] In the embodiment of the present invention, the aerosol prediction system inputs the preprocessed multi-modal data into the spatio-temporal feature fusion network for deep feature extraction.

[0032] For example, the spatial feature extraction module uses the Vision Transformer architecture to process the two-dimensional meteorological field. For example, the wind temperature and humidity grid data at the 850 hPa height level are divided into a sequence of 16×16 tiles, and the sea-land breeze circulation spatial correlation pattern between a certain plain and a certain bay is captured through the multi-head self-attention mechanism.

[0033] The U-Net encoder-decoder structure is used to extract multi-scale features of the aerosol concentration field. In the encoding stage, the 200 km regional scale features of the pollution diffusion of a certain urban agglomeration are captured through 4 times of downsampling, and in the decoding stage, it is gradually restored to the original resolution to retain the 50 km fine structure of the local emission sources.

[0034] Furthermore, the temporal correlation modeling part uses a spatio-temporal encoder to process the meteorological sequence of 72 consecutive hours, where the gated recurrent unit network learns the daily variation period of the boundary layer parameters, and the temporal convolutional network extracts the minute-level fluctuation features of the wind speed pulsation.

[0035] In addition, the global attention mechanism can dynamically fuse different modality features. For example, it can automatically increase the weight of PM2.5 data from ground monitoring stations under static and stable weather conditions, while emphasizing the cross-modal correlation between three-dimensional wind fields and turbulent kinetic energy during the passage of a cold front, and finally generate a spatio-temporal feature matrix containing multi-dimensional feature vectors.

[0036] In the embodiments of the present invention, the spatio-temporal feature matrix is a multi-dimensional data structure constructed by fusing the spatial and temporal dimension features of multi-modal meteorological data. It can be understood that the spatio-temporal feature matrix integrates multi-scale spatial distribution features extracted by a deep neural network (such as the macro-circulation pattern of the regional meteorological field and the fine structure of local pollution sources), as well as cross-modal temporal correlation features (such as the dynamic interaction between different meteorological elements and the law of their evolution over time). The core of the spatio-temporal feature matrix is to use models such as the attention mechanism and the encoder-decoder architecture to map heterogeneous meteorological observation data into a unified feature representation, which includes both topological associations in space (such as the pollutant diffusion path affected by terrain) and continuous dynamics in time (such as the periodic changes and sudden fluctuations of boundary layer parameters), thereby providing high-dimensional feature support for subsequent predictions.

[0037] S140: Perform multi-stage recursive optimization on the spatio-temporal feature matrix through a hierarchical time-domain aggregation algorithm to generate an aerosol concentration prediction sequence for the target area.

[0038] In the embodiments of the present invention, the aerosol prediction system uses a hierarchical time-domain aggregation algorithm to recursively optimize the feature matrix. For example, the 24-hour scale aggregator in the first stage captures the daily cycle change law of aerosol concentration through a bidirectional long short-term memory network. For example, the impact of the typical daytime dissipation and nighttime reconstruction process of the ground radiation inversion layer in a certain area on the vertical diffusion of pollutants.

[0039] For another example, the 6-hour scale aggregation module in the second stage uses a temporal self-attention mechanism to strengthen the feature representation during the pollution accumulation stage, and focuses on modeling the coupling effect of evening traffic peak emissions and a sudden drop in the boundary layer height. The hourly predictor in the final stage integrates multi-scale features and retains the key physical process information in the original spatio-temporal feature matrix through residual connections. For example, when predicting the aerosol concentration in the next 72 hours, the aerosol prediction system cross-temporally fuses the daily-scale changes in the atmospheric circulation background field, the hourly-scale fluctuations in local emission source strength, and the minute-scale turbulent mixing features, and outputs an aerosol concentration prediction field with a spatial resolution of 1 km and a time interval of 1 hour. The above aggregation process can adopt an autoregressive training strategy and achieve rolling optimization of the prediction results through a sliding time window mechanism, significantly improving the prediction stability of continuous pollution processes.

[0040] Based on this, in the embodiments of the present invention, multi-stage recursive optimization is a hierarchical and progressive time series modeling method, aiming to improve the prediction accuracy through feature aggregation and iterative correction at different time granularities. This process adopts a hierarchical time-domain aggregation algorithm to process the spatio-temporal feature matrix in stages.

[0041] At the macro scale stage, long-term evolution laws are captured (such as the inter-daily changes in the background of the atmospheric circulation), and the global time dependence is modeled through a recurrent neural network or an attention mechanism.

[0042] At the meso scale stage, the feature enhancement of key time periods is focused on (such as the meteorological coupling effect during the pollution accumulation period), and the time series self-attention is used to screen and enhance specific dynamic patterns.

[0043] At the micro scale stage: multi-scale features are integrated and the details of physical processes are retained (such as the instantaneous impact of turbulent mixing), and fine prediction is achieved through residual connection and cross-time series fusion mechanism.

[0044] It can be understood that each stage realizes the rolling update of parameters through the recursive feedback and sliding time window mechanism, ensuring the continuous optimization of the prediction results in terms of time series coherence and stability.

[0045] In the embodiments of the present invention, the aerosol concentration prediction sequence of the target area is based on the spatio-temporal feature matrix and the continuous spatio-temporal distribution result output by multi-stage optimization, characterizing the dynamic evolution of the aerosol concentration in the target area within a specific future time period. This sequence is generated by fusing multi-scale spatio-temporal features (such as regional transport, local emissions, and vertical diffusion processes), and has a clear spatial coverage range and time continuity. The characteristics of the aerosol concentration prediction sequence at least include:

[0046] (1) Spatial expression: presented in the form of a regular grid or geographical coding, retaining the spatial heterogeneity of pollutant diffusion (such as the concentration gradient difference between urban agglomerations and natural landforms).

[0047] (2) Temporal dynamics: The predicted values are output at a preset time interval (such as hourly), reflecting the non-linear response of the aerosol concentration to meteorological conditions (such as the accumulation process under static and stable weather and the rapid clearance after the frontal passage).

[0048] (3) Uncertainty control: Through cross-modal feature association and recursive optimization mechanism, the prediction deviation caused by data missing or model error is reduced, ensuring the robustness of the sequence under complex meteorological scenarios.

[0049] In summary, the embodiments of the present invention significantly improve the spatio-temporal accuracy and model generalization ability of aerosol concentration prediction through the collaborative optimization of multi-modal meteorological data fusion and deep learning architecture.

[0050] First, integrate multi-source heterogeneous data such as ground stations, satellite remote sensing, and meteorological reanalysis to construct a three-dimensional observation system covering the surface to the boundary layer, strengthening the completeness of data representation; solve the problems of resolution differences and geographical missing of multi-modal data through an adaptive spatio-temporal correction algorithm, ensuring the spatio-temporal consistency of the input data.

[0051] Secondly, extract multi-scale spatial features based on the hybrid architecture of Vision Transformer and U-Net, and capture cross-modal temporal correlations by combining gated recurrent units and temporal convolutional networks, breaking through the limitations of traditional models in modeling local features and long-range dependencies. Further adopt a hierarchical temporal aggregation algorithm to fuse the diurnal cycle law, pollution accumulation dynamics, and instantaneous turbulence effects across scales through staged recursive optimization, realizing a refined simulation of the meteorological-pollution coupling process.

[0052] The finally output prediction sequence has both high spatio-temporal resolution and physical consistency, can accurately depict the aerosol diffusion law under scenarios such as complex terrain and weather mutations, and provides highly robust scientific support for regional pollution early warning and prevention and control decisions. The embodiment of the present invention systematically solves the modeling problem of the non-linear interaction between meteorological elements and aerosol concentration through the technical loop of "multi-source data - feature fusion - cross-scale prediction", promoting the evolution of aerosol prediction technology towards intelligence and high precision.

[0053] In a preferred embodiment, the spatio-temporal heterogeneity correction of the original multi-modal meteorological observation dataset in S120 to generate a spatio-temporally aligned target multi-modal meteorological observation dataset includes:

[0054] S121: Identify the spatial resolution differences between the vertical stratification data of the three-dimensional atmospheric state variables and the surface aerosol concentration sequence.

[0055] In the embodiment of the present invention, the aerosol prediction system identifies the spatial resolution differences between the vertical stratification data of the three-dimensional atmospheric state variables and the surface aerosol concentration sequence. Specifically, the three-dimensional wind speed field and temperature field provided by the meteorological reanalysis data use a 0.25° longitude-latitude grid, and its vertical direction is divided into 37 pressure layers from the surface to the tropopause, while the surface aerosol concentration data comes from the discrete point observations of ground monitoring stations and the 1km spatial resolution grid data of satellite inversion products.

[0056] For example, in the aerosol prediction task in a certain plain area, the system detects that the grid spacing of the horizontal wind field at the 850 hPa altitude layer is about 27.8 km, while the spacing between ground PM2.5 monitoring stations is 10 - 15 km, and there are significant differences in spatial coverage density between the two. The aerosol prediction system calculates the matching degree between the grid point spacing of each vertical layer data in the horizontal projection direction and the coverage density of surface observation data, and identifies that the spatial resolution difference between the atmospheric state variables and the surface aerosol concentration within the altitude range of 925 hPa to 700 hPa in the boundary layer exceeds the preset threshold, and spatial interpolation processing of the vertical stratified data is required.

[0057] S122: Perform spatial interpolation on the vertical stratified data based on the earth curvature compensation algorithm to generate equi - longitude and latitude grid data matching the surface aerosol concentration sequence.

[0058] In the embodiment of the present invention, the aerosol prediction system performs spatial interpolation on the vertical stratified data based on the earth curvature compensation algorithm to generate equi - longitude and latitude grid data matching the surface aerosol concentration sequence. The earth curvature compensation algorithm performs geometric correction on the traditional plane interpolation method by introducing an arc length correction factor under the ellipsoidal earth model.

[0059] For example, when interpolating the original 1°×1° longitude - latitude grid in meteorological re - analysis data to the same 0.1° grid as the satellite aerosol optical depth data, the system first calculates the geodetic coordinates of the target grid point on the surface of the earth ellipsoid, and then adjusts the interpolation weight according to the spatial curvature of adjacent grid points. For the vertical wind speed field data in a certain bay area, the algorithm automatically compensates for the meridional convergence effect caused by the increase in latitude during the north - south interpolation process, ensuring that the grid spacing of the wind field data at the 925 hPa altitude layer after interpolation is consistent with the spatial distribution of surface monitoring stations in the coastal urban agglomeration, and eliminating the spatial resolution deviation caused by projection deformation.

[0060] S123: For the missing area of the detected boundary layer dynamic parameters, generate a filling mask for the missing area using the eddy covariance data of adjacent meteorological stations.

[0061] In the embodiment of the present invention, the aerosol prediction system generates a filling mask for the missing area of the detected boundary layer dynamic parameters using the eddy covariance data of adjacent meteorological stations. The filling mask is a binary matrix containing spatial weight distribution and data credibility indicators, and is used to identify the relationship between the area to be filled and its adjacent effective observation points.

[0062] For example, when the data of the mixing layer height observed by lidar is continuously missing due to terrain occlusion on the east side of a certain mountain range, the system selects the eddy covariance data of 3 meteorological stations on the west side of the mountain range and calculates the spatio-temporal variation characteristics of its turbulent kinetic energy flux. By analyzing the spatial transfer law of the turbulence intensity during the development process of the valley breeze circulation in the morning, the system constructs a filling mask bounded by the ridge line, where the weight coefficient of the western region of the mask is directly assigned based on the measured data, and the interpolation weight for the missing eastern region is generated according to the terrain slope and distance attenuation function, forming a spatial distribution template of boundary layer parameters covering the entire target area.

[0063] S124: Perform dynamic sliding window calibration according to the timestamp deviation between the filling mask and the equal longitude and latitude grid data to generate a spatio-temporally aligned target multi-modal meteorological observation dataset.

[0064] In the embodiment of the present invention, the aerosol prediction system performs dynamic sliding window calibration according to the timestamp deviation between the filling mask and the equal longitude and latitude grid data to generate a spatio-temporally aligned target multi-modal meteorological observation dataset. Specifically, the system takes the current moment as a reference and dynamically adjusts the start and end times of the time window to compensate for the time synchronization error between different data sources.

[0065] For example, during the winter pollution process in a plain urban agglomeration, the timestamp of the ground monitoring station data is 08:00:00 UTC on December 16th, while the timestamp of the meteorological reanalysis data is 08:02:30 UTC due to calculation delay. After the system detects a time deviation of 2.5 minutes, it extends the sliding window to 08:00:00 - 08:05:00, performs cubic spline interpolation on the reanalysis data within this window, and generates a dataset at 8:00:00 that is strictly aligned with the ground data. In response to the rapid change in the wind speed field during the cold front passage, the system automatically shortens the window step size from 5 minutes to 1 minute to ensure that the sudden change characteristics of the wind direction at the moment of the front arrival are accurately captured and synchronized to other modal data.

[0066] In a preferred embodiment, the performing dynamic sliding window calibration according to the timestamp deviation between the filling mask and the equal longitude and latitude grid data in S124 to generate a spatio-temporally aligned target multi-modal meteorological observation dataset includes:

[0067] S1241: Detect the time span of the continuously missing area in the filling mask.

[0068] In an embodiment of the present invention, the aerosol prediction system detects the time span of continuous missing regions in the filling mask. The time span refers to the duration between the start and end times of the missing data in the time series. For example, during a continuous stable weather process in winter in a certain urban agglomeration, due to cloudy weather, multiple lidars were unable to obtain boundary layer height data from 08:00 on December 5th to 14:00 on December 7th. The system determined through timestamp continuity analysis that the span of this missing period was 54 hours. The system further divided the 54 hours into 6 time windows of 9 hours each, and detected the spatial distribution range of missing grid points and the coverage density of adjacent valid data within each window, identifying a regional data missing caused by foggy weather from 02:00 to 11:00 on December 6th, with a time span of 9 hours and an influence range exceeding 60% of the total area.

[0069] S1242: Extract a similarity template from the historical synchronic longitude-latitude grid data according to the time span.

[0070] In an embodiment of the present invention, the aerosol prediction system extracts a similarity template from the historical synchronic longitude-latitude grid data according to the time span. The similarity template refers to a historical period dataset with meteorological conditions similar to the current missing period, and is used to provide the spatial distribution pattern required for filling.

[0071] For example, for the missing boundary layer parameters from 02:00 to 11:00 on December 6th mentioned above, the system retrieves the lidar observation records in the same period in the past 5 years (from December 1st to 15th), and screens out two historical periods on December 8th, 2019 and December 4th, 2021, whose correlation coefficients of ground meteorological observation data (wind speed, temperature, humidity) and the spatial distribution of the sea level pressure field in the current missing period exceed 0.85. The system extracts the boundary layer height data of these two historical periods within the same UTC time range, and generates a similarity template containing the daily variation characteristics of the mixed layer height under typical stable weather through spatial normalization processing.

[0072] S1243: Use the dynamic time warping algorithm to perform path matching between the similarity template and the current missing region.

[0073] In an embodiment of the present invention, the aerosol prediction system uses the dynamic time warping algorithm to perform path matching between the similarity template and the current missing region. The dynamic time warping algorithm solves the problem of time series phase shift caused by differences in the evolution rate of meteorological processes by constructing a non-linear alignment path between time series.

[0074] For example, when matching the historical template on December 4, 2021 with the ground temperature time series data of the current missing period, the algorithm detects that the dissipation time of the morning inversion layer in the historical template is 2 hours earlier than that of the current missing period. By calculating the Euclidean distance matrix at each time point between the two sequences, the dynamic time warping algorithm finds a minimum cumulative distance path to align the 09:00 data of the historical template with the 11:00 data of the current missing period, thereby correcting the time series misalignment caused by the difference in the moving speed of weather systems.

[0075] S1244: Adjust the step size parameter of the sliding window according to the slope change of the matching path.

[0076] In the embodiment of the present invention, the aerosol prediction system adjusts the step size parameter of the sliding window according to the slope change of the matching path. The step size parameter of the sliding window determines the data sampling interval in the time dimension during the filling process and needs to be adaptively adjusted according to the local slope of the dynamic time warping path.

[0077] For example, when the matching path shows a 45-degree slope in the period from 08:00 to 10:00 (indicating that the historical and current time series are synchronized), the system uses a fixed 1-hour step size for data filling; while in the period from 10:00 to 12:00, the path slope increases to 60 degrees (indicating that the current time series process speeds up), and the system shortens the step size to 0.5 hours to improve the time resolution. For a certain cold front passing process, the system detects a sudden change in the slope of the matching path at the moment when the front arrives, and accordingly dynamically adjusts the step size parameter to ensure that the data filling frequency in the strong wind area behind the front is increased to a 15-minute interval.

[0078] S1245: Perform two-way linear interpolation on the current missing area based on the adjusted step size parameter to generate a target multi-modal meteorological observation data set with spatio-temporal alignment.

[0079] In the embodiment of the present invention, the aerosol prediction system performs two-way linear interpolation on the current missing area based on the adjusted step size parameter to generate a target multi-modal meteorological observation data set with spatio-temporal alignment. The two-way linear interpolation fills the data in both the forward and backward directions along the time axis, and then obtains the final result through weighted averaging.

[0080] For example, when filling in the boundary layer height data at 08:30 on December 6 in a certain valley area, the system obtains the forward 08:00 historical template data and the backward 09:00 satellite inversion data at an adjusted 30-minute time step, and performs linear interpolation on 11 grid points on both sides of the valley axis along the longitude direction. During the interpolation process, the inhibitory effect of terrain elevation on the development of the mixed layer is considered, and an elevation correction factor is introduced at the ridge line, so that the generated data shows the characteristic that the mixed layer height on the leeward slope is lower than that on the windward slope in terms of spatial distribution, which is consistent with the spatial variation law of the measured data. After the above processing, all modal data are unified to the UTM-50N grid coordinate system of 500m×500m, and the time reference is aligned to Coordinated Universal Time, forming a spatio-temporal consistent dataset that can be directly input into the prediction model.

[0081] In a preferred embodiment, the inputting the target multi-modal meteorological observation dataset into the spatio-temporal feature fusion network in S130 to extract multi-scale spatial distribution features and cross-modal temporal correlation features, and generating a spatio-temporal feature matrix based on the multi-scale spatial distribution features and the cross-modal temporal correlation features includes:

[0082] S131: Locally enhance the texture of the surface aerosol concentration sequence through a convolutional attention module to generate a super-resolution spatial feature map.

[0083] In the embodiment of the present invention, the aerosol prediction system locally enhances the texture of the surface aerosol concentration sequence through a convolutional attention module to generate a super-resolution spatial feature map. Specifically, the system inputs the 1km-resolution surface PM2.5 concentration grid data after spatio-temporal correction into the convolutional attention module, and this module first enhances the spatial gradient amplitude using the Gaussian difference filtering algorithm.

[0084] For example, in the industrial agglomeration area of a certain plain urban agglomeration, the aerosol concentration field shows significant spatial heterogeneity within a range of 5km. After the system smooths the original data through a 3×3 Gaussian kernel, it uses the Laplace operator to detect the areas with sudden changes in the concentration gradient around the factory area, generating a gradient amplitude map reflecting the spatial distribution differences of pollution source strengths. Subsequently, the system generates a regional significance weight matrix according to the local variance distribution of the gradient amplitude map, and focuses on enhancing the texture details around traffic arteries and industrial point sources.

[0085] In the above process, the system uses deformable convolutional kernels to extract multi-directional features from the concentration field. For example, for the pollutant diffusion path dominated by sea-land breezes in a certain bay area, the deformable convolutional kernel automatically adjusts the sampling offset direction and captures the plume diffusion characteristics from the coastal sewage outfall to the offshore area along the southeast-northwest axis. Finally, the system performs an element-wise multiplication of the direction-sensitive feature map and the regional saliency weight, and uses a transposed convolutional layer to increase the spatial resolution from 1 km to 500 m, generating a super-resolution spatial feature map that can clearly represent the spatial structure of local emission sources.

[0086] S132: Traverse the vertical profile data of the three-dimensional atmospheric state variables using a three-dimensional dilated convolutional kernel to extract the correlation features between the boundary layer thickness and the aerosol diffusion path.

[0087] In the embodiment of the present invention, the aerosol prediction system traverses the vertical profile data of the three-dimensional atmospheric state variables using a three-dimensional dilated convolutional kernel to extract the correlation features between the boundary layer thickness and the aerosol diffusion path. The three-dimensional dilation convolution captures long-range dependencies in the vertical direction by expanding the receptive field.

[0088] For example, when processing the three-dimensional field data of temperature, humidity, and wind speed over a certain urban agglomeration, the system uses a convolutional kernel with a dilation rate of 3 to slide along the vertical direction and detects the sudden change feature of the temperature gradient caused by the inversion layer near the top of the boundary layer. This feature is highly correlated with the height of the inflection point of the vertical profile of the surface PM2.5 concentration. Based on this, the system establishes a quantitative correlation model between the dynamic change of the boundary layer height and the vertical diffusion ability of pollutants, and accurately predicts the near-surface pollution accumulation caused by the insufficient mixing layer height of less than 200 meters in the early morning under static and stable weather conditions.

[0089] S133: Concatenate the super-resolution spatial feature map and the correlation features in the channel dimension to generate a multi-modal fusion feature tensor.

[0090] In the embodiment of the present invention, the aerosol prediction system concatenates the super-resolution spatial feature map and the correlation features in the channel dimension to generate a multi-modal fusion feature tensor. The channel dimension concatenation realizes the parallel expression and information complementation of different modal features.

[0091] For example, concatenate the super-resolution concentration feature map with a resolution of 250 m (including 32 feature channels) and the boundary layer correlation features (including 16 channels) in the channel dimension to form a 48-channel fusion feature tensor. In the scenario of the sea breeze front passing through a coastal city, this tensor retains both the fine structure of the sewage outfall depicted by the super-resolution feature map and the vertical transport features extracted by the three-dimensional convolution, providing a multi-modal information basis for the model to take into account both the horizontal diffusion details and the vertical mixing process.

[0092] S134: Slide and intercept the time dimension of the multi-modal fusion feature tensor based on a bidirectional gated recurrent unit to generate cross-modal temporal correlation features.

[0093] In the embodiments of the present invention, the aerosol prediction system slides and intercepts the time dimension of the multi-modal fusion feature tensor based on a bidirectional gated recurrent unit to generate cross-modal temporal correlation features. The bidirectional gated recurrent unit captures temporal dependence relationships through forward and backward time flows.

[0094] For example, when processing the multi-modal fusion feature tensor for 72 consecutive hours, the system slides and intercepts the input sequence with a 6-hour time step. In an event of a cold front passing through, the forward network learns the pollutant accumulation trend during the pre-frontal static and stable stage, and the backward network captures the scavenging effect during the strong wind process after the front. The temporal correlation features generated after splicing the hidden states of both accurately quantify the time lag relationship between the sudden rise in the boundary layer height and the sudden drop in the PM2.5 concentration, providing key temporal features for predicting the turning point of the pollution process.

[0095] S135: Perform spatial position encoding on the cross-modal temporal correlation features and the multi-modal fusion feature tensor to generate a spatio-temporal feature matrix.

[0096] In the embodiments of the present invention, the aerosol prediction system performs spatial position encoding on the cross-modal temporal correlation features and the multi-modal fusion feature tensor to generate a spatio-temporal feature matrix. Spatial position encoding injects absolute and relative position information through the sine position embedding algorithm.

[0097] For example, for each spatial unit of a 500m×500m grid in a certain area, the system generates a 64-dimensional position encoding vector containing longitude and latitude information. This vector is added to the cross-modal temporal features position by position, enabling the spatio-temporal feature matrix to simultaneously contain the evolution law of meteorological elements, the pollutant diffusion path, and the topological constraints of geographical coordinates. In complex terrain areas, position encoding enables the model to automatically distinguish the differences in the impact of different geomorphic units such as valleys and ridges on pollutant transport, improving the spatial rationality of the prediction results.

[0098] In a preferred embodiment, the local texture enhancement of the surface aerosol concentration sequence in S131 by the convolutional attention module to generate a super-resolution spatial feature map includes:

[0099] S1311: Perform Gaussian difference filtering on the surface aerosol concentration sequence to extract the spatial gradient magnitude map.

[0100] In the embodiments of the present invention, the aerosol prediction system performs Gaussian difference filtering on the surface aerosol concentration sequence to extract the spatial gradient magnitude map. Gaussian difference filtering highlights the spatial gradient features through the difference in the convolution results of Gaussian kernels at different scales.

[0101] For example, when processing the aerosol optical depth data retrieved from satellites in a certain river valley area, the system convolves the original 1-km grid data using double Gaussian kernels with σ = 1.5 km and σ = 3 km, and calculates the difference between the two to obtain a gradient magnitude map. This map shows high gradient values in the piedmont area on the west side of the river valley, accurately reflecting the sharp concentration boundary formed by the accumulation of pollutants at the foot of the mountain due to the nocturnal inversion layer. The spatial distribution of the gradient magnitude map is highly consistent with the sudden change events of the hourly PM2.5 concentration measured by ground monitoring stations, verifying the ability of the filtering process to capture the pollutant diffusion front.

[0102] S1312: Generate regional significance weights according to the local variance distribution of the spatial gradient magnitude map.

[0103] In the embodiment of the present invention, the aerosol prediction system generates regional significance weights according to the local variance distribution of the spatial gradient magnitude map. The regional significance weights are used to quantify the feature importance of different spatial positions, and their calculation is based on the statistical characteristics within a sliding window.

[0104] For example, in the central urban area of a certain megacity, the system traverses the gradient magnitude map with a 500 m × 500 m window and calculates the variance of the gradient values within each window. The results show that the window variance at the intersections of the main road network reaches 12.5, which is significantly higher than 2.3 of the residential area windows. Based on this, the system generates a weight coefficient matrix, sets the weight of the intersection area to 0.9, and reduces the weight of the residential area to 0.2. This weight matrix effectively strengthens the spatial identification of traffic source emissions and provides guiding information for subsequent feature extraction.

[0105] S1313: Use deformable convolutional kernels to perform multi-directional feature extraction on the surface aerosol concentration sequence to generate direction-sensitive feature maps.

[0106] In the embodiment of the present invention, the aerosol prediction system uses deformable convolutional kernels to perform multi-directional feature extraction on the surface aerosol concentration sequence to generate direction-sensitive feature maps. The deformable convolutional kernels achieve direction-sensitive feature capture by adaptively adjusting the sampling point positions.

[0107] For example, in the leeward slope area of a certain mountain range, the system initializes the sampling offset parameter set of the deformable convolutional kernel so that it can adjust the receptive field along the direction of the topographic contour line. By calculating the spatial autocorrelation function of the concentration field, the system identifies that the dominant diffusion direction is 30 degrees northeast, and accordingly sets the convolutional kernel to offset 1.2 pixels in the positive X-axis direction and 0.8 pixels in the negative Y-axis direction. The bilinear interpolation layer performs feature mapping on the offset sampling positions, and the finally generated direction-sensitive feature map clearly shows the pollutant retention zone caused by the leeward slope vortex, and its spatial form highly coincides with the low-altitude circulation structure observed by the wind profiler radar.

[0108] S1314: Multiply the regional saliency weight and the direction-sensitive feature map element-wise to generate a texture-enhanced feature map.

[0109] In an embodiment of the present invention, the aerosol prediction system multiplies the regional saliency weight and the direction-sensitive feature map element-wise to generate a texture-enhanced feature map. The regional saliency weight matrix identifies spatial key regions through local variance analysis, and the direction-sensitive feature map is extracted by a deformable convolutional kernel along the dominant diffusion direction.

[0110] For example, around a certain petrochemical park, the system detects that the local variance in the area of the southeast plant boundary reaches 28.5, generating a weight coefficient of 0.92, while the weight in the northwest farmland area is only 0.15. The direction-sensitive feature map shows plume textures of pollutant diffusion along the southeast-northwest axis. After element-wise multiplication of the two, the texture amplitude in the plant boundary area increases from 12.7 to 11.7 (12.7×0.92), while that in the farmland area decreases from 3.2 to 0.48 (3.2×0.15), effectively enhancing the visualization of the industrial point source emission path.

[0111] S1315: Enhance the resolution of the texture-enhanced feature map through a transposed convolutional layer to generate a super-resolution spatial feature map.

[0112] In an embodiment of the present invention, the aerosol prediction system enhances the resolution of the texture-enhanced feature map through a transposed convolutional layer to generate a super-resolution spatial feature map. The transposed convolution realizes resolution enhancement through a learnable upsampling kernel.

[0113] For example, input a texture-enhanced feature map with a resolution of 500m in the suburbs of a certain city into the transposed convolutional layer. The system uses a 4×4 convolutional kernel for 2-fold upsampling to generate a super-resolution feature map with a resolution of 250m. This feature map reveals pollution mass structures at scales of 200 - 300m that cannot be resolved by traditional interpolation methods around key industrial point sources, and the spatial matching degree with the driving observations of high-precision mobile monitoring vehicles reaches 92%, significantly improving the ability to identify local pollution sources.

[0114] In a more specific embodiment, the multi-directional feature extraction of the surface aerosol concentration sequence in S1313 using the deformable convolutional kernel to generate a direction-sensitive feature map includes:

[0115] S13131: Initialize the set of sampling offset parameters of the deformable convolutional kernel.

[0116] In an embodiment of the present invention, the aerosol prediction system multiplies the regional saliency weight and the direction-sensitive feature map element-wise to generate a texture-enhanced feature map. The regional saliency weight matrix identifies spatial key regions through local variance analysis, and the direction-sensitive feature map is extracted by a deformable convolutional kernel along the dominant diffusion direction.

[0117] For example, around a petrochemical industrial park, the system detects that the local variance in the area of the factory boundary on the southeast side reaches 28.5, generating a weight coefficient of 0.92, while the weight in the farmland area on the northwest side is only 0.15. The direction-sensitive feature map shows the plume texture of pollutant diffusion along the southeast-northwest axis. After multiplying the two element by element, the texture amplitude in the factory boundary area increases from 12.7 to 11.7 (12.7×0.92), while that in the farmland area decreases from 3.2 to 0.48 (3.2×0.15), effectively enhancing the visual expression of the industrial point source emission path.

[0118] S13132: Calculate the weight coefficient in each direction according to the spatial autocorrelation function of the surface aerosol concentration sequence.

[0119] In the embodiment of the present invention, the aerosol prediction system calculates the weight coefficient in each direction according to the spatial autocorrelation function of the surface aerosol concentration sequence. The spatial autocorrelation function is used to quantify the concentration change trend in different directions.

[0120] For example, in a valley basin, the system calculates the spatial autocorrelation coefficients in four directions of 0 degree, 45 degrees, 90 degrees, and 135 degrees, and finds that the correlation coefficient in the 45-degree direction (along the valley trend) reaches 0.78, significantly higher than other directions. Accordingly, the system increases the weight coefficient of the deformable convolution kernel in the 45-degree direction to 0.65, and reduces the weights in other directions accordingly, so that the feature extraction process focuses on the dominant diffusion direction guided by the terrain.

[0121] S13133: Perform position-sensitive feature mapping on the sampling offset parameter through a bilinear interpolation layer.

[0122] In the embodiment of the present invention, the aerosol prediction system performs position-sensitive feature mapping on the sampling offset parameter through a bilinear interpolation layer. Bilinear interpolation ensures that continuous spatial features can be obtained at the offset sampling positions.

[0123] For example, when a 1.5-pixel offset occurs at the sampling point in the northwest direction of an industrial park for the deformable convolution kernel, the system takes the target position as the center and performs weighted averaging on the concentration values of the surrounding 4 original grid points. The weights are dynamically calculated according to the geometric relationship between the offset position and the original grid, so that the feature mapping result can not only reflect the direction sensitivity brought by the sampling point offset but also maintain spatial continuity. This process successfully identifies the fan-shaped pollution area affected by the dominant wind direction when capturing the pollutant diffusion front of a sudden leakage event in a chemical industrial park.

[0124] S13134: Multiply the mapped feature by the weight coefficient to generate a direction-sensitive feature map.

[0125] In an embodiment of the present invention, the aerosol prediction system multiplies the mapped features by the weight coefficients to generate a direction-sensitive feature map. The matrix multiplication operation realizes the direction-selective fusion of features.

[0126] For example, in a certain monsoon transition zone, the system assigns weight coefficients of 0.7 and 0.3 to the feature mapping results in the two dominant directions of southeast and northwest respectively. After fusion through matrix multiplication, the generated feature map accurately shows the pollutant transport channels brought by the southeast wind in the afternoon and the local accumulation effect caused by the northwest wind at night. The spatial distribution of the direction-sensitive feature map is physically consistent with the diurnal variation characteristics of the boundary layer height observed by lidar, verifying the effectiveness of the weight allocation strategy.

[0127] In a preferred embodiment, the multi-stage recursive optimization of the spatio-temporal feature matrix by the hierarchical time-domain aggregation algorithm in S140 to generate the aerosol concentration prediction sequence of the target area includes:

[0128] S141: Divide the spatio-temporal feature matrix into a first periodic fluctuation component and a second periodic fluctuation component; wherein, the period length of the first periodic fluctuation component is less than that of the second periodic fluctuation component.

[0129] In S141, the aerosol prediction system divides the spatio-temporal feature matrix into a first periodic fluctuation component and a second periodic fluctuation component. Specifically, the system uses the empirical mode decomposition algorithm to extract the intrinsic mode functions of the time dimension of the spatio-temporal feature matrix. The first periodic fluctuation component corresponds to the short-period changes on the hourly to daily scale (such as the vertical diffusion fluctuations of pollutants caused by the diurnal variation of the boundary layer height), and the second periodic fluctuation component represents the long-period evolution on the weekly to ten-day scale (such as the change in regional transport intensity caused by the adjustment of the atmospheric circulation).

[0130] For example, in a winter heavy pollution case in a certain plain urban agglomeration, after the spatio-temporal feature matrix is decomposed, the first periodic fluctuation component shows an obvious 24-hour periodic oscillation during the period from 08:00 on December 10 to 20:00 on December 11, reflecting the modulation effect of the diurnal rise and fall of the inversion layer on the ground PM2.5 concentration; while the second periodic fluctuation component shows the characteristics of the maintenance of the static-stability weather caused by the continuous southward pressure of the Siberian high since December 5, with a period length of 120 hours, dominating the regional pollution accumulation trend.

[0131] S142: Dynamically attenuate the frequency-domain energy of the first periodic fluctuation component through an adaptive weight allocator.

[0132] In S142, the aerosol prediction system dynamically attenuates the frequency-domain energy of the first periodic fluctuation component through an adaptive weight allocator. The adaptive weight allocator constructs an energy spectral density function based on the wavelet transform coefficients and suppresses the high-frequency noise components according to the preset threshold.

[0133] For example, when processing the first periodic fluctuation component dominated by the diurnal variation of land-sea breeze in a coastal city, the system detects that the turbulent pulsation energy at the 30-minute scale accounts for more than 15% in the frequency domain. By designing a Butterworth digital filter, the energy of the periodic components below 2 hours is attenuated by 40%. At the same time, the system retains the energy distribution in the key frequency band of 4 - 12 hours to ensure the integrity of the characteristics of the sharp drop in pollutant concentration caused by the advancement of the sea breeze front (usually lasting 3 - 5 hours). This processing reduces the standard deviation of high-frequency fluctuations in the downwind area of an industrial park from 28.7 μg / m³ to 16.3 μg / m³, significantly improving the smoothness of the prediction sequence.

[0134] S143: Use the residual connection structure to superimpose the attenuated first periodic fluctuation component and the second periodic fluctuation component to generate an optimized feature vector.

[0135] In S143, the aerosol prediction system uses the residual connection structure to superimpose the attenuated first periodic fluctuation component and the second periodic fluctuation component to generate an optimized feature vector. The residual connection retains the key information of the physical process in the original spatio-temporal feature matrix through cross-layer direct connection.

[0136] For example, when superimposing and processing the optimized features in a valley area, the system linearly superimposes the attenuated 24-hour periodic component (characterizing the diurnal variation of mountain-valley breeze) and the second periodic component at the 10-day scale (reflecting the cold air activity cycle), and introduces the terrain dynamic uplift effect parameter in the original feature matrix through the residual connection. This operation enables the optimized feature vector to retain both the 3-hour scale pollutant removal fluctuation at the moment of frontal passage and inherit the leeward slope pollution retention trend caused by the terrain during the cold front passage event on December 15th, realizing the organic integration of multi-scale features.

[0137] S144: Input the optimized feature vector into the mixture density network to generate the probability distribution parameters of the aerosol concentration.

[0138] In S144, the aerosol prediction system inputs the optimized feature vector into the mixture density network to generate the probability distribution parameters of the aerosol concentration. The mixture density network adopts a multi-branch fully connected structure to model the uncertainty effects of meteorological conditions and pollutant source emissions respectively.

[0139] For example, during the Spring Festival in a certain megacity, the optimized feature vector includes complex factors such as the shutdown of industrial sources and the sharp increase in traffic flow. The mixture density network outputs the weights, means, and variance parameters of three Gaussian distribution components. In the prediction on January 25th (the third day of the first lunar month), the network generates distribution parameters with a mean of 65 μg / m³ and a variance of 12.3 μg² / m 6 to accurately reflect the statistical characteristics of the sharp fluctuations in concentration during the concentrated discharge period of fireworks and firecrackers.

[0140] S145: Perform Monte Carlo sampling on the historical observation sequence according to the probability distribution parameters to generate a predicted sequence of aerosol concentrations in the target area.

[0141] In S145, the aerosol prediction system performs Monte Carlo sampling on the historical observation sequence according to the probability distribution parameters to generate a predicted sequence of aerosol concentrations in the target area. Monte Carlo sampling simulates various possible meteorological-emission scenarios through a random number generator. For example, during the activation of the heavy pollution emergency plan in a certain area, the system performs 500 samplings to generate a probability prediction set. In the prediction on January 15th, the sampling sequence shows that the opening probability of the southwest transmission channel is 63%, and the median PM2.5 concentration reaches 153 μg / m³, with a matching degree of 89% with the spatial transmission path of the subsequent actual pollution process. The 1-hour resolution feature of the predicted sequence successfully captures the concentration peak caused by the accidental emission in an industrial park, providing minute-level early warning support for emergency control.

[0142] In a preferred embodiment, the inputting the optimized feature vector into the mixture density network to generate the probability distribution parameters of the aerosol concentration in S144 includes:

[0143] S1441: Divide the optimized feature vector into a meteorological driving component and a pollution source contribution component.

[0144] In S1441, the aerosol prediction system divides the optimized feature vector into a meteorological driving component and a pollution source contribution component. The division process is based on the correlation analysis between the feature channels and prior physical knowledge. For example, the features related to the three-dimensional wind field and temperature field are classified as the meteorological driving component, while the industrial point source emission intensity and traffic flow characteristics are classified as the pollution source contribution component. In a case of a port city, the meteorological driving component includes the sea breeze front transit time and vertical turbulence intensity parameters, and the pollution source contribution component covers the time series data of the ship emission SO2 concentration and diesel vehicle flow. The system verifies through mutual information quantification and confirms that this division makes the information overlap degree between the two components less than 8%, meeting the requirements of independent modeling.

[0145] S1442: Perform a non-linear transformation on the meteorological driving component through a fully connected layer to generate mean and variance estimates.

[0146] In S1442, the aerosol prediction system performs a non - linear transformation on the meteorological driving components through a fully - connected layer to generate mean and variance estimates. The fully - connected layer uses the LeakyReLU activation function to introduce non - linear mapping ability. For example, it transforms features such as 850hPa wind speed, boundary layer height, and relative humidity into the conditional mean of PM2.5 concentration. In a certain dust transport case, based on the position parameter of the 700hPa jet axis in the meteorological driving components, the system outputs a mean that jumps from the background value of 35μg / m³ to 182μg / m³, and the variance expands synchronously to 45μg² / m 6 , accurately reflecting the concentration mutation characteristics before and after the arrival of the dust front.

[0147] S1443: Use a gated linear unit to model the time - lag effect of the pollution source contribution component to generate a skewness coefficient.

[0148] In S1443, the aerosol prediction system uses a gated linear unit to model the time - lag effect of the pollution source contribution component to generate a skewness coefficient. The gated linear unit captures the cumulative effect of emission source strength through temporal convolution. For example, the change in PM2.5 emissions caused by the start - up and shut - down operations of a sintering machine in a steel enterprise. The system models its 6 - 8 - hour lag effect and generates a positive skewness distribution parameter. In the maintenance event on March 12th, the pollution source contribution component shows that the SO2 emissions drop suddenly by 70%. The gated linear unit outputs a skewness coefficient of - 0.37, accurately characterizing the left - skewed characteristic of the concentration distribution, and the skewness error between the prediction result and the observed value at the ground monitoring station is less than 0.05.

[0149] S1444: Combine the mean, the variance, and the skewness coefficient into a three - parameter log - normal distribution, and generate probability distribution parameters of aerosol concentration through the three - parameter log - normal distribution.

[0150] In S1444, the aerosol prediction system combines the mean, variance, and skewness coefficient into a three - parameter log - normal distribution and generates probability distribution parameters of aerosol concentration. The three - parameter log - normal distribution adapts to the asymmetric concentration distribution by introducing a location parameter. For example, in the case of an accidental emission from a coal - fired power plant resulting in a right - skewed concentration distribution (skewness + 1.2), the system adjusts the location parameter to make the distribution peak coincide with the monitored value. In the prediction on December 28th, the distribution generates a 5 - 95% confidence interval of [48, 215]μg / m³, completely covering the measured maximum value of 209μg / m³ on that day, and the coverage rate of the prediction interval is increased by 22 percentage points compared with the traditional normal distribution.

[0151] In an alternative embodiment, after generating the aerosol concentration prediction sequence of the target area, it further includes:

[0152] S210: Obtain real - time meteorological re - analysis data stream and satellite aerosol optical depth observation values.

[0153] In the embodiment of the present invention, the aerosol prediction system acquires real-time meteorological reanalysis data streams and satellite aerosol optical depth observations. The real-time meteorological reanalysis data stream is output every half hour by the high-resolution numerical model of the European Centre for Medium-Range Weather Forecasts, and includes parameters such as 10-meter wind speed, 2-meter temperature, and boundary layer height of a 0.1° grid in the target area. For example, during the heavy pollution emergency response in a coastal urban agglomeration, the system accesses the ERA5 reanalysis data stream from 08:00 on December 15th to 08:00 on December 16th in real time, and synchronously acquires the 550nm aerosol optical depth observations updated every 10 minutes by the Himawari-8 geostationary satellite. After the satellite data is geographically corrected, it forms spatial complementarity with the PM2.5 hourly concentration data of ground monitoring stations, covering the traditional monitoring blind spots in the offshore ship emission area.

[0154] S220: Input the real-time meteorological reanalysis data stream into a pre-trained spatio-temporal feature fusion network to extract a real-time spatio-temporal feature matrix.

[0155] In the embodiment of the present invention, the aerosol prediction system inputs the real-time meteorological reanalysis data stream into a pre-trained spatio-temporal feature fusion network to extract a real-time spatio-temporal feature matrix. The spatio-temporal feature fusion network uses the same architecture parameters as in the offline training stage to perform end-to-end feature encoding on the real-time input data. For example, when processing the real-time data at 06:00 on December 16th, the three-dimensional dilated convolutional layer in the network captures the feature of the sudden rise in the boundary layer height caused by the passage of a cold air front, and its vertical profile shows that the mixed layer height rapidly rises from 200 meters at 06:00 to 850 meters at 06:30. At the same time, the Vision Transformer module identifies the interaction between the urban heat island circulation and the sea breeze circulation, and generates a 64-dimensional feature vector representing the advancing speed of the sea breeze front in the real-time spatio-temporal feature matrix.

[0156] S230: Dynamically align the real-time spatio-temporal feature matrix with the historical spatio-temporal feature matrix through a sliding time window to obtain a dynamic alignment result.

[0157] In an embodiment of the present invention, the aerosol prediction system dynamically aligns the real-time spatio-temporal feature matrix with the historical spatio-temporal feature matrix through a sliding time window to obtain a dynamic alignment result. The sliding time window is based on the current moment and slides forward by 6 hours to construct a time series segment. For example, in the real-time prediction task at 07:00 on December 16th, the system selects the spatio-temporal feature matrix from 07:00 to 13:00 on December 10th, 2021 (corresponding to a cold front passing process) in the historical database, and non-linearly aligns the 06:00 - 12:00 segment of the real-time data with the 07:00 - 13:00 segment of the historical data through the dynamic time warping algorithm. The alignment result shows that the current frontal movement speed is 1.2 times faster than the historical case. Based on this, the system adjusts the time offset, reducing the timestamp deviation of the front reaching the coastline from 38 minutes to within 5 minutes.

[0158] S240: Based on the dynamic alignment result, use the Kalman filter algorithm to iteratively update the error covariance matrix of the aerosol concentration prediction sequence.

[0159] In an embodiment of the present invention, the aerosol prediction system iteratively updates the error covariance matrix of the aerosol concentration prediction sequence based on the dynamic alignment result using the Kalman filter algorithm. The Kalman filter algorithm fuses the statistical characteristics of the predicted value and the observed value through a state space model. For example, a difference analysis is performed on the satellite aerosol optical depth observation value (0.82 ± 0.15) at 06:30 on December 16th and the model predicted value (0.79 ± 0.21). The system calculates the spatial distribution characteristics of the prediction error at the current moment and finds that the root mean square error in the downwind area of the industrial park reaches 28 μg / m³, which is significantly higher than other areas. Based on this, a directional update mechanism for the covariance matrix is activated.

[0160] S250: Online correct the aerosol concentration prediction sequence of the target area according to the iteratively updated error covariance matrix.

[0161] In an embodiment of the present invention, the aerosol prediction system online corrects the aerosol concentration prediction sequence of the target area according to the iteratively updated error covariance matrix. After being updated by the Kalman filter, the error covariance matrix reflects the prediction uncertainty distribution of each spatial grid.

[0162] For example, the predicted error variance matrix at 09:00 in a certain port city shows that the variance in the container terminal area reaches , while that in the residential area is only 。Based on the updated Kalman gain coefficients (0.68 for the dock area and 0.32 for the residential area), the system linearly fuses the initial predicted value of 135 μg / m³ in the dock area with the observed value of 148 μg / m³ to obtain a corrected value of 135×0.32 + 148×0.68 = 143.2 μg / m³, reducing the absolute error from the ship emission monitoring value of 145 μg / m³ at 09:10 to 1.8 μg / m³. The root mean square error of the corrected prediction sequence for the entire area is reduced from 18.7 μg / m³ to 9.3 μg / m³.

[0163] In a preferred embodiment, the iterative update of the error covariance matrix of the aerosol concentration prediction sequence in S240 includes:

[0164] S241: Generate parameterized expressions for the aerosol concentration observation model and the state transition model.

[0165] In S241, the aerosol prediction system generates parameterized expressions for the aerosol concentration observation model and the state transition model. The observation model establishes a linear regression relationship between the PM2.5 concentration and the satellite aerosol optical depth. For example, the conversion coefficient calibrated based on the 3-year observation data of a certain urban agglomeration is 89.3 μg / (m³·AOD). The state transition model describes the temporal evolution of the pollutant concentration through an autoregressive equation, and its order is determined to be 3 according to the Bayesian information criterion, reflecting that the weights of the influence of the concentrations in the previous three hours on the current moment are 0.55, 0.32, and 0.13 respectively. The model parameters are updated hourly to adapt to the dynamic changes in meteorological conditions.

[0166] S242: Determine the innovation covariance matrix based on the real-time satellite aerosol optical depth observation values.

[0167] In S242, the aerosol prediction system determines the innovation covariance matrix based on the real-time satellite aerosol optical depth observation values. The innovation covariance matrix characterizes the statistical characteristics of the difference between the observed value and the predicted value, and its calculation is based on the historical residual data within a sliding window. For example, when the Himawari-8 satellite passed by at 07:10 on December 16th, the system compared the predicted concentration (mean value of 112 μg / m³) and the retrieved concentration (mean value of 127 μg / m³) of 15 grids in the coastal area, and calculated the regional innovation variance as , and the diagonal elements of the covariance matrix show that the error variance in the port area is as high as , indicating that this area needs to be corrected with emphasis.

[0168] S243: Detect the positive definiteness of the innovation covariance matrix through Cholesky decomposition.

[0169] In S243, the aerosol prediction system detects the positive definiteness of the innovation covariance matrix through Cholesky decomposition. Cholesky decomposition requires the matrix to be symmetric and positive definite, and decomposition failure may occur when numerical calculation errors or abnormal observation noises appear. For example, in the update cycle at 07:30 on December 16th, the system performed decomposition on the 6×6 innovation covariance matrix in the port area and detected that the element in the 3rd row and 3rd column made the matrix non-positive definite due to the sudden fluctuation of ship emissions. The system recorded the minimum eigenvalue during the decomposition process as -0.17, triggering the regularization processing flow.

[0170] S244: When it is detected that the innovation covariance matrix is non-positive definite, a regularization factor is used to perform diagonal loading on the prior estimate to obtain the target covariance matrix.

[0171] In S244, when the aerosol prediction system detects that the innovation covariance matrix is non-positive definite, it uses a regularization factor to perform diagonal loading on the prior estimate to obtain the target covariance matrix. Diagonal loading ensures the positive definiteness of the matrix by adding a small positive value to the diagonal of the covariance matrix. For example, for the above port area problem, the system selects the regularization factor λ = 0.05 and increases each diagonal element of the original covariance matrix by 5% of its absolute value. The minimum eigenvalue of the processed matrix is increased to 0.12, and it successfully passes the Cholesky decomposition verification, forming a target covariance matrix that can be used for Kalman gain calculation.

[0172] S245: Update the Kalman gain coefficient according to the target covariance matrix to generate a corrected aerosol concentration prediction sequence.

[0173] In S245, the aerosol prediction system updates the Kalman gain coefficient according to the target covariance matrix to generate a corrected aerosol concentration prediction sequence. The Kalman gain coefficient determines the correction weight of the observed value to the prediction result, and its calculation is based on the updated covariance matrix.

[0174] For example, the innovation covariance in the port area is , and the prediction error variance is . The calculated Kalman gain coefficient is 0.47. The system applies this coefficient to the prediction value at 08:00 on December 16th, correcting the original prediction value of 135 μg / m³ to 135 + 0.47×(127 - 135) = 131.2 μg / m³. The deviation from the measured value of 130 μg / m³ at the ground monitoring station at 08:10 is reduced from 5 μg / m³ to 1.2 μg / m³. After full-region iterative update, in the verification at 09:00 on December 16th, the spatial correlation coefficient of the finally output corrected prediction sequence is increased from 0.82 to 0.91, and the root mean square error is reduced by 32%, achieving online optimization of the prediction accuracy.

[0175] In a preferred embodiment, the online correction of the aerosol concentration prediction sequence of the target area according to the iteratively updated error covariance matrix in S250 includes:

[0176] S251: Obtain the aerosol concentration observation value at the current moment in the real-time meteorological reanalysis data stream and the initial prediction value of the aerosol concentration prediction sequence.

[0177] In S251, the aerosol prediction system obtains the aerosol concentration observation value at the current moment in the real-time meteorological reanalysis data stream and the initial prediction value of the aerosol concentration prediction sequence. Specifically, the system extracts the boundary layer height and three-dimensional wind speed field data at 08:00 on December 16 from the real-time data interface of the European Centre for Medium-Range Weather Forecasts and synchronously accesses the PM2.5 hourly average observation data of the ground environmental monitoring station network. For example, at the monitoring site downwind of the industrial park in a port city, the measured PM2.5 concentration at the current moment is 138 μg / m³, while the initial prediction value is 125 μg / m³. The system records the spatial distribution difference between the observation value and the prediction value at this moment as the benchmark input for subsequent correction.

[0178] S252: Generate a weight allocation ratio for the aerosol concentration observation value at the current moment according to the Kalman gain coefficient of the error covariance matrix; perform linear weighted fusion on the initial prediction value and the aerosol concentration observation value based on the weight allocation ratio to generate the corrected aerosol concentration value at the current moment.

[0179] In S252, the aerosol prediction system generates a weight allocation ratio for the aerosol concentration observation value at the current moment according to the Kalman gain coefficient of the error covariance matrix and performs linear weighted fusion on the initial prediction value and the observation value. The Kalman gain coefficient is determined by the inverse operation of the covariance matrix. For example, the gain coefficient calculated in the port area is 0.62, indicating that the weight of the observation value accounts for 62%. The system applies this coefficient to the initial prediction value of 125 μg / m³ and the observation value of 138 μg / m³ at 08:00 to generate a corrected concentration value of 125×0.38 + 138×0.62 = 132.9 μg / m³. This value reduces the initial prediction deviation by 67%, and the absolute error from the measured value of 133 μg / m³ by the 08:10 minute-level mobile monitoring vehicle is less than 0.5 μg / m³.

[0180] S253: Extract the residual component of the corrected aerosol concentration value and the spatial distribution residual of the satellite aerosol optical depth observation value; perform feature mapping on the spatial distribution residual through the spatio-temporal feature fusion network to generate a residual spatio-temporal correlation matrix.

[0181] In S253, the aerosol prediction system extracts the spatial distribution residuals of the residual component of the corrected aerosol concentration value and the satellite aerosol optical depth observation value, and generates a residual spatio-temporal correlation matrix through a spatio-temporal feature fusion network. The residual component is calculated from the spatial difference between the corrected prediction value and the ground measured value. For example, the residual at 09:00 in an industrial park is +15.2 μg / m³, while the satellite inversion residual is -0.23 AOD. The system inputs both into the spatio-temporal fusion network. The three-dimensional convolutional layer in the network detects an abnormal boundary layer height in the vertical direction (actual value 850m vs predicted value 720m), generating a 32-dimensional feature vector representing insufficient vertical diffusion ability. The attention mechanism module strengthens the residual correlation within a 5-km range downwind of the industrial park. The finally output residual spatio-temporal correlation matrix shows a residual positive correlation area with a duration of 3 hours and an expanding spatial range in this region, providing directional feedback for subsequent model optimization.

[0182] S254: Input the residual spatio-temporal correlation matrix into the hierarchical time-domain aggregation algorithm to extract the periodic fluctuation pattern of the residual component; dynamically adjust the time decay factor of the Kalman gain coefficient according to the periodic fluctuation pattern.

[0183] In S254, the aerosol prediction system dynamically adjusts the time decay factor of the Kalman gain coefficient according to the periodic fluctuation pattern. The time decay factor controls the decay rate of the contribution of historical data to the current correction. For example, after detecting a quasi-period of 1.5 hours, the system adjusts the decay factor from 0.85 to 0.72, extending the effective historical window to 3 cycles (4.5 hours). During the correction process at 08:30, this adjustment increases the weight of the residual pattern during the morning traffic peak at 07:00 by 15%, improving the prediction adaptability to the current evening traffic peak.

[0184] S255: Update the sliding window length of the error covariance matrix using the adjusted time decay factor; perform a moving average filter on the aerosol concentration prediction sequence at the next moment based on the updated sliding window length to generate a smoothed aerosol concentration prediction value.

[0185] In S255, the aerosol prediction system updates the sliding window length of the error covariance matrix using the adjusted time decay factor and performs a moving average filter on the prediction sequence at the next moment. The sliding window expands from the default 6 hours to 8 hours, covering 3 complete diurnal cycles of the boundary layer. The system performs a weighted average of 12 samples within the window on the 09:00 prediction value, where the weight of the 08:00 correction value accounts for 41%. The standard deviation of the filtered prediction value in the residential area drops from 18.7 μg / m³ to 9.3 μg / m³, effectively suppressing short-term fluctuations caused by meteorological mutations.

[0186] S256: Concatenate the smoothed aerosol concentration prediction value and the corrected aerosol concentration value in a time series to generate an online corrected aerosol concentration prediction sequence for the target area.

[0187] In S256, the aerosol prediction system concatenates the smoothed aerosol concentration prediction value and the corrected aerosol concentration value in a time series to generate an online corrected aerosol concentration prediction sequence for the target area. For example, the corrected value of 132.9 μg / m³ at 08:00 and the smoothed prediction value of 128.4 μg / m³ at 09:00 are concatenated to form a continuous sequence. In the verification of the ground monitoring value of 129 μg / m³ at 09:30, the absolute error of this sequence is 0.6 μg / m³, and the spatial correlation coefficient remains above 0.93, achieving the optimization of the temporal coherence and spatial consistency of the prediction results.

[0188] In a more specific embodiment, the generating, by the spatio-temporal feature fusion network, of a residual spatio-temporal correlation matrix by feature mapping of the spatial distribution residual in S253 includes:

[0189] S2531: Decompose the spatial distribution residual into a multi-channel residual tensor of a spatial residual component and a temporal residual component; perform a multi-scale dilated convolution operation on the spatial residual component of the multi-channel residual tensor to extract spatial gradient mutation features under different receptive fields.

[0190] In S2531, the aerosol prediction system decomposes the spatial distribution residual into a multi-channel residual tensor of a spatial residual component and a temporal residual component, and performs a multi-scale dilated convolution operation on the spatial residual component. The spatial residual component reflects the spatial difference between the corrected concentration field and the satellite inversion value. For example, the residual in the coastal area is +5.2 μg / m³ while that in the inland industrial area is -3.8 μg / m³. The system uses three groups of dilated convolution kernels with dilation rates of 2, 4, and 8 to extract the pollutant diffusion front features at scales of 500 m, 1 km, and 2 km in the port area respectively. In the data at 08:00 on December 16th, the convolution kernel with a dilation rate of 4 captures the 1-km scale concentration gradient mutation band caused by the advancement of the sea breeze front, and its spatial gradient amplitude reaches 28 μg / (m³·km).

[0191] S2532: Generate a spatial attention weight matrix according to the local direction consistency of the spatial gradient mutation features.

[0192] In S2532, the aerosol prediction system generates a spatial attention weight matrix based on the local direction consistency of the spatial gradient mutation features. The local direction consistency is determined by calculating the angular variance of the gradient vectors within a 3×3 window. For example, the gradient direction variance in a certain transportation hub area is 15 degrees, generating a weight coefficient of 0.92; while the gradient direction in the farmland area is disordered (variance of 85 degrees), and the weight drops to 0.31. The system multiplies this weight matrix with the multi-scale gradient features to strengthen the residual feature expression of the dominant diffusion direction and weaken the interference of random fluctuations.

[0193] S2533: Perform a temporal causal convolution operation on the temporal residual components of the multi-channel residual tensor to extract the lag time correlation features; generate a temporal convolution feature map based on the periodic intensity distribution of the lag time correlation features.

[0194] In S2533, the aerosol prediction system performs a temporal causal convolution operation on the temporal residual components of the multi-channel residual tensor to extract the lag time correlation features. The temporal causal convolution uses a convolution kernel with a width of 5 and only accesses the data at the current and historical moments. For example, when processing the temporal residual at 08:00, the convolution kernel detects that the residual has been growing positively continuously during the period from 07:30 to 07:50, generating a strongly correlated feature with a lag of 2 time steps. The system further calculates the periodic intensity of this feature and finds that the residual fluctuation period corresponding to the hourly traffic peak is 24 hours, and accordingly generates a temporal convolution feature map.

[0195] S2534: Perform cross-modal feature splicing on the spatial attention weight matrix and the temporal convolution feature map to generate a spatio-temporal cross-correlation feature block; perform adaptive normalization processing on the spatio-temporal cross-correlation feature block in the channel dimension to generate a normalized spatio-temporal feature vector; model the forward and backward dependence relationships of the normalized spatio-temporal feature vector through a bidirectional long short-term memory network to extract global temporal dependence features.

[0196] In S2534, the aerosol prediction system performs cross-modal feature splicing on the spatial attention weight matrix and the temporal convolution feature map to generate a spatio-temporal cross-correlation feature block. The spliced feature block contains 32 spatial channels and 16 temporal channels, and the system performs adaptive normalization processing on it in the channel dimension. For example, the maximum value of the spatial channels in the port area is scaled to 0.87, and the mean value of the temporal channels is adjusted to 0.12 to eliminate the dimension difference. The normalized feature vector is input into the bidirectional long short-term memory network to capture the forward cumulative and backward propagation effects of the residual evolution.

[0197] S2535: Aggregate the topological structure of the spatial adjacency relationship of the global temporal dependence features using a graph convolutional layer to generate spatial node embedding features; perform a feature addition operation on the spatial node embedding features and the global temporal dependence features to generate an enhanced spatio-temporal feature sequence; generate a three-dimensional feature encoding vector based on the spatio-temporal dimension distribution of the enhanced spatio-temporal feature sequence.

[0198] In S2535, the aerosol prediction system aggregates the topological structure of the spatial adjacency relationship of the global temporal dependence features using a graph convolutional layer. The graph convolution constructs a spatial adjacency matrix based on Delaunay triangulation. For example, monitoring points in ports, industrial areas, and residential areas are used as graph nodes, and the pollutant transmission probability between nodes is calculated as the edge weight. The system aggregates neighborhood features through 3 layers of graph convolution, and the generated spatial node embedding features show that the residual contribution rate of the industrial area to the downwind residential area reaches 73%. This feature is added to the temporal features output by the bidirectional long short-term memory network to form an enhanced spatio-temporal feature sequence.

[0199] S2536: Perform a non-linear projection on the three-dimensional feature encoding vector through a multi-layer perceptron to generate a residual spatio-temporal correlation matrix; wherein, each element of the residual spatio-temporal correlation matrix represents the residual correlation strength between different spatial positions and time steps.

[0200] In S2536, the aerosol prediction system performs a non-linear projection on the three-dimensional feature encoding vector through a multi-layer perceptron to generate a residual spatio-temporal correlation matrix. The multi-layer perceptron contains 128 hidden units and uses the GeLU activation function. The elements of the projected matrix represent the residual correlation strength between different spatial positions and time steps. For example, the correlation coefficient between the residual at 08:00 in the port area and the residual in the industrial area at 07:30 is 0.68, while the correlation coefficient with the residual in the residential area at 07:00 is only 0.12. This matrix accurately quantifies the spatio-temporal delay effect of the pollution transmission path.

[0201] S2537: Input the residual spatio-temporal correlation matrix into the hierarchical time-domain aggregation algorithm to extract the periodic fluctuation pattern of the residual component.

[0202] In S2537, the aerosol prediction system inputs the residual spatio-temporal correlation matrix into the hierarchical time-domain aggregation algorithm to extract the periodic fluctuation pattern of the residual component. The algorithm uses empirical wavelet transform to decompose 3 intrinsic mode functions, where the 0.5-hour scale mode reflects traffic emission pulses, and the 24-hour mode corresponds to the daily variation period of the boundary layer. The system identifies that there is a 1.5-hour quasi-periodic fluctuation in the residual of the industrial area during the period from 08:00 to 09:00, and adjusts the time decay factor of the Kalman gain accordingly.

[0203] As an optional but non-limiting embodiment, the step of performing spatial position encoding on the cross-modal temporal correlation feature and the multimodal fusion feature tensor in S135 to generate a spatiotemporal feature matrix includes:

[0204] S1351: Decompose the cross-modal temporal correlation features into a temporal correlation high-frequency component and a temporal correlation low-frequency component.

[0205] S1352: Generate a two-dimensional position encoding matrix according to the spatial dimension distribution of the multimodal fusion feature tensor.

[0206] S1353: Superimpose the time-series associated high-frequency component and the two-dimensional position coding matrix channel by channel to generate a high-frequency enhanced position feature map; perform a three-dimensional separable convolution operation on the high-frequency enhanced position feature map to extract spatial high-frequency texture features.

[0207] S1354: Perform feature channel splicing on the temporal correlation low-frequency component and the multimodal fusion feature tensor to generate a low-frequency fusion feature block; perform a cross-channel attention weighted operation on the low-frequency fusion feature block to generate a spatial low-frequency correlation feature map.

[0208] S1355: Perform multi-scale feature fusion on the spatial high-frequency texture feature and the spatial low-frequency associated feature map to generate a multi-resolution spatial coding feature; generate a dynamic channel weight vector according to the number of channels of the multi-resolution spatial coding feature; use the dynamic channel weight vector to perform channel dimension weighted summation on the multi-resolution spatial coding feature to generate a weighted spatial coding feature map.

[0209] S1356: Expand the weighted spatial coding feature map into a spatiotemporal coding sequence along the time dimension; perform a bidirectional dilated convolution operation on the spatiotemporal coding sequence to extract forward and backward temporal dependency features; splice the forward temporal dependency features and the backward temporal dependency features in the time dimension to generate a full temporal correlation feature vector.

[0210] S1357: Perform a spatial dimension reorganization operation on the full temporal correlation feature vector to restore the original spatial resolution; perform a residual connection on the reorganized full temporal correlation feature vector and the weighted spatial coding feature map to generate an enhanced spatiotemporal coding matrix.

[0211] S1358: Perform a local response normalization operation on the enhanced spatiotemporal coding matrix to generate a normalized spatiotemporal coding tensor; input the normalized spatiotemporal coding tensor into a multilayer perceptron for nonlinear projection to generate an initial estimate of the spatiotemporal feature matrix.

[0212] S1359: Detect the region where the spatial continuity of the initial estimate of the spatio-temporal feature matrix is interrupted; generate a spatial interpolation mask template according to the geometric center coordinates of the interrupted region; use the spatial interpolation mask template to perform edge-guided repair on the initial estimate of the spatio-temporal feature matrix to generate the final spatio-temporal feature matrix.

[0213] In S1351, the aerosol prediction system decomposes the cross-modal temporal correlation features into temporal correlation high-frequency components and temporal correlation low-frequency components. Specifically, the system uses empirical wavelet transform to perform frequency-domain decomposition on the temporal features. The high-frequency components capture the rapid changes at the minute-to-hour scale (such as the instantaneous concentration fluctuations caused by turbulent pulsations), and the low-frequency components characterize the slow evolution above the inter-day scale (such as the adjustment of the atmospheric circulation background field). For example, during the advancement of the sea breeze front in a port city, the high-frequency components show fluctuations with a 10-minute period in the 08:30 - 09:00 period, reflecting the rapid drop and recovery process of PM2.5 concentration caused by the passing of the gust front; the low-frequency components show the continuously strengthening southeast wind background field since December 15th, dominating the offshore transport trend of pollutants.

[0214] In S1352, the aerosol prediction system generates a two-dimensional position encoding matrix according to the spatial dimension distribution of the multi-modal fusion feature tensor. The position encoding matrix uses the sine position embedding algorithm to convert geographical coordinates into a high-dimensional vector space representation. For example, for a 500m×500m UTM grid in a certain urban agglomeration, the system generates a 64-dimensional position encoding vector containing information of longitude 48.25°E and latitude 32.17°N for each grid cell. This matrix automatically distinguishes land and sea grids in the coastal area, and the cosine similarity of the encoding vectors is lower than 0.2, ensuring that the model can identify the blocking effect of the coastline on pollutant diffusion.

[0215] In S1353, the aerosol prediction system stacks the temporal correlation high-frequency components and the two-dimensional position encoding matrix channel by channel to generate a high-frequency enhanced position feature map, and performs a three-dimensional separable convolution operation to extract spatial high-frequency texture features. The three-dimensional separable convolution reduces the number of parameters through the cascade of depth convolution and pointwise convolution. For example, a 3×3×3 convolution kernel is used to slide along the spatial-frequency domain dimension. In the case of a sudden emission event in an industrial park, this operation captures the serrated diffusion edge of the pollution plume along the dominant wind direction at a 1km scale, and its spatial gradient amplitude is increased by 23% compared with traditional convolution, accurately reflecting the concentration mutation characteristics around the factory boundary.

[0216] In S1354, the aerosol prediction system splices the temporal correlation low-frequency component and the multi-modal fusion feature tensor in channels to generate a low-frequency fusion feature block, and performs a cross-channel attention weighting operation to generate a spatial low-frequency correlation feature map. The cross-channel attention mechanism dynamically assigns weights by calculating the correlation between channels. For example, the attention weights of the boundary layer height feature channel and the three-dimensional wind field channel are increased to 0.75. In a certain valley area, this feature map clearly shows the accumulation zone of pollutants at the foot of the mountain caused by the development of the nocturnal inversion layer, and its spatial form is highly consistent with the vertical distribution of pollutants observed by the ground lidar.

[0217] In S1355, the aerosol prediction system performs multi-scale feature fusion on the spatial high-frequency texture feature and the spatial low-frequency correlation feature map to generate a multi-resolution spatial encoded feature, and performs weighted summation through a dynamic channel weight vector. The multi-scale fusion adopts a pyramid pooling structure. For example, features are extracted at three scales of 1 km, 2 km, and 4 km and upsampled to the original resolution. The dynamic weights are adjusted according to the information entropy of the feature channels. The weight of the 1 km fine-grained feature in a certain transportation hub area reaches 0.68, which is significantly higher than 0.12 of the 4 km regional background field, ensuring that the model focuses on the fine structure of the key pollution source area.

[0218] In S1356, the aerosol prediction system unfolds the weighted spatial encoded feature map along the time dimension into a spatio-temporal encoded sequence, and performs a bidirectional dilated convolution operation to extract a full-time correlation feature vector. The bidirectional dilated convolution uses a causal convolution kernel with a dilation rate of 2. The forward flow captures historical dependencies, and the backward flow models future trends. For example, when processing the sequence from 08:00 to 09:00 on December 16th, the forward convolution identifies the scavenging effect triggered by the cold front passing at 07:50, and the backward convolution predicts the pollution backflow caused by the decline of the sea breeze at 09:10. The feature vector generated after splicing the two accurately quantifies the impact period of 3 hours before and after the frontal passage.

[0219] In S1357, the aerosol prediction system performs a spatial dimension reorganization operation on the full-time correlation feature vector to restore the original resolution, and generates an enhanced spatio-temporal encoding matrix through residual connection with the weighted spatial encoded feature map. The spatial reorganization is achieved through transposed convolution. For example, a feature tensor of 64×64×256 is converted into a matrix of 512×512×128. The residual connection retains the topographic dynamic uplift parameters in the original encoding, so that the vortex retention feature in the leeward slope area of a certain mountain range is doubly enhanced in the enhanced matrix, and its activation value is increased by 1.8 times compared with the baseline.

[0220] In S1358, the aerosol prediction system performs local response normalization on the enhanced spatio-temporal coding matrix and inputs it into a multi-layer perceptron to generate an initial estimate of the spatio-temporal feature matrix. Local response normalization suppresses the interference of channels with excessively high activation values. For example, it compresses the maximum value of the channels around a certain industrial point source from 12.7 to 4.3. The multi-layer perceptron uses the GeLU activation function for non-linear projection, mapping the 256-dimensional input to a 128-dimensional feature space. The generated feature matrix accurately represents the coupling effect of ship emissions and sea breeze transportation in the port area.

[0221] In S1359, the aerosol prediction system detects the regions with interrupted spatial continuity in the initial estimate of the spatio-temporal feature matrix and generates a spatial interpolation mask template for edge-guided repair. The interrupted regions refer to the bands of eigenvalue mutations caused by data loss or model errors. For example, a 500m×500m blank area appears in a certain cross-sea bridge area due to the failure of satellite inversion. The system extracts the minimum bounding rectangle of the interrupted region through morphological operations and generates a direction-sensitive interpolation mask along the bridge body direction to ensure that the repair result conforms to the actual diffusion path.

[0222] It can be understood that S1351 - S1359 significantly improve the spatio-temporal modeling ability of aerosol concentration prediction through multi-scale spatio-temporal feature fusion and dynamic coding optimization. First, by decomposing the high-frequency and low-frequency components of the temporal features, the minute-level fluctuations and the daily evolution laws are captured respectively, and the geographical coordinate constraints are strengthened by combining spatial position coding. Three-dimensional separable convolution is used to extract spatial high-frequency textures, and cross-channel attention weighted aggregation is used to combine low-frequency correlation features to achieve multi-resolution feature fusion. The dynamic channel weight vector adaptively allocates the contribution degrees of different-scale features, and bidirectional dilated convolution is used to model the full-time series dependence to enhance the response ability to meteorological mutation events. Residual connections retain the details of the original physical process, local response normalization suppresses noise interference, and the multi-layer perceptron completes non-linear projection, finally generating a high-dimensional spatio-temporal feature matrix. By iteratively detecting and repairing the regions with interrupted spatial continuity, using edge-gradient-guided interpolation masks and Poisson reconstruction, the artifacts caused by data loss or model errors are eliminated, ensuring the spatial coherence and physical consistency of the feature matrix and providing high-precision input for subsequent predictions.

[0223] As an optional but non-limiting embodiment, the edge-guided repair of the initial estimate of the spatio-temporal feature matrix using the spatial interpolation mask template in S1359 to generate the final spatio-temporal feature matrix includes:

[0224] S13591: Detect the regions with interrupted spatial continuity in the initial estimate of the spatio-temporal feature matrix, and generate an edge gradient magnitude map according to the pixel distribution of the interrupted region; perform a morphological closing operation on the edge gradient magnitude map to fill the holes, generate a closed edge contour, and extract the minimum bounding rectangle of the closed edge contour.

[0225] S13592: Determine the geometric center coordinates of the interruption area based on the intersection point of the diagonals of the minimum circumscribed rectangle, and generate a radial basis interpolation weight distribution map according to the geometric center coordinates; Multiply the radial basis interpolation weight distribution map and the spatial interpolation mask template pixel by pixel to generate a direction-sensitive interpolation mask.

[0226] S13593: Perform adaptive anisotropic interpolation on the neighborhood pixels of the interruption area according to the weight values of the pixels in the direction-sensitive interpolation mask to generate a preliminary repaired feature map; Extract the set of pixel points overlapping with the closed edge contour in the preliminary repaired feature map, and calculate the local texture consistency measure of the pixel point set based on the direction angle of the edge gradient magnitude map.

[0227] S13594: Adjust the weight distribution of the direction-sensitive interpolation mask according to the local texture consistency measure to generate an optimized interpolation mask; Use the optimized interpolation mask to perform edge-guided Poisson equation reconstruction on the preliminary repaired feature map to generate an intermediate repaired feature map.

[0228] S13595: Perform non-local mean filtering on the intermediate repaired feature map to suppress interpolation noise, and calculate the residual energy between the filtered feature map and the initial estimate of the spatio-temporal feature matrix; Dynamically adjust the filtering intensity parameter according to the spatial distribution of the residual energy to generate a smoothed repaired feature map.

[0229] S13596: Replace the pixel values corresponding to the interruption area in the smoothed repaired feature map into the initial estimate of the spatio-temporal feature matrix to generate an updated spatio-temporal feature matrix; Detect the remaining spatial discontinuity areas in the updated spatio-temporal feature matrix, and iteratively execute the steps from generating the edge gradient magnitude map to replacing the smoothed repaired feature map until all interruption areas are repaired.

[0230] S13597: Use the updated spatio-temporal feature matrix generated in the last iteration as the final spatio-temporal feature matrix.

[0231] In S13591, the aerosol prediction system detects the spatially continuous interruption areas in the initial estimate and generates an edge gradient magnitude map, and performs a morphological closing operation to extract the closed edge contour. For example, a 200m-wide strip-shaped interruption occurs on the west side of an industrial park due to a monitoring station failure, and the gradient magnitude map shows concentration gradient mutation edges on both the north and south sides. A 5×5 circular structuring element is used for the closing operation to fill the internal holes and extract the minimum circumscribed rectangle of the contour, accurately locating the geometric center of the interruption area.

[0232] In S13592, the aerosol prediction system generates a radial basis interpolation weight distribution map based on the geometric center coordinates, and multiplies it with the spatial interpolation mask to generate a direction-sensitive interpolation mask. The radial basis function takes the geometric center as the origin and distributes weights according to the exponential decay law. The mask in a certain valley interruption area has its weight increased to 0.9 in the downwind direction of the dominant wind and decreased to 0.3 in the upwind direction, ensuring that the interpolation process conforms to the directionality of pollutant transport.

[0233] In S13593, the aerosol prediction system performs adaptive anisotropic interpolation according to the direction-sensitive mask to generate a preliminary repaired feature map, and calculates the local texture consistency metric. Anisotropic interpolation extends the sampling distance along the gradient direction. For example, in a road network interruption area, an interpolation kernel with an aspect ratio of 3:1 is set along the vehicle flow direction. The texture consistency metric identifies artifact regions for secondary correction by calculating the variance of the gradient direction within a 30° sector window.

[0234] In S13594, the aerosol prediction system adjusts the weight distribution of the interpolation mask and performs Poisson equation reconstruction to generate an intermediate repaired feature map. Poisson reconstruction uses the normal area as the boundary condition and solves the Laplace equation within the interruption area. For example, after the interruption repair in a certain residential area, the eigenvalue transition is smooth and continuous with the concentration field gradient of the surrounding area, and the maximum residual decreases from 35 μg / m³ to 7 μg / m³.

[0235] In S13595, the aerosol prediction system performs non-local mean filtering on the intermediate repaired feature map to suppress noise and dynamically adjusts the filtering intensity to generate a smooth repaired feature map. Non-local mean performs weighted averaging by searching for similar image patches. After the repair in an industrial area, the checkerboard artifacts generated by interpolation are effectively eliminated, and the peak signal-to-noise ratio is increased by 8.6 dB.

[0236] In S13596, the aerosol prediction system replaces the smooth repaired feature with the initial estimate and detects the remaining interruption areas, and iteratively performs repair until all discontinuities are eliminated. For example, after 3 iterations in a certain coastal area, the area of the interruption area is reduced from the initial 15 km² to 0.2 km², and the spatial continuity index reaches 0.98.

[0237] In S13597, the aerosol prediction system takes the result of the last iteration as the final spatio-temporal feature matrix. In the verification at 09:00 on December 16th, the spatial correlation coefficient of this matrix with the drone traverse monitoring data reaches 0.95, and the root mean square error is lower than 5 μg / m³, completely retaining the fine structure of the pollution source and the regional transport characteristics, providing high-precision input for subsequent predictions.

[0238] Designed in this way, the above S13591 - S13597 propose an edge - guided iterative repair mechanism for the spatially discontinuous regions in the spatio - temporal feature matrix, effectively improving the geometric rationality and robustness of the prediction results. By performing morphological closing operations to extract the closed contours of the interrupted regions, combining radial basis interpolation to generate direction - sensitive masks, and adaptive anisotropic interpolation to ensure that the repaired results conform to the actual diffusion paths. Poisson equation reconstruction uses the normal regions as boundary conditions to restore the smooth transition features of the interrupted regions, and non - local mean filtering suppresses the interpolation noise, dynamically adjusting the filtering intensity to balance denoising and detail preservation. The iterative repair mechanism gradually eliminates the residual discontinuous regions until the spatial continuity index meets the standard. This method performs excellently in scenarios with complex terrains and data missing. For example, after repairing the interrupted regions of coastal bridges, the pollutant diffusion paths are consistent with the coastline trends, and the artifact elimination rate in industrial areas is significantly improved, thus significantly increasing the spatial correlation coefficient of the generated spatio - temporal feature matrix and reducing the root - mean - square error, providing a reliable basis for high - precision prediction.

[0239] Based on the same inventive concept, an embodiment of the present invention further provides an aerosol prediction system. Refer to Figure 2 As shown, it is a schematic structural diagram of a possible aerosol prediction system provided in an embodiment of the present invention. Figure 2 In it, the aerosol prediction system 200 includes: a processor 210 and a memory 220. Among them, the memory 220 stores a computer program executable by the processor 210, and the processor 210 can execute the steps of the above - mentioned artificial - intelligence - based aerosol prediction method by executing the instructions stored in the memory 220.

[0240] Based on the same inventive concept, an embodiment of the present invention provides a computer - readable storage medium, which includes a computer program. When the computer program runs on the aerosol prediction system, the computer program is used to make the aerosol prediction system execute the steps of the above - mentioned artificial - intelligence - based aerosol prediction method. In some possible implementation manners, various aspects of the artificial - intelligence - based aerosol prediction method provided by the present invention can also be implemented in the form of a program product, which includes a computer program. When the program product runs on the aerosol prediction system, the computer program is used to make the aerosol prediction system execute the steps in the above - mentioned artificial - intelligence - based aerosol prediction method. For example, the aerosol prediction system can execute the steps as shown in Figure 1 shown.

Claims

1. An artificial intelligence-based aerosol prediction method, characterized in that, The method is applied to an aerosol prediction system, and the method includes: Obtain an original multi-modal meteorological observation data set, where the original multi-modal meteorological observation data set includes a surface aerosol concentration sequence, three-dimensional atmospheric state variables, and boundary layer dynamic parameters; Perform spatio-temporal heterogeneity correction on the original multi-modal meteorological observation data set to generate a spatio-temporally aligned target multi-modal meteorological observation data set; Input the target multi-modal meteorological observation data set into a spatio-temporal feature fusion network to extract multi-scale spatial distribution features and cross-modal temporal correlation features, and generate a spatio-temporal feature matrix based on the multi-scale spatial distribution features and the cross-modal temporal correlation features; Perform multi-stage recursive optimization on the spatio-temporal feature matrix through a hierarchical time-domain aggregation algorithm to generate an aerosol concentration prediction sequence for the target area; The step of inputting the target multi-modal meteorological observation data set into a spatio-temporal feature fusion network to extract multi-scale spatial distribution features and cross-modal temporal correlation features, and generating a spatio-temporal feature matrix based on the multi-scale spatial distribution features and the cross-modal temporal correlation features includes: Perform local texture enhancement on the surface aerosol concentration sequence through a convolutional attention module to generate a super-resolution spatial feature map; Traverse the vertical profile data of the three-dimensional atmospheric state variables using a three-dimensional dilated convolutional kernel to extract the correlation features between the boundary layer thickness and the aerosol diffusion path; Perform channel dimension splicing on the super-resolution spatial feature map and the correlation features to generate a multi-modal fusion feature tensor; Based on a bidirectional gated recurrent unit, perform sliding interception on the time dimension of the multi-modal fusion feature tensor to generate cross-modal temporal correlation features; Perform spatial position encoding on the cross-modal temporal correlation features and the multi-modal fusion feature tensor to generate a spatio-temporal feature matrix.

2. The method according to claim 1, wherein The step of performing spatio-temporal heterogeneity correction on the original multi-modal meteorological observation data set to generate a spatio-temporally aligned target multi-modal meteorological observation data set includes: Identify the spatial resolution difference between the vertical stratification data of the three-dimensional atmospheric state variables and the surface aerosol concentration sequence; Perform spatial interpolation on the vertical stratification data based on the earth curvature compensation algorithm to generate equi-latitude and equi-longitude grid data matching the surface aerosol concentration sequence; For the missing area of the detected boundary layer dynamic parameters, generate a filling mask for the missing area using the eddy covariance data of adjacent meteorological stations; Perform dynamic sliding window calibration according to the timestamp deviation between the filling mask and the equi-latitude and equi-longitude grid data to generate a spatio-temporally aligned target multi-modal meteorological observation data set.

3. The method according to claim 2, characterized in that, The step of performing dynamic sliding window calibration according to the timestamp deviation between the filling mask and the equi-latitude and equi-longitude grid data to generate a spatio-temporally aligned target multi-modal meteorological observation data set includes: Detect the time span of continuous missing areas in the filling mask; Extract a similarity template from the equi-latitude and equi-longitude grid data of the same historical period according to the time span; Use the dynamic time warping algorithm to perform path matching between the similarity template and the current missing area; Adjust the step size parameter of the sliding window according to the slope change of the matching path; Perform bidirectional linear interpolation on the current missing area based on the adjusted step size parameter to generate a spatiotemporally aligned target multi-modal meteorological observation dataset.

4. The method according to claim 1, wherein The local texture enhancement of the surface aerosol concentration sequence by the convolutional attention module to generate a super-resolution spatial feature map includes: Perform Gaussian difference filtering on the surface aerosol concentration sequence to extract the spatial gradient magnitude map; Generate regional saliency weights according to the local variance distribution of the spatial gradient magnitude map; Use a deformable convolutional kernel to perform multi-directional feature extraction on the surface aerosol concentration sequence to generate a direction-sensitive feature map; Perform an element-wise product of the regional saliency weights and the direction-sensitive feature map to generate a texture-enhanced feature map; Enhance the resolution of the texture-enhanced feature map through a transposed convolutional layer to generate a super-resolution spatial feature map; The use of a deformable convolutional kernel to perform multi-directional feature extraction on the surface aerosol concentration sequence to generate a direction-sensitive feature map includes: Initialize the sampling offset parameter set of the deformable convolutional kernel; Calculate the weight coefficients in each direction according to the spatial autocorrelation function of the surface aerosol concentration sequence; Perform position-sensitive feature mapping on the sampling offset parameters through a bilinear interpolation layer; Multiply the mapped features by the weight coefficients in matrix form to generate a direction-sensitive feature map.

5. The method according to claim 1, wherein The multi-stage recursive optimization of the spatio-temporal feature matrix by the hierarchical time-domain aggregation algorithm to generate an aerosol concentration prediction sequence for the target area includes: Divide the spatio-temporal feature matrix into a first periodic fluctuation component and a second periodic fluctuation component; wherein, the period length of the first periodic fluctuation component is less than the period length of the second periodic fluctuation component; Dynamically attenuate the frequency-domain energy of the first periodic fluctuation component through an adaptive weight allocator; Use a residual connection structure to superimpose the attenuated first periodic fluctuation component and the second periodic fluctuation component to generate an optimized feature vector; Input the optimized feature vector into a mixture density network to generate probability distribution parameters of the aerosol concentration; Perform Monte Carlo sampling on the historical observation sequence according to the probability distribution parameters to generate an aerosol concentration prediction sequence for the target area; The input of the optimized feature vector into the mixture density network to generate probability distribution parameters of the aerosol concentration includes: Divide the optimized feature vector into a meteorological driving component and a pollution source contribution component; Perform a non-linear transformation on the meteorological driving component through a fully connected layer to generate mean and variance estimates; Model the time lag effect of the pollution source contribution component using a gated linear unit to generate a skewness coefficient; Combine the mean, the variance, and the skewness coefficient into a three-parameter lognormal distribution, and generate probability distribution parameters of the aerosol concentration through the three-parameter lognormal distribution.

6. The method according to claim 1, characterized in that, After generating the aerosol concentration prediction sequence for the target area, it further includes: Obtain the real-time meteorological reanalysis data stream and satellite aerosol optical depth observations; Input the real-time meteorological reanalysis data stream into a pre-trained spatio-temporal feature fusion network to extract a real-time spatio-temporal feature matrix; Dynamically align the real-time spatio-temporal feature matrix with the historical spatio-temporal feature matrix by sliding a time window to obtain a dynamic alignment result; Based on the dynamic alignment result, use the Kalman filter algorithm to iteratively update the error covariance matrix of the aerosol concentration prediction sequence; Online correct the aerosol concentration prediction sequence of the target area according to the iteratively updated error covariance matrix.

7. The method according to claim 6, characterized in that, The iterative update of the error covariance matrix of the aerosol concentration prediction sequence using the Kalman filter algorithm includes: Generate parameterized expressions of the aerosol concentration observation model and the state transition model; Determine the innovation covariance matrix according to the real-time satellite aerosol optical depth observation value; Detect the positive definiteness of the innovation covariance matrix through Cholesky decomposition; When it is detected that the innovation covariance matrix is non-positive definite, use a regularization factor to perform diagonal loading on the prior estimate to obtain the target covariance matrix; Update the Kalman gain coefficient according to the target covariance matrix to generate a corrected aerosol concentration prediction sequence.

8. The method according to claim 6, characterized in that, The online correction of the aerosol concentration prediction sequence of the target area according to the iteratively updated error covariance matrix includes: Obtain the aerosol concentration observation value at the current moment in the real-time meteorological reanalysis data stream and the initial prediction value of the aerosol concentration prediction sequence; Generate the weight allocation ratio of the aerosol concentration observation value at the current moment according to the Kalman gain coefficient of the error covariance matrix; perform linear weighted fusion on the initial prediction value and the aerosol concentration observation value based on the weight allocation ratio to generate the corrected aerosol concentration value at the current moment; Extract the residual component of the corrected aerosol concentration value and the spatial distribution residual of the satellite aerosol optical depth observation value; perform feature mapping on the spatial distribution residual through the spatio-temporal feature fusion network to generate a residual spatio-temporal correlation matrix; Input the residual spatio-temporal correlation matrix into the hierarchical time-domain aggregation algorithm to extract the periodic fluctuation pattern of the residual component; dynamically adjust the time decay factor of the Kalman gain coefficient according to the periodic fluctuation pattern; Update the sliding window length of the error covariance matrix using the adjusted time decay factor; perform moving average filtering on the aerosol concentration prediction sequence at the next moment based on the updated sliding window length to generate a smoothed aerosol concentration prediction value; Perform time series splicing on the smoothed aerosol concentration prediction value and the corrected aerosol concentration value to generate an online corrected aerosol concentration prediction sequence for the target area; The generating of the residual spatio-temporal correlation matrix by performing feature mapping on the spatial distribution residual through the spatio-temporal feature fusion network includes: Decompose the spatial distribution residual into a multi-channel residual tensor of a spatial residual component and a temporal residual component; perform multi-scale dilated convolution operations on the spatial residual component of the multi-channel residual tensor to extract spatial gradient mutation features under different receptive fields; Generate a spatial attention weight matrix according to the local direction consistency of the spatial gradient mutation features; Performing a time series causal convolution operation on the time series residual component of the multi-channel residual tensor to extract lag time correlation features; generating a time series convolution feature map based on the periodic intensity distribution of the lag time correlation features; The spatial attention weight matrix and the temporal convolution feature map are cross-modally concatenated to generate a spatiotemporal cross-correlation feature block; the spatiotemporal cross-correlation feature block is adaptively normalized in the channel dimension to generate a normalized spatiotemporal feature vector; the normalized spatiotemporal feature vector is modeled with forward and backward dependencies through a bidirectional long short-term memory network to extract global temporal dependency features; Using a graph convolution layer to perform topological structure aggregation on the spatial adjacency relationship of the global temporal dependency feature to generate a spatial node embedding feature; performing a feature addition operation on the spatial node embedding feature and the global temporal dependency feature to generate an enhanced spatiotemporal feature sequence; generating a three-dimensional feature encoding vector based on the spatiotemporal dimension distribution of the enhanced spatiotemporal feature sequence; The three-dimensional feature encoding vector is nonlinearly projected through a multilayer perceptron to generate a residual spatiotemporal correlation matrix; wherein each element of the residual spatiotemporal correlation matrix represents the residual correlation strength of different spatial positions and time steps.

9. An aerosol prediction system, characterized in that, The method comprises a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor is enabled to perform the steps of any one of the methods of claims 1 to 8.

Images