Aerosol prediction method and system based on artificial intelligence

Abstract

The invention relates to the technical field of meteorological data analysis, and provides an aerosol prediction method and system based on artificial intelligence, which are used for accurately identifying an abnormal aerosol gathering mode in an extreme meteorological event. The method comprises the steps that an original multi-mode meteorological observation data set is acquired, and the original multi-mode meteorological observation data set comprises an earth surface aerosol concentration sequence, three-dimensional atmosphere state variables and boundary layer dynamic parameters; performing space-time heterogeneity correction on the original multi-modal meteorological observation data set to generate a space-time aligned target multi-modal meteorological observation data set; inputting the target multi-modal meteorological observation data set into a spatial-temporal feature fusion network to extract multi-scale spatial distribution features and cross-modal time sequence correlation features, and generating a spatial-temporal feature matrix based on the multi-scale spatial distribution features and the cross-modal time sequence correlation features; and performing multi-stage recursive optimization on the spatial-temporal characteristic matrix through a hierarchical time domain aggregation algorithm to generate an aerosol concentration prediction sequence of the target region.

Description

Technical Field

[0001] The present invention belongs to the technical field of meteorological data analysis, and specifically 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 environmental monitoring and pollution control. Traditional methods mainly rely on ground station observation data or single meteorological models for statistical analysis. However, in actual application, traditional methods often have problems such as data heterogeneity constraints, model architecture limitations, insufficient time domain modeling, and failure to predict extreme events.

[0003] The above defects lead to bottlenecks in existing aerosol prediction technology, such as low temporal and spatial resolution, insufficient modeling of multi-physical process coupling, and poor continuous prediction stability, which makes 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. Summary of the invention

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

[0006] In a first aspect, an embodiment of the present invention provides an artificial intelligence-based aerosol prediction method, which is applied to an aerosol prediction system, and the method includes: obtaining an original multimodal meteorological observation dataset, wherein the original multimodal meteorological observation dataset includes a surface aerosol concentration sequence, three-dimensional atmospheric state variables, and boundary layer dynamic parameters; performing spatiotemporal heterogeneity correction on the original multimodal meteorological observation dataset to generate a spatiotemporal aligned target multimodal meteorological observation dataset; inputting the target multimodal meteorological observation dataset into a spatiotemporal feature fusion network to extract multi-scale spatial distribution features and cross-modal time series correlation features, and generating a spatiotemporal feature matrix based on the multi-scale spatial distribution features and the cross-modal time series correlation features; performing multi-stage recursive optimization on the spatiotemporal feature matrix through a hierarchical time domain aggregation algorithm to generate an aerosol concentration prediction sequence for the 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, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes 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 is run on an aerosol prediction system, the computer program is used to enable the aerosol prediction system to perform the steps of the above method.

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

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

[0011] On this basis, the spatiotemporal feature fusion network dynamically perceives the nonlinear relationship between atmospheric parameters at different altitudes and the surface environment, realizes multi-scale spatial feature analysis from micro-turbulence to macro-circulation, and enhances the modeling ability of aerosol distribution laws under complex meteorological conditions.

[0012] In addition, the hierarchical time-domain aggregation framework uses a progressive optimization strategy to simultaneously take into account the instantaneous response characteristics and long-term evolution trends of the meteorological system, significantly improving the temporal consistency of continuous forecasts while ensuring real-time forecasting performance.

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

[0014] Figure 1 A schematic flow chart of an aerosol prediction method based on artificial intelligence provided in an embodiment of the present invention.

[0015] Figure 2 A schematic diagram of the structure of an aerosol prediction system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0016] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the technical solution of the present invention, rather than all the embodiments. Based on the embodiments recorded in the present invention document, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the technical solution of the present invention.

[0017] See also Figure 1 , which is an aerosol prediction method based on artificial intelligence provided in an embodiment of the present invention. The method can be applied to an aerosol prediction system. The specific process is as follows S110-S140: S110: Acquire an original multimodal meteorological observation dataset, where the original multimodal meteorological observation dataset includes a surface aerosol concentration sequence, three-dimensional atmospheric state variables, and boundary layer dynamic parameters.

[0018] In an embodiment of the present invention, the aerosol prediction system first obtains an original multi-modal meteorological observation dataset from a multi-source observation platform.

[0019] For example, the surface aerosol concentration series comes from the ground environmental monitoring station network and the MODIS sensor of the polar-orbiting satellite, such as the PM2.5 and PM10 mass concentration data uploaded every few minutes by multiple air quality monitoring stations distributed in a plain area, and the 550nm aerosol optical thickness inversion product obtained every hour by the geostationary satellite.

[0020] The three-dimensional atmospheric state variables are generated by fusing meteorological reanalysis data with sounding observations, including three-dimensional wind speed field, temperature field, and humidity field grid data with a spatial resolution of 0.25° provided by a medium-term weather forecast center, and boundary layer vertical profile data collected at regular intervals by multiple wind profiler radars deployed in the target area.

[0021] The dynamic parameters of the boundary layer involve characteristic quantities such as the atmospheric mixing layer height and turbulent kinetic energy flux, which are obtained by assimilation calculation of lidar network observations and WRF models. For example, several Doppler lidars deployed in a certain urban agglomeration output the boundary layer structure parameters with a vertical resolution of 1 km according to the set time step.

[0022] The above multi-source heterogeneous data are formatted and standardized through the data acquisition module of the aerosol prediction system to form an original multimodal meteorological observation dataset containing timestamps, geographic coordinates, and numerical quality identifiers.

[0023] S120: Performing spatiotemporal heterogeneity correction on the original multimodal meteorological observation dataset to generate a spatiotemporal aligned target multimodal meteorological observation dataset.

[0024] In an embodiment of the present invention, the aerosol prediction system performs correction processing for the temporal and spatial heterogeneity of the original data set.

[0025] For example, during a typical pollution process in a plain in winter, the time resolution of ground station data is 5 minutes while that of satellite data is at the hourly level. The aerosol prediction system uses an adaptive time alignment algorithm to perform cubic spline interpolation on the MODIS aerosol optical thickness to generate a 5-minute interval data set synchronized with ground monitoring.

[0026] In addition, in terms of spatial dimension, in view of the scale difference between the 0.25° grid of meteorological reanalysis data and the 1km 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 that matches the reanalysis field.

[0027] 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 activated the data filling mechanism for the earth's boundary area. Based on the dynamic uplift effect of mountain terrain simulated by the WRF model and the daily variation of the mixing layer height observed by lidar in the adjacent area, a spatially continuous boundary layer parameter field was generated.

[0028] It can be understood that the datasets after time and space correction are unified in the UTM-50N coordinate system to form a regular grid of 500m×500m, and the time base adopts Coordinated Universal Time to ensure strict alignment of different modal data in time and space dimensions.

[0029] S130: Inputting the target multimodal meteorological observation dataset into a spatiotemporal feature fusion network to extract multi-scale spatial distribution features and cross-modal temporal correlation features, and generating a spatiotemporal feature matrix based on the multi-scale spatial distribution features and the cross-modal temporal correlation features.

[0030] In an embodiment of the present invention, the aerosol prediction system inputs the preprocessed multimodal data into a spatiotemporal feature fusion network for deep feature extraction.

[0031] For example, the spatial feature extraction module uses the VisionTransformer architecture to process two-dimensional meteorological fields. For example, it divides the wind, temperature and humidity grid data at the 850hPa altitude layer into a 16×16 tile sequence, and captures the spatial correlation pattern of the sea-land breeze circulation between a plain and a bay through a multi-head self-attention mechanism.

[0032] The U-Net encoder-decoder structure is used to extract the multi-scale characteristics of the aerosol concentration field. In the encoding stage, the 200km regional scale characteristics of pollution diffusion in a certain urban agglomeration are captured by downsampling four times. In the decoding stage, the resolution is gradually restored to the original resolution to retain the 50km fine structure of the local emission source.

[0033] Furthermore, the time series association modeling part uses a spatiotemporal encoder to process the 72-hour continuous meteorological sequence, in which the gated recurrent unit network learns the daily variation cycle of the boundary layer parameters and the temporal convolutional network extracts the minute-level fluctuation characteristics of the wind speed pulsation.

[0034] In addition, the global attention mechanism can dynamically integrate different modal features. For example, it can automatically increase the weight of PM2.5 data from ground monitoring stations under calm weather conditions, and focus on the cross-modal correlation between three-dimensional wind fields and turbulent kinetic energy when a cold front passes, ultimately generating a spatiotemporal feature matrix containing multi-dimensional feature vectors.

[0035] In an embodiment of the present invention, the spatiotemporal feature matrix is ​​a multidimensional data structure constructed by fusing the spatial and temporal dimension features of multimodal meteorological data. It can be understood that the spatiotemporal feature matrix integrates the multi-scale spatial distribution features extracted by the deep neural network (such as the macroscopic circulation pattern of the regional meteorological field and the fine structure of the local pollution source) and the cross-modal time series correlation features (such as the dynamic interaction between different meteorological elements and their evolution over time). The core of the spatiotemporal 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 expression, which includes both spatial topological associations (such as the diffusion path of pollutants under the influence of terrain) and continuous dynamics in time (such as periodic changes and sudden fluctuations in boundary layer parameters), thereby providing high-dimensional feature support for subsequent predictions.

[0036] S140: Perform multi-stage recursive optimization on the spatiotemporal feature matrix through a hierarchical time-domain aggregation algorithm to generate an aerosol concentration prediction sequence for the target area.

[0037] In the embodiment 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 variation of aerosol concentration through a bidirectional long short-term memory network, such as the influence of the daytime dissipation and nighttime reconstruction of a typical ground radiation inversion layer in a certain area on the vertical diffusion of pollutants.

[0038] For another example, the 6-hour scale aggregation module in the second stage uses the temporal self-attention mechanism to strengthen the feature expression of the pollution accumulation stage, focusing on modeling the coupling effect of evening traffic peak emissions and the sudden drop in boundary layer height. The hourly predictor in the final stage integrates multi-scale features and retains the key physical process information in the original spatiotemporal feature matrix through residual connections. For example, when predicting aerosol concentrations in the next 72 hours, the aerosol prediction system integrates the daily-scale atmospheric circulation background field changes, the hourly-scale local emission source intensity fluctuations, and the minute-scale turbulent mixing characteristics across time series, 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 to achieve rolling optimization of the prediction results through a sliding time window mechanism, which significantly improves the prediction stability of continuous pollution processes.

[0039] Based on this, in the embodiment of the present invention, multi-stage recursive optimization is a hierarchical and progressive time series modeling method, which aims to improve the prediction accuracy through feature aggregation and iterative correction of different time granularities. The process adopts a hierarchical time domain aggregation algorithm to process the spatiotemporal feature matrix in stages.

[0040] At the macroscale stage, long-term evolution laws (such as the diurnal changes in the atmospheric circulation background) are captured, and global time dependencies are modeled through recurrent neural networks or attention mechanisms.

[0041] At the mesoscale stage, we focus on the feature enhancement of key periods (such as the meteorological coupling effect during the pollution accumulation period) and use temporal self-attention to screen and enhance specific dynamic patterns.

[0042] At the microscale stage: integrate multi-scale features and retain the details of physical processes (such as the instantaneous impact of turbulent mixing), and achieve refined predictions through residual connections and cross-time series fusion mechanisms.

[0043] It can be understood that each stage realizes rolling update of parameters through recursive feedback and sliding time window mechanism, ensuring continuous optimization of prediction results in terms of temporal consistency and stability.

[0044] In the embodiment of the present invention, the aerosol concentration prediction sequence of the target area is a continuous spatiotemporal distribution result based on the spatiotemporal feature matrix and multi-stage optimization output, which characterizes the dynamic evolution of the aerosol concentration in the target area within a specific period of time in the future. The sequence is generated by integrating multi-scale spatiotemporal features (such as regional transmission, local emissions and vertical diffusion processes) and has a clear spatial coverage and time continuity. The characteristics of the aerosol concentration prediction sequence include at least:

[0045] (1) Spatial expression: Presented in the form of a regular grid or geocoding, retaining the spatial heterogeneity of pollutant diffusion (such as the difference in concentration gradients between urban agglomerations and natural landforms).

[0046] (2) Temporal dynamics: The predicted values ​​are output at preset time intervals (e.g., hourly), reflecting the nonlinear response of aerosol concentration to changes in meteorological conditions (e.g., accumulation in calm weather and rapid clearance after a frontal passage).

[0047] (3) Uncertainty control: Through cross-modal feature association and recursive optimization mechanism, the prediction bias caused by missing data or model errors is reduced to ensure the robustness of the sequence in complex meteorological scenarios.

[0048] In summary, the embodiments of the present invention significantly improve the spatiotemporal accuracy and model generalization ability of aerosol concentration prediction through the collaborative optimization of multimodal meteorological data fusion and deep learning architecture.

[0049] First, we integrate multi-source heterogeneous data such as ground stations, satellite remote sensing, and meteorological reanalysis to build a three-dimensional observation system covering the surface to the boundary layer, thereby enhancing the completeness of data representation. We solve the resolution differences and geographical missing problems of multimodal data through an adaptive spatiotemporal correction algorithm, thereby ensuring the spatiotemporal consistency of the input data.

[0050] Secondly, the hybrid architecture of VisionTransformer and U-Net is used to extract multi-scale spatial features, and the gated recurrent unit and time convolutional network are combined to capture cross-modal temporal associations, breaking through the limitations of traditional models in modeling local features and long-range dependencies. The hierarchical time-domain aggregation algorithm is further used to integrate the daily cycle law, pollution accumulation dynamics and instantaneous turbulence effects across scales through staged recursive optimization, thus achieving a refined simulation of the meteorological-pollution coupling process.

[0051] The final output prediction sequence has both high spatiotemporal resolution and physical consistency, and can accurately depict the aerosol diffusion law in complex terrain, sudden weather changes and other scenarios, providing highly robust scientific support for regional pollution warning and prevention and control decisions. The embodiment of the present invention systematically solves the modeling problem of nonlinear interaction between meteorological elements and aerosol concentration through the technical closed loop of "multi-source data-feature fusion-cross-scale prediction", and promotes the evolution of aerosol prediction technology towards intelligence and high precision.

[0052] In a preferred embodiment, the step of performing spatiotemporal heterogeneity correction on the original multimodal meteorological observation dataset in S120 to generate a spatiotemporal aligned target multimodal meteorological observation dataset includes: S121: Identify the spatial resolution difference between the vertical layered data of the three-dimensional atmospheric state variable and the surface aerosol concentration series.

[0053] In an embodiment of the present invention, the aerosol prediction system identifies the spatial resolution difference between the vertical layered data of the three-dimensional atmospheric state variables and the surface aerosol concentration series. Specifically, the three-dimensional wind speed field and temperature field provided by the meteorological reanalysis data adopt a 0.25° longitude and latitude grid, which is vertically divided into 37 pressure layers from the surface to the top of the troposphere, while the surface aerosol concentration data comes from the discrete point observations of the ground monitoring station and the 1km spatial resolution grid data of the satellite inversion product.

[0054] For example, in an aerosol prediction task in a plain area, the system detected that the horizontal wind field grid spacing at the 850hPa altitude layer was about 27.8km, while the spacing between ground PM2.5 monitoring stations was 10-15km, and there was a significant difference in spatial coverage density between the two. The aerosol prediction system calculates the matching degree between the grid spacing of each vertical layer data in the horizontal projection direction and the coverage density of the surface observation data, and identifies that the spatial resolution difference between the atmospheric state variables and the surface aerosol concentration in the altitude range of 925hPa to 700hPa in the boundary layer exceeds the preset threshold, and spatial interpolation of vertical layered data is required.

[0055] S122: Performing spatial interpolation on the vertical layered data based on an earth curvature compensation algorithm to generate equal longitude and latitude grid data matching the surface aerosol concentration sequence.

[0056] In an embodiment of the present invention, the aerosol prediction system performs spatial interpolation on vertical layered data based on an earth curvature compensation algorithm to generate equal longitude and latitude grid data that matches the surface aerosol concentration sequence. The earth curvature compensation algorithm geometrically corrects the traditional plane interpolation method by introducing an arc length correction factor under an ellipsoid earth model.

[0057] For example, when interpolating the original 1°×1° longitude and latitude grid in the meteorological reanalysis data to the same 0.1° grid as the satellite aerosol optical thickness data, the system first calculates the geodetic coordinates of the target grid point on the surface of the Earth's ellipsoid, and then adjusts the interpolation weight according to the spatial curvature of the adjacent grid points. For the vertical wind speed field data in a certain bay area, the algorithm automatically compensates for the meridian convergence effect caused by the increase in latitude during the north-south interpolation process, ensuring that the grid spacing of the interpolated 925hPa altitude layer wind field data in the coastal urban agglomeration is consistent with the spatial distribution of the surface monitoring stations, eliminating the spatial resolution deviation caused by projection deformation.

[0058] S123: For the detected missing area of ​​the boundary layer dynamic parameter, generate a filling mask for the missing area using the eddy covariance data of the neighboring meteorological stations.

[0059] In an 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 the neighboring meteorological stations. The filling mask is a binary matrix containing spatial weight distribution and data credibility index, which is used to identify the relationship between the area to be filled and its neighboring valid observation points.

[0060] For example, when the height data of the mixed layer observed by the lidar is continuously missing due to terrain obstruction on the east side of a mountain range, the system selects the eddy covariance data of three meteorological stations on the west side of the mountain range to calculate the spatiotemporal variation characteristics of its turbulent kinetic energy flux. By analyzing the spatial transmission law of turbulence intensity during the development of valley wind circulation in the morning period, the system constructs a filling mask bounded by the ridge line, in which the weight coefficient of the western area of ​​the mask is directly assigned based on the measured data, and the interpolation weight of the missing area in the east is generated according to the terrain slope and distance attenuation function, forming a boundary layer parameter spatial distribution template covering the entire target area.

[0061] 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 target multimodal meteorological observation dataset aligned in time and space.

[0062] In the embodiment of the present invention, the aerosol prediction system performs dynamic sliding window calibration based on the timestamp deviation of the filling mask and the equal longitude and latitude grid data to generate a target multimodal meteorological observation data set aligned in time and space. Specifically, the system dynamically adjusts the start and end time of the time window based on the current time to compensate for the time synchronization error between different data sources.

[0063] For example, during the winter pollution process of a plain urban agglomeration, the ground monitoring station data was timestamped at 08:00:00 UTC on December 16, while the meteorological reanalysis data was timestamped at 08:02:30 UTC due to calculation delays. After the system detected a 2.5-minute time deviation, it expanded the sliding window to 08:00:00-08:05:00, performed cubic spline interpolation on the reanalysis data within the window, and generated a dataset at 8:00:00 that was strictly aligned with the ground data. In response to the rapid changes in the wind speed field during the passage of the cold front, the system automatically shortened the window step from 5 minutes to 1 minute to ensure that the sudden change in wind direction at the time of the frontal arrival was accurately captured and synchronized to other modal data.

[0064] In a preferred embodiment, the step of 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 target multimodal meteorological observation dataset aligned in time and space includes: S1241: Detect the time span of the continuous missing regions in the filling mask.

[0065] In an embodiment of the present invention, the aerosol prediction system detects the time span of the continuous missing areas in the 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 the continuous calm weather in a certain urban agglomeration in winter, due to cloudy weather, multiple lidars were unable to obtain boundary layer height data from 08:00 on December 5 to 14:00 on December 7. The system determined that the span of the missing period was 54 hours through timestamp continuity analysis. The system further divides 54 hours into 6 9-hour time windows, and detects the spatial distribution range of the missing grid points and the coverage density of the adjacent valid data in each window, and identifies the regional data missing due to foggy weather from 02:00 on December 6 to 11:00 on December 6, with a time span of 9 hours and an impact range of more than 60% of the total area.

[0066] S1242: Extracting similarity templates from the equal longitude and latitude grid data of the same historical period according to the time span.

[0067] In an embodiment of the present invention, the aerosol prediction system extracts similarity templates from the historical grid data of equal latitude and longitude according to the time span. The similarity template refers to a historical period data set with similar meteorological conditions to the current missing period, which is used to provide the spatial distribution pattern required for filling.

[0068] For example, in view of the missing boundary layer parameters from 02:00 to 11:00 on December 6, the system retrieved the lidar observation records of the same period in the past five years (December 1 to 15), and screened out two historical periods, December 8, 2019 and December 4, 2021, whose ground meteorological observation data (wind speed, temperature, humidity) and the spatial distribution of the sea level pressure field in the current missing period had a correlation coefficient of more than 0.85. The system extracted the boundary layer height data in the same UTC time range for these two historical periods, and generated a similarity template containing the daily variation characteristics of the mixing layer height under typical calm weather through spatial standardization.

[0069] S1243: Perform path matching between the similarity template and the current missing region using a dynamic time warping algorithm.

[0070] In an embodiment of the present invention, the aerosol prediction system uses a dynamic time warping algorithm to perform path matching between the similarity template and the current missing area. The dynamic time warping algorithm solves the problem of time series phase offset caused by differences in the evolution rate of meteorological processes by constructing nonlinear alignment paths between time series.

[0071] For example, when matching the ground temperature time series data of the historical template on December 4, 2021 with the current missing period, the algorithm detected that the morning inversion layer in the historical template dissipated 2 hours earlier than the current missing period. By calculating the Euclidean distance matrix of 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 the weather system.

[0072] S1244: Adjust the step size parameter of the sliding window according to the slope change of the matching path.

[0073] In an 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 sliding window step size parameter determines the data sampling interval of the time dimension during the filling process and needs to be adaptively adjusted according to the local slope of the dynamic time warping path.

[0074] For example, when the matching path has a slope of 45 degrees from 08:00 to 10:00 (indicating that the history and current time series are synchronized), the system uses a fixed 1-hour step size for data filling; and when the path slope increases to 60 degrees from 10:00 to 12:00 (indicating that the current time series is accelerating), the system shortens the step size to 0.5 hours to improve the temporal resolution. For a certain cold front passing process, the system detected that the matching path had a sudden change in slope when the front arrived, and dynamically adjusted the step size parameters accordingly to ensure that the data filling frequency in the strong wind area behind the front was increased to 15 minutes.

[0075] S1245: Perform bidirectional linear interpolation on the current missing area based on the adjusted step size parameter to generate a target multimodal meteorological observation dataset that is aligned in time and space.

[0076] In an embodiment of the present invention, the aerosol prediction system performs bidirectional linear interpolation on the current missing area based on the adjusted step size parameter to generate a target multimodal meteorological observation data set aligned in time and space. Bidirectional linear interpolation fills data in the forward and backward directions along the time axis, and then obtains the final result through weighted averaging.

[0077] For example, when filling in the boundary layer height data of a valley area at 08:30 on December 6, the system obtains the forward 08:00 historical template data and the backward 09:00 satellite inversion data with an adjusted 30-minute step size, and performs linear interpolation on the 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 that the height of the mixed layer on the leeward slope is lower than that on the windward slope in 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 500m×500m UTM-50N grid coordinate system, and the time base is aligned to the coordinated universal time, forming a time-space consistent data set that can be directly input into the prediction model.

[0078] In a preferred embodiment, the step of inputting the target multimodal meteorological observation dataset into a spatiotemporal feature fusion network in S130 to extract multiscale spatial distribution features and cross-modal temporal association features, and generating a spatiotemporal feature matrix based on the multiscale spatial distribution features and the cross-modal temporal association features, includes: S131: Performing local texture enhancement on the surface aerosol concentration sequence through a convolutional attention module to generate a super-resolution spatial feature map.

[0079] In an embodiment of the present invention, the aerosol prediction system performs local texture enhancement on 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 raster data that has been time-space corrected into the convolutional attention module, which first uses a Gaussian difference filtering algorithm to enhance the spatial gradient amplitude.

[0080] For example, in an industrial cluster of a plain city group, the aerosol concentration field shows significant spatial heterogeneity within a range of 5 km. After smoothing the original data with a 3×3 Gaussian kernel, the system uses the Laplace operator to detect the concentration gradient mutation area around the factory area and generates a gradient amplitude map reflecting the spatial distribution differences of the pollution source. Subsequently, the system generates a regional significance weight matrix based on the local variance distribution of the gradient amplitude map, focusing on strengthening the texture details around traffic arteries and industrial point sources.

[0081] In the above process, the system uses a deformable convolution kernel to extract multi-directional features of the concentration field. For example, for the pollutant diffusion path dominated by sea and land breezes in a certain bay area, the deformable convolution kernel automatically adjusts the sampling offset direction to capture the feather-like diffusion characteristics from the coastal sewage outlet to the offshore area along the southeast-northwest axis. Finally, the system multiplies the direction-sensitive feature map by the regional significance weight element by element, and increases the spatial resolution from 1km to 500m through the transposed convolution layer, generating a super-resolution spatial feature map that can clearly characterize the spatial structure of local emission sources.

[0082] S132: Using a three-dimensional expansion convolution kernel to traverse the vertical profile data of the three-dimensional atmospheric state variable, extracting correlation features between the boundary layer thickness and the aerosol diffusion path.

[0083] In an embodiment of the present invention, the aerosol prediction system uses a three-dimensional dilated convolution kernel to traverse the vertical profile data of the three-dimensional atmospheric state variable to extract the correlation characteristics between the boundary layer thickness and the aerosol diffusion path. The three-dimensional dilated convolution captures the long-range dependency in the vertical direction by expanding the receptive field.

[0084] For example, when processing the three-dimensional field data of temperature, humidity, and wind speed over a certain urban agglomeration, the system uses a convolution kernel with an expansion rate of 3 to slide in the vertical direction, and detects the temperature gradient mutation characteristics caused by the inversion layer near the top of the boundary layer. This feature is highly correlated with the inflection point position of the vertical profile of the PM2.5 concentration on the surface. Based on this, the system establishes a quantitative correlation model between the dynamic change of the boundary layer height and the vertical diffusion capacity of pollutants, and accurately predicts the near-ground pollution accumulation caused by the mixing layer height of less than 200 meters in the early morning under calm weather.

[0085] S133: Concatenate the super-resolution spatial feature map and the associated features in channel dimension to generate a multimodal fusion feature tensor.

[0086] In an embodiment of the present invention, the aerosol prediction system performs channel dimension splicing on the super-resolution spatial feature map and the associated features to generate a multi-modal fusion feature tensor. Channel dimension splicing realizes the parallel expression and information complementarity of different modal features.

[0087] For example, the 250m resolution super-resolution concentration feature map (including 32 feature channels) and the boundary layer correlation feature (including 16 channels) are spliced ​​in the channel dimension to form a 48-channel fusion feature tensor. In the scene of the sea breeze front passing through the coastal city, this tensor retains the fine structure of the sewage outlet depicted by the super-resolution feature map and the vertical transport features extracted by three-dimensional convolution, providing the model with a multi-modal information basis that takes into account both the horizontal diffusion details and the vertical mixing process.

[0088] S134: Slidingly intercepting the time dimension of the multimodal fusion feature tensor based on a bidirectional gated recurrent unit to generate a cross-modal temporal correlation feature.

[0089] In an embodiment of the present invention, the aerosol prediction system generates cross-modal temporal correlation features by slidingly intercepting the time dimension of the multimodal fusion feature tensor based on a bidirectional gated recurrent unit. The bidirectional gated recurrent unit captures temporal dependencies through two time streams, forward and backward.

[0090] For example, when processing a multimodal fusion feature tensor for 72 consecutive hours, the system slides and intercepts the input sequence with a time step of 6 hours. In a cold front crossing event, the forward network learns the pollutant accumulation trend in the pre-front calm stage, and the backward network captures the removal effect of the strong wind process after the front. The time series correlation features generated by splicing the hidden states of the two accurately quantify the time lag relationship between the sudden increase in boundary layer height and the sudden drop in PM2.5 concentration, providing key time series features for predicting the turning point of the pollution process.

[0091] S135: Perform spatial position encoding on the cross-modal temporal correlation features and the multimodal fusion feature tensor to generate a spatiotemporal feature matrix.

[0092] In an embodiment 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 spatiotemporal feature matrix. The spatial position encoding injects absolute and relative position information through a sinusoidal position embedding algorithm.

[0093] 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 time series features position by position, so that the spatiotemporal feature matrix simultaneously contains the evolution law of meteorological elements, the diffusion path of pollutants, and the topological constraints of geographic 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 transmission, thereby improving the spatial rationality of the prediction results.

[0094] In a preferred embodiment, the step of performing local texture enhancement on the surface aerosol concentration sequence by a convolutional attention module in S131 to generate a super-resolution spatial feature map includes: S1311: Perform Gaussian difference filtering on the surface aerosol concentration sequence to extract a spatial gradient amplitude map.

[0095] In an embodiment of the present invention, the aerosol prediction system performs Gaussian difference filtering on the surface aerosol concentration sequence to extract the spatial gradient amplitude map. The Gaussian difference filtering highlights the spatial gradient characteristics through the difference of Gaussian kernel convolution results of different scales.

[0096] For example, when processing satellite-inverted aerosol optical thickness data in a river valley area, the system uses double Gaussian kernels of σ=1.5km and σ=3km to convolve the original 1km grid data, and calculates the difference between the two to obtain a gradient amplitude map. The map shows high gradient values ​​in the front of the mountain on the west side of the valley, accurately reflecting the concentration abrupt change boundary formed by the accumulation of pollutants at the foot of the mountain due to the nighttime inversion layer. The spatial distribution of the gradient amplitude map is highly consistent with the PM2.5 hourly concentration mutation events measured by the ground monitoring station, verifying the ability of the filtering process to capture the front of pollutant diffusion.

[0097] S1312: Generate regional significance weights according to the local variance distribution of the spatial gradient amplitude map.

[0098] In an embodiment of the present invention, the aerosol prediction system generates regional significance weights according to the local variance distribution of the spatial gradient amplitude map. The regional significance weights are used to quantify the importance of features at different spatial locations, and the calculation is based on the statistical characteristics within the sliding window.

[0099] For example, in the central urban area of ​​a megacity, the system traverses the gradient amplitude map with a 500m×500m window and calculates the variance of the gradient value in each window. The results show that the window variance of the main road network intersection reaches 12.5, which is significantly higher than the 2.3 of the residential area window. The system generates a weight coefficient matrix based on this, setting the weight of the intersection area to 0.9 and the residential area weight to 0.2. This weight matrix effectively strengthens the spatial identification of traffic source emissions and provides guidance information for subsequent feature extraction.

[0100] S1313: Using a deformable convolution kernel to perform multi-directional feature extraction on the surface aerosol concentration sequence to generate a direction-sensitive feature map.

[0101] In an embodiment of the present invention, the aerosol prediction system uses a deformable convolution kernel to extract multi-directional features of the surface aerosol concentration sequence to generate a direction-sensitive feature map. The deformable convolution kernel achieves direction-sensitive feature capture by adaptively adjusting the sampling point position.

[0102] For example, in the leeward slope area of ​​a mountain range, the system initializes the sampling offset parameter set of the deformable convolution kernel so that it can adjust the receptive field along the direction of the terrain contour. 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 convolution kernel to offset 1.2 pixels in the positive direction of the X-axis and 0.8 pixels in the negative direction of the Y-axis. The bilinear interpolation layer performs feature mapping on the offset sampling position, and the resulting direction-sensitive feature map clearly shows the pollutant retention zone caused by the vortex on the leeward slope, and its spatial morphology is highly consistent with the low-altitude circulation structure observed by the wind profile radar.

[0103] S1314: performing element-by-element multiplication of the regional saliency weight and the direction-sensitive feature map to generate a texture enhancement feature map.

[0104] In an embodiment of the present invention, the aerosol prediction system generates a texture enhancement feature map by performing an element-by-element product of the regional significance weight and the direction-sensitive feature map. The regional significance weight matrix identifies the spatial key areas through local variance analysis, and the direction-sensitive feature map is extracted by a deformable convolution kernel along the dominant diffusion direction.

[0105] For example, around a petrochemical park, the system detected that the local variance of the southeast plant boundary area reached 28.5, generating a weight coefficient of 0.92, while the weight of the northwest farmland area was only 0.15. The direction-sensitive feature map shows the feather-like texture of pollutants diffusing along the southeast-northwest axis. After multiplying the two element by element, the texture amplitude of the plant boundary area increased from 12.7 to 11.7 (12.7×0.92), while the farmland area decreased from 3.2 to 0.48 (3.2×0.15), effectively enhancing the visualization of the emission path of industrial point sources.

[0106] S1315: Performing resolution enhancement on the texture enhancement feature map through a transposed convolution layer to generate a super-resolution spatial feature map.

[0107] In an embodiment of the present invention, the aerosol prediction system generates a super-resolution spatial feature map by performing resolution enhancement on the texture enhancement feature map through a transposed convolution layer. The transposed convolution achieves resolution enhancement through a learnable upsampling kernel.

[0108] For example, a 500m resolution texture-enhanced feature map of a city suburb was input into the transposed convolution layer, and the system used a 4×4 convolution kernel for 2x upsampling to generate a 250m resolution super-resolution feature map. This feature map reveals the 200-300m scale pollution cluster structure around key industrial point sources that cannot be distinguished by traditional interpolation methods, and the spatial matching degree with the cruise observation results of the high-precision mobile monitoring vehicle reached 92%, significantly improving the ability to identify local pollution sources.

[0109] In a more specific embodiment, the step of performing multi-directional feature extraction on the surface aerosol concentration sequence using a deformable convolution kernel in S1313 to generate a direction-sensitive feature map includes: S13131: Initialize the sampling offset parameter set of the deformable convolution kernel.

[0110] In an embodiment of the present invention, the aerosol prediction system generates a texture enhancement feature map by performing an element-by-element product of the regional significance weight and the direction-sensitive feature map. The regional significance weight matrix identifies the spatial key areas through local variance analysis, and the direction-sensitive feature map is extracted by a deformable convolution kernel along the dominant diffusion direction.

[0111] For example, around a petrochemical park, the system detected that the local variance of the southeast plant boundary area reached 28.5, generating a weight coefficient of 0.92, while the weight of the northwest farmland area was only 0.15. The direction-sensitive feature map shows the feather-like texture of pollutants diffusing along the southeast-northwest axis. After multiplying the two element by element, the texture amplitude of the plant boundary area increased from 12.7 to 11.7 (12.7×0.92), while the farmland area decreased from 3.2 to 0.48 (3.2×0.15), effectively enhancing the visualization of the emission path of industrial point sources.

[0112] S13132: Calculate the weight coefficients in each direction according to the spatial autocorrelation function of the surface aerosol concentration sequence.

[0113] In an embodiment of the present invention, the aerosol prediction system calculates the weight coefficients of each direction according to the spatial autocorrelation function of the surface aerosol concentration sequence. The spatial autocorrelation function is used to quantify the concentration variation trend in different directions.

[0114] For example, in a valley basin, the system calculated the spatial autocorrelation coefficients in four directions: 0, 45, 90, and 135 degrees, and found that the correlation coefficient in the 45-degree direction (along the valley direction) reached 0.78, which was significantly higher than that in other directions. Based on this, the system increased the weight coefficient of the deformable convolution kernel in the 45-degree direction to 0.65, and reduced the weights of the other directions accordingly, so that the feature extraction process focuses on the dominant diffusion direction guided by the terrain.

[0115] S13133: Perform position-sensitive feature mapping on the sampling offset parameters through a bilinear interpolation layer.

[0116] In an embodiment of the present invention, the aerosol prediction system performs position-sensitive feature mapping on the sampling offset parameter through a bilinear interpolation layer. The bilinear interpolation ensures that the offset sampling position can obtain continuous spatial features.

[0117] For example, when the deformable convolution kernel is offset by 1.5 pixels at the sampling point in the northwest direction of an industrial park, the system takes the target position as the center and performs a weighted average of the concentration values ​​of the four surrounding original grid points. The weight is dynamically calculated based on the geometric relationship between the offset position and the original grid, so that the feature mapping results can reflect the directional sensitivity caused by the sampling point offset while maintaining spatial continuity. This process successfully identified the fan-shaped pollution area affected by the dominant wind direction when capturing the pollutant diffusion front of a sudden leakage incident in a chemical park.

[0118] S13134: Perform matrix multiplication on the mapped features and the weight coefficients to generate a direction-sensitive feature map.

[0119] In the embodiment of the present invention, the aerosol prediction system performs matrix multiplication on the mapped features and weight coefficients to generate a direction-sensitive feature map. The matrix multiplication operation realizes the direction-selective fusion of features.

[0120] For example, in a certain monsoon transition zone, the system assigns weight coefficients of 0.7 and 0.3 to the feature mapping results of the two dominant directions of southeast and northwest, respectively. The feature map generated by matrix multiplication and fusion accurately shows the pollutant transmission channel 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 daily variation characteristics of the boundary layer height observed by the lidar, which verifies the effectiveness of the weight allocation strategy.

[0121] In a preferred embodiment, the step of performing multi-stage recursive optimization on the spatiotemporal feature matrix by a hierarchical time-domain aggregation algorithm in S140 to generate an aerosol concentration prediction sequence for the target area includes: S141: Divide the space-time 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 smaller than the period length of the second periodic fluctuation component.

[0122] In S141, the aerosol prediction system divides the spatiotemporal characteristic matrix into the first-cycle fluctuation component and the second-cycle fluctuation component. Specifically, the system uses the empirical mode decomposition algorithm to extract the intrinsic mode function of the time dimension of the spatiotemporal characteristic matrix, where the first-cycle fluctuation component corresponds to the short-cycle changes on the hour to day scale (such as the vertical diffusion fluctuation of pollutants caused by the daily change of the boundary layer height), and the second-cycle fluctuation component represents the long-cycle evolution on the week to ten-day scale (such as the change of regional transmission intensity caused by the adjustment of atmospheric circulation).

[0123] For example, in a case of severe winter pollution in a plain urban agglomeration, after the decomposition of the spatiotemporal characteristic matrix, the first-cycle fluctuation component showed obvious 24-hour periodic oscillation from 08:00 on December 10 to 20:00 on December 11, reflecting the modulation effect of the day and night changes of the inversion layer on the ground PM2.5 concentration; while the second-cycle fluctuation component showed the maintenance of calm and stable weather caused by the continued southward pressure of the Siberian high pressure since December 5, with a cycle length of 120 hours, dominating the regional pollution accumulation trend.

[0124] S142: Dynamically attenuate the frequency domain energy of the first periodic fluctuation component through an adaptive weight allocator.

[0125] 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 wavelet transform coefficients and suppresses high-frequency noise components according to a preset threshold.

[0126] For example, when processing the first-cycle fluctuation component dominated by the daily changes in sea and land breezes in a coastal city, the system detected that the turbulent pulsation energy at the 30-minute scale accounted for more than 15% in the frequency domain, and the energy of the components with a period of less than 2 hours was attenuated by 40% by designing a Butterworth digital filter. At the same time, the system retains the energy distribution of the key frequency band of 4-12 hours to ensure the characteristic integrity of the sudden drop in pollutant concentration caused by the advancement of the sea breeze front (usually lasting 3-5 hours). This processing reduced the standard deviation of high-frequency fluctuations in the downwind area of ​​a certain industrial park from 28.7μg / m³ to 16.3μg / m³, significantly improving the smoothness of the forecast sequence.

[0127] S143: Using a residual connection structure, the attenuated first periodic fluctuation component and the second periodic fluctuation component are superimposed to generate an optimized feature vector.

[0128] In S143, the aerosol prediction system uses a residual connection structure to superimpose the attenuated first period fluctuation component and the second period fluctuation component to generate an optimized feature vector. The residual connection retains the key information of the physical process in the original spatiotemporal feature matrix through cross-layer direct connection.

[0129] For example, when superimposing the optimized features of a river valley area, the system linearly superimposes the attenuated 24-hour period component (characterizing the diurnal variation of valley wind) with the 10-day scale second period component (reflecting the cold air activity cycle), and introduces the terrain dynamic lifting effect parameters in the original feature matrix through residual connection. This operation enables the optimized feature vector to retain the 3-hour scale pollutant removal fluctuations at the time of the cold front crossing on December 15, and inherit the pollution retention trend on the leeward slope caused by the terrain, realizing the organic integration of multi-scale features.

[0130] S144: Input the optimized feature vector into a mixed density network to generate probability distribution parameters of aerosol concentration.

[0131] In S144, the aerosol prediction system inputs the optimized feature vector into the mixed density network to generate the probability distribution parameters of the aerosol concentration. The mixed density network adopts a multi-branch fully connected structure to model the uncertainty effects of meteorological conditions and pollution source emissions respectively.

[0132] For example, during the Spring Festival in a megacity, the optimized feature vector includes complex factors such as industrial shutdown and traffic surge. The mixed density network outputs the weights, mean and variance parameters of the three Gaussian distribution components. The network generated a mean of 65μg / m³ and a variance of 12.3μg² / m in the forecast on January 25 (the third day of the first lunar month). 6 The distribution parameters can accurately reflect the statistical characteristics of the dramatic fluctuations in concentration during the concentrated discharge period of fireworks and firecrackers.

[0133] S145: Performing 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.

[0134] In S145, the aerosol prediction system performs Monte Carlo sampling on the historical observation sequence according to the probability distribution parameters to generate the aerosol concentration prediction sequence for the target area. Monte Carlo sampling simulates a variety of possible meteorological-emission scenarios through a random number generator. For example, during the activation of a heavy pollution emergency plan in a certain area, the system performs 500 samplings to generate a probability prediction set. In the forecast on January 15, the sampling sequence showed that the probability of the southwest transmission channel being opened was 63%, corresponding to a median PM2.5 concentration of 153μg / m³, and the matching degree with the spatial transmission path of the subsequent actual pollution process was 89%. The 1-hour resolution feature of the prediction sequence successfully captured the concentration spike caused by accidental emissions in a certain industrial park, providing minute-level warning support for emergency control.

[0135] In a preferred embodiment, the step of inputting the optimized feature vector into a mixed density network to generate probability distribution parameters of aerosol concentration in S144 includes: S1441: Divide the optimized eigenvector into a meteorological driving component and a pollution source contribution component.

[0136] In S1441, the aerosol prediction system divides the optimized feature vector into a meteorological driven component and a pollution source contribution component. The division process is based on the correlation analysis of characteristic channels and prior physical knowledge. For example, the three-dimensional wind field and temperature field related characteristics are classified as meteorological driven components, while the industrial point source emission intensity and traffic flow characteristics are classified as pollution source contribution components. In the case of a port city, the meteorological driven 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 ship emission SO2 concentration and diesel vehicle flow. The system verifies through mutual information quantification and confirms that the division makes the information overlap between the two components less than 8%, meeting the requirements of independent modeling.

[0137] S1442: Perform nonlinear transformation on the meteorological driving component through a fully connected layer to generate mean and variance estimates.

[0138] In S1442, the aerosol prediction system uses a fully connected layer to perform nonlinear transformation on the meteorological driving component to generate mean and variance estimates. The fully connected layer uses the LeakyReLU activation function to introduce nonlinear mapping capabilities, such as transforming the 850hPa wind speed, boundary layer height and relative humidity characteristics into the conditional mean of PM2.5 concentration. In a dust transmission case, the system output mean value jumped from the background value of 35μg / m³ to 182μg / m³ based on the 700hPa jet axis position parameter in the meteorological driving component, and the variance was simultaneously expanded to 45μg² / m 6, accurately reflecting the concentration mutation characteristics before and after the arrival of the dust front.

[0139] S1443: Modeling the time lag effect of the pollution source contribution component using a gated linear unit to generate a skewness coefficient.

[0140] In S1443, the aerosol prediction system uses gated linear units 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 the emission source strength through time series convolution. For example, the PM2.5 emission changes caused by the start-up and shutdown operations of the sintering machine of a steel company, the system models its 6-8 hour lag effect and generates positive skewed distribution parameters. In the maintenance event on March 12, the pollution source contribution component showed a sudden drop of 70% in SO2 emissions. The gated linear unit output skewness coefficient of -0.37, accurately characterizing the left-skewed characteristics of the concentration distribution, and the skewness error between the predicted results and the ground monitoring station observations was less than 0.05.

[0141] S1444: Combining the mean, the variance, and the skewness coefficient into a three-parameter lognormal distribution, and generating probability distribution parameters of aerosol concentration through the three-parameter lognormal distribution.

[0142] In S1444, the aerosol prediction system combines the mean, variance, and skewness coefficients into a three-parameter lognormal distribution and generates probability distribution parameters for aerosol concentrations. The three-parameter lognormal distribution adapts to asymmetric concentration distributions by introducing position parameters. For example, accidental emissions from a coal-fired power plant cause the concentration distribution to be right-skewed (skewness +1.2). The system adjusts the position parameters to make the distribution peak match the monitored value. In the December 28 forecast, the distribution generated a 5-95% confidence interval of [48,215] μg / m³, which fully covered the maximum value of 209 μg / m³ measured on that day, an increase of 22 percentage points over the traditional normal distribution forecast interval coverage.

[0143] In an alternative embodiment, after generating the aerosol concentration prediction sequence of the target area, the method further includes: S210: Obtain real-time meteorological reanalysis data stream and satellite aerosol optical thickness observations.

[0144] In an embodiment of the present invention, an aerosol prediction system obtains real-time meteorological reanalysis data streams and satellite aerosol optical thickness observations. The real-time meteorological reanalysis data stream is output every half hour by the high-resolution numerical model of the European Center for Medium-Range Weather Forecasts, and contains parameters such as 10-meter wind speed, 2-meter temperature, and boundary layer height for a 0.1° grid in the target area. For example, during the emergency response to heavy pollution in a coastal urban agglomeration, the system accesses the ERA5 reanalysis data stream from 08:00 on December 15 to 08:00 on December 16 in real time, and simultaneously obtains the 550nm aerosol optical thickness observations updated every 10 minutes by the Sunflower 8 geostationary satellite. After being geographically corrected, the satellite data forms a spatial complement to the PM2.5 hourly concentration data of the ground monitoring station, covering the traditional monitoring blind spots in offshore ship emission areas.

[0145] S220: Inputting the real-time meteorological reanalysis data stream into a pre-trained spatiotemporal feature fusion network to extract a real-time spatiotemporal feature matrix.

[0146] In an embodiment of the present invention, the aerosol prediction system inputs the real-time meteorological reanalysis data stream into a pre-trained spatiotemporal feature fusion network to extract a real-time spatiotemporal feature matrix. The spatiotemporal feature fusion network uses the same architectural parameters as the offline training phase 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 16, the three-dimensional dilated convolution layer in the network captures the sudden rise in boundary layer height caused by the passage of the cold air front, and its vertical profile shows that the mixing layer height rises rapidly from 200 meters at 06:00 to 850 meters at 06:30. At the same time, the VisionTransformer module identifies the interaction between the urban heat island circulation and the sea breeze circulation, and generates a 64-dimensional feature vector representing the advancement speed of the sea breeze front in the real-time spatiotemporal feature matrix.

[0147] S230: Dynamically aligning the real-time spatiotemporal feature matrix with the historical spatiotemporal feature matrix through a sliding time window to obtain a dynamic alignment result.

[0148] In an embodiment of the present invention, the aerosol prediction system dynamically aligns the real-time spatiotemporal feature matrix with the historical spatiotemporal feature matrix through a sliding time window to obtain a dynamic alignment result. The sliding time window is based on the current time and slides forward 6 hours to construct a time series segment. For example, in the real-time prediction task at 07:00 on December 16, the system selects the spatiotemporal feature matrix from 07:00 to 13:00 on December 10, 2021 in the historical database (corresponding to a cold front transit process), and uses a dynamic time warping algorithm to non-linearly align the real-time data 06:00-12:00 segment with the historical data 07:00-13:00 segment. The alignment results show that the current front movement speed is 1.2 times faster than the historical case. The system adjusts the time offset accordingly, so that the timestamp deviation of the front reaching the coastline is reduced from 38 minutes to less than 5 minutes.

[0149] S240: Based on the dynamic alignment result, the error covariance matrix of the aerosol concentration prediction sequence is iteratively updated using a Kalman filter algorithm.

[0150] In an embodiment of the present invention, the aerosol prediction system uses the Kalman filter algorithm to iteratively update the error covariance matrix of the aerosol concentration prediction sequence based on the dynamic alignment result. The Kalman filter algorithm fuses the statistical characteristics of the predicted value and the observed value through the state space model. For example, the satellite aerosol optical thickness observation value (0.82±0.15) at 06:30 on December 16 is analyzed for differences with 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 28μg / m³, which is significantly higher than other areas. Based on this, the directional update mechanism of the covariance matrix is ​​initiated.

[0151] S250: performing online correction on the aerosol concentration prediction sequence of the target area according to the iteratively updated error covariance matrix.

[0152] In an embodiment of the present invention, the aerosol prediction system performs online correction on the aerosol concentration prediction sequence of the target area according to the iteratively updated error covariance matrix. After the error covariance matrix is ​​updated by the Kalman filter, it reflects the prediction uncertainty distribution of each spatial grid.

[0153] For example, the forecast error variance matrix of a port city at 09:00 shows that the variance of the container terminal area is , while the residential area is only . Based on the updated Kalman gain coefficient (0.68 for docks and 0.32 for residential areas), the system linearly merges the initial predicted value of 135μg / m³ in the dock area with the observed value of 148μg / m³, and obtains the corrected value of 135×0.32+148×0.68=143.2μg / m³, which reduces the absolute error with the 09:10 ship emission monitoring value of 145μg / m³ to 1.8μg / m³. The root mean square error of the corrected prediction sequence for the entire area is reduced from 18.7μg / m³ to 9.3μg / m³.

[0154] In a preferred embodiment, the iterative updating of the error covariance matrix of the aerosol concentration prediction sequence by using the Kalman filter algorithm in S240 includes: S241: Generate parameterized expressions of aerosol concentration observation model and state transition model.

[0155] In S241, the aerosol prediction system generates parameterized expressions of the aerosol concentration observation model and the state transition model. The observation model establishes a linear regression relationship between PM2.5 concentration and satellite aerosol optical thickness. For example, the conversion coefficient calibrated based on 3 years of observation data of a certain urban agglomeration is 89.3μg / (m³·AOD). The state transition model describes the time evolution of pollutant concentration through an autoregressive equation, and its order is determined to be 3rd order according to the Bayesian information criterion, reflecting that the weights of the current concentration affected by the state of the previous three hours are 0.55, 0.32, and 0.13 respectively. The model parameters are updated every hour to adapt to the dynamic changes in meteorological conditions.

[0156] S242: Determine the new information covariance matrix based on the real-time satellite aerosol optical thickness observations.

[0157] In S242, the aerosol prediction system determines the innovation covariance matrix based on the real-time satellite aerosol optical thickness observations. The innovation covariance matrix characterizes the statistical characteristics of the difference between the observed and predicted values, and its calculation is based on the historical residual data in the sliding window. For example, when the Sunflower-8 satellite passed at 07:10 on December 16, the system compared the predicted concentration (mean 112μg / m³) and the inverted concentration (mean 127μg / m³) of 15 grids in the coastal area, and calculated the regional innovation variance to be , the diagonal elements of the covariance matrix show that the error variance in the port area is as high as , indicating that this area requires focused correction.

[0158] S243: Detect the positive definiteness of the innovation covariance matrix through Cholesky decomposition.

[0159] In S243, the aerosol prediction system detects the positive definiteness of the new information covariance matrix through Cholesky decomposition. Cholesky decomposition requires that the matrix is ​​symmetric and positive definite, which may cause decomposition failure when numerical calculation errors or abnormal observation noise occur. For example, in the update cycle of 07:30 on December 16, the system performed a decomposition on the 6×6 new information covariance matrix of the port area and detected that the element in the 3rd row and 3rd column was not positive definite due to the sudden fluctuation of ship emissions. The system records the minimum eigenvalue during the decomposition process as -0.17, triggering the regularization process.

[0160] S244: When it is detected that the new information covariance matrix is ​​non-positive, a regularization factor is used to diagonally load the prior estimate to obtain a target covariance matrix.

[0161] In S244, when the aerosol prediction system detects that the new information covariance matrix is ​​non-positive, it uses a regularization factor to diagonally load the prior estimate to obtain the target covariance matrix. Diagonal loading ensures the positive definiteness of the matrix by adding small positive values ​​to the diagonal of the covariance matrix. For example, for the above-mentioned port area problem, the system selects the regularization factor λ=0.05 and increases the absolute value of each diagonal element of the original covariance matrix by 5%. The minimum eigenvalue of the processed matrix is ​​increased to 0.12, which is successfully verified by Cholesky decomposition to form a target covariance matrix that can be used for Kalman gain calculation.

[0162] S245: Update the Kalman gain coefficient according to the target covariance matrix to generate a corrected aerosol concentration prediction sequence.

[0163] 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 observation value to the prediction result, and its calculation is based on the updated covariance matrix.

[0164] For example, the innovation covariance of the port area is , the prediction error variance is , the Kalman gain coefficient was calculated to be 0.47. The system applied this coefficient to the forecast value at 08:00 on December 16, correcting the original forecast value of 135μg / m³ to 135+0.47×(127-135)=131.2μg / m³, and the deviation from the actual value of 130μg / m³ measured at the ground monitoring station at 08:10 was reduced from 5μg / m³ to 1.2μg / m³. After iterative updates in the entire region, the final output of the revised forecast sequence was verified at 09:00 on December 16, and the spatial correlation coefficient was increased from 0.82 to 0.91, and the root mean square error was reduced by 32%, achieving online optimization of the forecast accuracy.

[0165] In a preferred embodiment, the step of performing online correction on the aerosol concentration prediction sequence of the target area according to the iteratively updated error covariance matrix in S250 includes: S251: Obtaining 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.

[0166] 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 Center for Medium-Range Weather Forecasts, and simultaneously accesses the PM2.5 hourly mean observation data of the ground environmental monitoring station network. For example, at the downwind monitoring station in the industrial park of a port city, the actual measured PM2.5 concentration is 138μg / m³ at the current moment, while the initial prediction value is 125μg / m³. The system records the spatial distribution difference between the observed value and the predicted value at that moment as the benchmark input for subsequent corrections.

[0167] S252: generating a weight distribution ratio of the aerosol concentration observation value at the current moment according to the Kalman gain coefficient of the error covariance matrix; performing linear weighted fusion of the initial prediction value and the aerosol concentration observation value based on the weight distribution ratio to generate a corrected aerosol concentration value at the current moment.

[0168] In S252, the aerosol prediction system generates the weight distribution ratio of 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 of 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 for 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 125μg / m³ and the observation value of 138μg / m³ at 08:00, and generates a corrected concentration value of 125×0.38+138×0.62=132.9μg / m³. This value is 67% lower than the initial prediction deviation, and the absolute error with the actual measured value of 133μg / m³ by the minute-level mobile monitoring vehicle at 08:10 is less than 0.5μg / m³.

[0169] S253: extracting the residual component of the corrected aerosol concentration value and the spatial distribution residual of the satellite aerosol optical thickness observation value; performing feature mapping on the spatial distribution residual through the spatiotemporal feature fusion network to generate a residual spatiotemporal correlation matrix.

[0170] In S253, the aerosol prediction system extracts the residual component of the corrected aerosol concentration value and the spatial distribution residual of the satellite aerosol optical thickness observation value, and generates a residual spatiotemporal correlation matrix through the spatiotemporal feature fusion network. The residual component is calculated by the spatial difference between the corrected predicted value and the ground measured value. For example, the residual of an industrial park at 09:00 is +15.2μg / m³, while the satellite inversion residual is -0.23AOD. The system inputs the two into the spatiotemporal fusion network. The three-dimensional convolution layer in the network detects the abnormal boundary layer height in the vertical direction (actual value 850mvs predicted value 720m), and generates a 32-dimensional feature vector representing the insufficient vertical diffusion capacity. The attention mechanism module strengthens the residual correlation within a range of 5km downwind of the industrial park. The residual spatiotemporal correlation matrix finally output shows that there is a residual positive correlation zone with a continuous 3-hour and extended spatial range in this area, providing directional feedback for subsequent model optimization.

[0171] S254: Inputting the residual spatiotemporal correlation matrix into the hierarchical time-domain aggregation algorithm to extract the periodic fluctuation pattern of the residual component; dynamically adjusting the time attenuation factor of the Kalman gain coefficient according to the periodic fluctuation pattern.

[0172] 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 1.5-hour quasi-period, the system adjusts the decay factor from 0.85 to 0.72, extending the effective historical window to 3 periods (4.5 hours). During the 08:30 correction process, this adjustment increased the residual mode weight of the 07:00 morning traffic rush hour by 15%, improving the forecast adaptability to the current evening traffic rush hour.

[0173] S255: using the adjusted time attenuation factor to update the sliding window length of the error covariance matrix; performing sliding 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.

[0174] In S255, the aerosol prediction system uses the adjusted time decay factor to update the sliding window length of the error covariance matrix and performs a sliding average filter on the next moment prediction sequence. The sliding window is extended from the default 6 hours to 8 hours, covering 3 complete daily boundary layer variation cycles. The system performs a weighted average of 12 samples in the window for the 09:00 prediction value, of which the 08:00 correction value has a weight of 41%. The standard deviation of the filtered prediction value in residential areas is reduced from 18.7μg / m³ to 9.3μg / m³, effectively suppressing short-term fluctuations caused by sudden meteorological changes.

[0175] S256: performing time series splicing on the smoothed aerosol concentration prediction value and the corrected aerosol concentration value to generate an online corrected target area aerosol concentration prediction sequence.

[0176] In S256, the aerosol prediction system splices the smoothed aerosol concentration prediction value with the corrected aerosol concentration value in time series to generate an online corrected target area aerosol concentration prediction sequence. For example, the 08:00 corrected value 132.9μg / m³ and the 09:00 smoothed prediction value 128.4μg / m³ are spliced ​​to form a continuous sequence. In the verification of the 09:30 ground monitoring value of 129μg / m³, the absolute error of this sequence is 0.6μg / m³, and the spatial correlation coefficient is maintained above 0.93, achieving the optimization of the temporal coherence and spatial consistency of the prediction results.

[0177] In a more specific embodiment, the step of performing feature mapping on the spatial distribution residuals through the spatiotemporal feature fusion network in S253 to generate a residual spatiotemporal correlation matrix includes: S2531: Decompose the spatial distribution residual into a multi-channel residual tensor of spatial residual components and temporal residual components; 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.

[0178] In S2531, the aerosol prediction system decomposes the spatial distribution residual into a multi-channel residual tensor of spatial residual components and time series residual components, and performs a multi-scale hole convolution operation on the spatial residual components. 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 +5.2μg / m³ and the residual in the inland industrial area is -3.8μg / m³. The system uses three sets of hole convolution kernels with expansion rates of 2, 4, and 8 to extract the characteristics of the pollutant diffusion front at the scales of 500m, 1km, and 2km in the port area, respectively. In the data at 08:00 on December 16, the convolution kernel with an expansion rate of 4 captured the 1km scale concentration gradient mutation zone caused by the advancement of the sea breeze front, and its spatial gradient amplitude reached 28μg / (m³·km).

[0179] S2532: Generate a spatial attention weight matrix based on the local directional consistency of the spatial gradient mutation feature.

[0180] In S2532, the aerosol prediction system generates a spatial attention weight matrix based on the local directional consistency of the spatial gradient mutation characteristics. The local directional consistency is determined by calculating the angular variance of the gradient vector in a 3×3 window. For example, the gradient direction variance in a transportation hub area is 15 degrees, generating a weight coefficient of 0.92; while the gradient direction in the farmland area is disordered (variance 85 degrees), and the weight is reduced to 0.31. The system multiplies the 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.

[0181] S2533: Perform a time series causal convolution operation on the time series residual component of the multi-channel residual tensor to extract the lag time correlation feature; generate a time series convolution feature map based on the periodic intensity distribution of the lag time correlation feature.

[0182] In S2533, the aerosol prediction system performs a time series causal convolution operation on the time series residual component of the multi-channel residual tensor to extract the lagged time correlation features. The time series causal convolution uses a convolution kernel with a width of 5 and only accesses the current and historical time data. For example, when processing the 08:00 time series residual, the convolution kernel detects that the residual continues to grow positively from 07:30 to 07:50, generating a strong correlation feature that lags 2 time steps. The system further calculates the periodic intensity of the feature and finds that the residual fluctuation period corresponding to the hourly traffic peak is 24 hours, based on which the time series convolution feature map is generated.

[0183] S2534: Perform cross-modal feature concatenation on the spatial attention weight matrix and the temporal convolutional feature map to generate a spatiotemporal cross-correlation feature block; perform adaptive normalization processing on the channel dimension of the spatiotemporal cross-correlation feature block to generate a normalized spatiotemporal feature vector; perform forward and backward dependency modeling on the normalized spatiotemporal feature vector through a bidirectional long short-term memory network to extract global temporal dependency features.

[0184] In S2534, the aerosol prediction system performs cross-modal feature splicing of the spatial attention weight matrix and the temporal convolution feature map to generate a spatiotemporal cross-correlation feature block. The spliced ​​feature block contains 32 spatial channels and 16 temporal channels, and the system performs adaptive normalization of the channel dimensions. For example, the maximum value of the spatial channel in the port area is scaled to 0.87, and the mean value of the temporal channel is adjusted to 0.12 to eliminate dimensional differences. The normalized feature vector is input into the bidirectional long short-term memory network to capture the forward accumulation and backward propagation effects of the residual evolution.

[0185] S2535: Use 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; perform feature addition operation on the spatial node embedding feature and the global temporal dependency feature to generate an enhanced spatiotemporal feature sequence; generate a three-dimensional feature encoding vector based on the spatiotemporal dimension distribution of the enhanced spatiotemporal feature sequence.

[0186] In S2535, the aerosol prediction system uses a graph convolution layer to aggregate the spatial adjacency relationship of global time-dependent features in a topological structure. 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 probability of pollutant transmission between nodes is calculated as the edge weight. The system aggregates neighborhood features through three layers of graph convolution, and the generated spatial node embedding features show that the residual contribution rate of industrial areas to downwind residential areas is 73%. This feature is added to the temporal features output by the bidirectional long short-term memory network to form an enhanced spatiotemporal feature sequence.

[0187] S2536: Perform nonlinear projection on the three-dimensional feature encoding vector through a multilayer perceptron to generate a residual spatiotemporal association matrix; wherein each element of the residual spatiotemporal association matrix represents the residual association strength of different spatial positions and time steps.

[0188] In S2536, the aerosol prediction system uses a multilayer perceptron to perform nonlinear projection on the three-dimensional feature encoding vector to generate a residual spatiotemporal correlation matrix. The multilayer perceptron contains 128 hidden units and uses the GeLU activation function. The projected matrix elements represent the residual correlation strength of different spatial positions and time steps. For example, the correlation coefficient between the residual of the port area at 08:00 and the residual of the industrial area at 07:30 is 0.68, while the correlation coefficient with the residual of the residential area at 07:00 is only 0.12. This matrix accurately quantifies the spatiotemporal delay effect of the pollution transmission path.

[0189] S2537: Input the residual spatiotemporal correlation matrix into the hierarchical time-domain aggregation algorithm to extract the periodic fluctuation pattern of the residual component.

[0190] In S2537, the aerosol prediction system inputs the residual spatiotemporal correlation matrix into the hierarchical time-domain aggregation algorithm to extract the periodic fluctuation pattern of the residual components. The algorithm uses the empirical wavelet transform to decompose three intrinsic mode functions, of which the 0.5-hour scale mode reflects the traffic emission pulse and the 24-hour mode corresponds to the daily variation cycle of the boundary layer. The system identifies that the residual in the industrial area has a 1.5-hour quasi-periodic fluctuation in the period of 08:00-09:00, and adjusts the time attenuation factor of the Kalman gain accordingly.

[0191] 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: S1351: Decomposing the cross-modal time series correlation features into a time series correlation high-frequency component and a time series correlation low-frequency component.

[0192] S1352: Generate a two-dimensional position encoding matrix according to the spatial dimension distribution of the multimodal fusion feature tensor.

[0193] 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.

[0194] S1354: Perform feature channel splicing on the temporally correlated 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.

[0195] 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.

[0196] S1356: Expand the weighted spatial coding feature map along the time dimension into a spatiotemporal coding sequence; 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.

[0197] 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.

[0198] 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.

[0199] S1359: Detect the spatial continuity interruption area of ​​the initial estimate of the space-time feature matrix; generate a spatial interpolation mask template according to the geometric center coordinates of the interruption area; use the spatial interpolation mask template to perform edge-guided repair on the initial estimate of the space-time feature matrix to generate a final space-time feature matrix.

[0200] In S1351, the aerosol prediction system decomposes the cross-modal time series correlation features into time series correlation high-frequency components and time series correlation low-frequency components. Specifically, the system uses empirical wavelet transform to decompose the time series features in the frequency domain, where the high-frequency components capture rapid changes on the scale of minutes to hours (such as instantaneous fluctuations in concentration caused by turbulent pulsations), and the low-frequency components represent slow evolutions above the daily scale (such as adjustments to the atmospheric circulation background field). For example, during the advancement of the sea breeze front in a port city, the high-frequency component fluctuates with a 10-minute period during the 08:30-09:00 period, reflecting the sudden drop-recovery process of PM2.5 concentration caused by the passage of the gust front; the low-frequency component shows the southeast wind background field that has continued to strengthen since December 15, dominating the offshore transport trend of pollutants.

[0201] In S1352, the aerosol prediction system generates a two-dimensional position encoding matrix based on the spatial dimension distribution of the multimodal fusion feature tensor. The position encoding matrix uses a sinusoidal position embedding algorithm to convert geographic coordinates into a high-dimensional vector space expression. For example, for a 500m×500m UTM grid of a certain urban agglomeration, the system generates a 64-dimensional position encoding vector containing longitude 48.25°E and latitude 32.17°N information for each grid cell. The matrix automatically distinguishes between land and sea grids in coastal areas, and the cosine similarity of the encoding vector is less than 0.2, ensuring that the model can recognize the blocking effect of the coastline on the diffusion of pollutants.

[0202] In S1353, the aerosol prediction system superimposes the time-series-correlated high-frequency components with 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 by cascading deep convolution and point-by-point convolution, for example, using a 3×3×3 convolution kernel to slide along the spatial-frequency domain dimension. In a sudden emission event in an industrial park, this operation captured the jagged diffusion edge of the pollution cluster along the dominant wind direction at a scale of 1 km, and its spatial gradient amplitude was 23% higher than that of traditional convolution, accurately reflecting the concentration mutation characteristics around the factory boundary.

[0203] In S1354, the aerosol prediction system performs channel splicing of the time-series correlated low-frequency components and the multimodal fusion feature tensor to generate a low-frequency fusion feature block, and performs a cross-channel attention weighted 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, increasing the attention weights of the boundary layer height feature channel and the three-dimensional wind field channel to 0.75. In a valley area, the feature map clearly shows the pollution accumulation zone at the foot of the mountain caused by the development of the night inversion layer, and its spatial morphology is highly consistent with the vertical distribution of pollutants observed by ground-based lidar.

[0204] In S1355, the aerosol prediction system performs multi-scale feature fusion of spatial high-frequency texture features and spatial low-frequency correlation feature maps to generate multi-resolution spatial coding features, and performs weighted summation through dynamic channel weight vectors. Multi-scale fusion adopts a pyramid pooling structure, for example, extracting features at three scales of 1km, 2km, and 4km and upsampling to the original resolution. The dynamic weight is dynamically adjusted according to the information entropy of the feature channel. The 1km fine-grained feature weight of a certain transportation hub area reaches 0.68, which is significantly higher than the 0.12 of the 4km regional background field, ensuring that the model focuses on the fine structure of key pollution source areas.

[0205] In S1356, the aerosol prediction system expands the weighted spatial coding feature map into a spatiotemporal coding sequence along the time dimension, and performs a bidirectional dilated convolution operation to extract the full time series correlation feature vector. The bidirectional dilated convolution uses a causal convolution kernel with an expansion 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 16, the forward convolution identifies the clearing effect triggered by the passage of the cold front at 07:50, and the backward convolution predicts the pollution return caused by the decline of the sea breeze at 09:10. The feature vector generated by splicing the two accurately quantifies the impact period of 3 hours before and after the front passes.

[0206] In S1357, the aerosol prediction system performs a spatial dimension reorganization operation on the full time series correlation feature vector to restore the original resolution, and generates an enhanced spatiotemporal coding matrix by residual connection with the weighted spatial coding feature map. Spatial reorganization is achieved through transposed convolution, for example, converting a 64×64×256 feature tensor into a 512×512×128 matrix. The residual connection retains the terrain dynamic lift parameters in the original code, so that the vortex retention characteristics in the leeward slope of a mountain range are doubly enhanced in the enhanced matrix, and its activation value is increased by 1.8 times compared with the baseline.

[0207] In S1358, the aerosol prediction system performs local response normalization on the enhanced spatiotemporal coding matrix and inputs it into the multilayer perceptron to generate an initial estimate of the spatiotemporal feature matrix. Local response normalization suppresses interference from channels with too high activation values, such as compressing the maximum value of the channel around an industrial point source from 12.7 to 4.3. The multilayer perceptron uses the GeLU activation function for nonlinear projection to map the 256-dimensional input to a 128-dimensional feature space. The generated feature matrix accurately characterizes the coupling effect of ship emissions and sea breeze transport in the port area.

[0208] In S1359, the aerosol prediction system detects the spatial continuity interruption area of ​​the initial estimation of the spatiotemporal feature matrix and generates a spatial interpolation mask template for edge-guided repair. The interruption area refers to the characteristic value mutation zone caused by data loss or model error, such as a 500m×500m blank area in a cross-sea bridge area due to satellite inversion failure. The system extracts the minimum circumscribed rectangle of the interruption area through morphological operations, generates a direction-sensitive interpolation mask along the direction of the bridge, and ensures that the repair result conforms to the actual diffusion path.

[0209] It can be understood that S1351-S1359 significantly improves the spatiotemporal modeling capability of aerosol concentration prediction through multi-scale spatiotemporal feature fusion and dynamic coding optimization. First, by decomposing the high-frequency and low-frequency components of the time series features, the minute-level fluctuations and the daily evolution laws are captured respectively, and the geographic 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 of low-frequency correlation features is used to achieve multi-resolution feature fusion. The dynamic channel weight vector adaptively allocates the contribution of features of different scales, and the bidirectional hole convolution modeling full temporal dependency enhances the response capability to meteorological mutation events. The residual connection retains the details of the original physical process, the local response normalization suppresses noise interference, and the multi-layer perceptron completes the nonlinear projection, and finally generates a high-dimensional spatiotemporal feature matrix. By iteratively detecting and repairing the interrupted areas of spatial continuity, the edge gradient-guided interpolation mask and Poisson reconstruction are used to eliminate artifacts caused by missing data or model errors, ensure the spatial coherence and physical consistency of the feature matrix, and provide high-precision input for subsequent predictions.

[0210] As an optional but non-limiting embodiment, the step of performing edge-guided repair on the initial estimate of the spatiotemporal feature matrix using the spatial interpolation mask template in S1359 to generate a final spatiotemporal feature matrix includes: S13591: Detect the spatial continuity interruption area in the initial estimation of the spatiotemporal feature matrix, and generate an edge gradient amplitude map based on the pixel distribution of the interruption area; perform a morphological closing operation on the edge gradient amplitude map to fill the holes, generate a closed edge contour, and extract the minimum circumscribed rectangle of the closed edge contour.

[0211] S13592: Determine the geometric center coordinates of the interruption area based on the intersection of the diagonals of the minimum circumscribed rectangle, and generate a radial basis interpolation weight distribution map based on the geometric center coordinates; multiply the radial basis interpolation weight distribution map by the spatial interpolation mask template pixel by pixel to generate a direction-sensitive interpolation mask.

[0212] S13593: According to the weight value of each pixel in the direction-sensitive interpolation mask, perform adaptive anisotropic interpolation on the neighborhood pixels of the interrupted area to generate a preliminary repair feature map; extract the pixel point set overlapping with the closed edge contour in the preliminary repair feature map, and calculate the local texture consistency measure of the pixel point set based on the direction angle of the edge gradient amplitude map.

[0213] S13594: adjusting the weight distribution of the direction-sensitive interpolation mask according to the local texture consistency metric to generate an optimized interpolation mask; performing edge-guided Poisson equation reconstruction on the preliminary repair feature map using the optimized interpolation mask to generate an intermediate repair feature map.

[0214] S13595: Perform non-local mean filtering on the intermediate repair feature map to suppress interpolation noise, and calculate the residual energy between the filtered feature map and the initial estimate of the spatiotemporal feature matrix; dynamically adjust the filter strength parameter according to the spatial distribution of the residual energy to generate a smooth repair feature map.

[0215] S13596: Replace the pixel values ​​corresponding to the interrupted area in the smooth repair feature map into the initial estimate of the spatiotemporal feature matrix to generate an updated spatiotemporal feature matrix; detect the remaining spatial discontinuous areas in the updated spatiotemporal feature matrix, and iteratively execute the steps of generating the edge gradient amplitude map and replacing it with the smooth repair feature map until all interrupted areas are repaired.

[0216] S13597: The updated spatiotemporal feature matrix generated by the last iteration is used as the final spatiotemporal feature matrix.

[0217] In S13591, the aerosol prediction system detects the spatial continuity interruption area in the initial estimate and generates the edge gradient amplitude map, and performs morphological closing operation to extract the closed edge contour. For example, a 200m wide strip interruption occurred on the west side of an industrial park due to a monitoring station failure. The gradient amplitude map shows that there are concentration gradient mutation edges on the north and south sides. A 5×5 circular structure element is used for closing operation, filling the internal holes and extracting the minimum circumscribed rectangle of the contour to accurately locate the geometric center of the interruption area.

[0218] 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 assigns weights according to the exponential decay law. The mask of a valley interruption area increases the weight to 0.9 downwind of the dominant wind and decreases to 0.3 upwind, ensuring that the interpolation process conforms to the directionality of pollutant transmission.

[0219] In S13593, the aerosol prediction system performs adaptive anisotropic interpolation based on the direction-sensitive mask to generate a preliminary repair 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, a 3:1 aspect ratio interpolation kernel is set along the direction of traffic flow. The texture consistency metric calculates the variance of the gradient direction in a 30° fan-shaped window and identifies the artifact area for secondary correction.

[0220] In S13594, the aerosol prediction system adjusts the interpolation mask weight distribution and performs Poisson equation reconstruction to generate an intermediate repair feature map. Poisson reconstruction uses the normal area as the boundary condition and solves the Laplace equation in the interruption area. For example, after the interruption repair of a residential area, the eigenvalue transition is smooth and continuous with the surrounding concentration field gradient, and the maximum residual is reduced from 35μg / m³ to 7μg / m³.

[0221] In S13595, the aerosol prediction system performs non-local mean filtering on the intermediate repair feature map to suppress noise, and dynamically adjusts the filter strength to generate a smooth repair feature map. The non-local mean performs weighted averaging by searching for similar image blocks. After the repair of a certain industrial area, the checkerboard artifacts caused by interpolation were effectively eliminated, and the peak signal-to-noise ratio was improved by 8.6dB.

[0222] In S13596, the aerosol prediction system replaces the smooth repair features with the initial estimate and detects the remaining discontinuity areas, iteratively performing the repair until all discontinuities are eliminated. For example, after three iterations in a coastal area, the area of ​​the discontinuity area was reduced from the initial 15km² to 0.2km², and the spatial continuity index reached 0.98.

[0223] In S13597, the aerosol prediction system uses the last iteration result as the final spatiotemporal feature matrix. In the verification at 09:00 on December 16, the spatial correlation coefficient of the matrix with the drone flight monitoring data reached 0.95, and the root mean square error was less than 5μg / m³, which completely retained the fine structure of the pollution source and the regional transmission characteristics, providing high-precision input for subsequent predictions.

[0224] With such a design, the above S13591-S13597 proposes an edge-guided iterative repair mechanism for the spatial discontinuous areas in the spatiotemporal feature matrix, which effectively improves the geometric rationality and robustness of the prediction results. The closed contour of the interrupted area is extracted by morphological closing operation, and the direction-sensitive mask is generated by combining radial basis interpolation. The adaptive anisotropic interpolation ensures that the repair result conforms to the actual diffusion path. The Poisson equation reconstruction uses the normal area as the boundary condition to restore the smooth transition characteristics of the interrupted area. The non-local mean filter suppresses the interpolation noise, and dynamically adjusts the filter intensity to balance denoising and detail retention. The iterative repair mechanism gradually eliminates the residual discontinuous areas until the spatial continuity index meets the standard. This method performs well in complex terrain and data missing scenarios. For example, after the coastal bridge interruption area is repaired, the pollutant diffusion path is consistent with the coastline, and the industrial area artifact elimination rate is significantly improved, thereby significantly improving the spatial correlation coefficient of the generated spatiotemporal feature matrix and reducing the root mean square error, providing a reliable foundation for high-precision prediction.

[0225] Based on the same inventive concept, an embodiment of the present invention also provides an aerosol prediction system. Figure 2 , which is a schematic diagram of a possible aerosol prediction system provided in an embodiment of the present invention. Figure 2 In the embodiment, the aerosol prediction system 200 includes: a processor 210 and a memory 220. The memory 220 stores a computer program executable by the processor 210, and the processor 210 can perform the steps of the aerosol prediction method based on artificial intelligence by executing the instructions stored in the memory 220.

[0226] 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 is run on an aerosol prediction system, the computer program is used to enable the aerosol prediction system to perform the steps of the aerosol prediction method based on artificial intelligence. In some possible embodiments, various aspects of the aerosol prediction method based on artificial intelligence 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 is run on an aerosol prediction system, the computer program is used to enable the aerosol prediction system to perform the steps of the aerosol prediction method based on artificial intelligence. For example, the aerosol prediction system can perform the following steps: Figure 1 Follow the steps shown in .

Claims

1. An aerosol prediction method based on artificial intelligence, characterized in that: The method is applied to an aerosol prediction system, and the method comprises: Acquiring an original multimodal meteorological observation dataset, wherein the original multimodal meteorological observation dataset includes a surface aerosol concentration sequence, a three-dimensional atmospheric state variable, and a boundary layer dynamic parameter; Performing spatiotemporal heterogeneity correction on the original multimodal meteorological observation dataset to generate a spatiotemporal aligned target multimodal meteorological observation dataset; Inputting the target multimodal meteorological observation dataset into a spatiotemporal feature fusion network to extract multiscale spatial distribution features and cross-modal temporal correlation features, and generating a spatiotemporal feature matrix based on the multiscale spatial distribution features and the cross-modal temporal correlation features; The spatiotemporal feature matrix is ​​recursively optimized in multiple stages through a hierarchical time-domain aggregation algorithm to generate an aerosol concentration prediction sequence for the target area.

2. The method according to claim 1, characterized in that The step of performing spatiotemporal heterogeneity correction on the original multimodal meteorological observation dataset to generate a spatiotemporal aligned target multimodal meteorological observation dataset includes: Identifying the spatial resolution difference between the vertically layered data of the three-dimensional atmospheric state variable and the surface aerosol concentration series; Performing spatial interpolation on the vertical layered data based on an earth curvature compensation algorithm to generate equal longitude and latitude grid data matching the surface aerosol concentration sequence; For the detected missing area of ​​the boundary layer dynamic parameter, using the eddy covariance data of the adjacent meteorological station to generate a filling mask for the missing area; Dynamic sliding window calibration is performed based on the timestamp deviation between the filling mask and the equal longitude and latitude grid data to generate a target multimodal meteorological observation dataset aligned in time and space.

3. The method according to claim 2, characterized in that The method of performing dynamic sliding window calibration based on the timestamp deviation between the filling mask and the equal longitude and latitude grid data to generate a target multimodal meteorological observation data set aligned in time and space includes: Detecting the time span of consecutive missing regions in the filling mask; Extracting similarity templates from the same historical period of equal latitude and longitude grid data according to the time span; Using a dynamic time warping algorithm to perform path matching between the similarity template and the current missing region; Adjust the step size parameter of the sliding window according to the slope change of the matching path; Based on the adjusted step size parameters, bidirectional linear interpolation is performed on the current missing area to generate a target multimodal meteorological observation dataset that is aligned in time and space.

4. The method according to claim 1, characterized in that: The target multimodal meteorological observation data set is input into a spatiotemporal feature fusion network to extract multiscale spatial distribution features and cross-modal temporal association features, and a spatiotemporal feature matrix is ​​generated based on the multiscale spatial distribution features and the cross-modal temporal association features, including: Performing local texture enhancement on the surface aerosol concentration sequence through a convolutional attention module to generate a super-resolution spatial feature map; Using a three-dimensional dilated convolution kernel to traverse the vertical profile data of the three-dimensional atmospheric state variable, extracting the correlation characteristics between the boundary layer thickness and the aerosol diffusion path; The super-resolution spatial feature map and the associated features are spliced ​​in channel dimension to generate a multimodal fusion feature tensor; Slidingly intercepting the time dimension of the multimodal fusion feature tensor based on a bidirectional gated recurrent unit to generate a cross-modal temporal correlation feature; The cross-modal temporal correlation features and the multimodal fusion feature tensor are spatially encoded to generate a spatiotemporal feature matrix.

5. The method according to claim 4, characterized in that The method of performing local texture enhancement on the surface aerosol concentration sequence by a convolutional attention module to generate a super-resolution spatial feature map includes: Performing Gaussian difference filtering on the surface aerosol concentration sequence to extract a spatial gradient amplitude map; generating a regional significance weight according to a local variance distribution of the spatial gradient magnitude map; Using a deformable convolution kernel to extract multi-directional features of the surface aerosol concentration sequence to generate a direction-sensitive feature map; Performing an element-by-element product of the regional saliency weight and the direction-sensitive feature map to generate a texture enhancement feature map; Improving the resolution of the texture enhancement feature map by a transposed convolution layer to generate a super-resolution spatial feature map; The method of extracting multi-directional features of the surface aerosol concentration sequence using a deformable convolution kernel to generate a direction-sensitive feature map includes: Initializing a sampling offset parameter set of the deformable convolution kernel; Calculating weight coefficients in each direction according to the spatial autocorrelation function of the surface aerosol concentration sequence; Performing position-sensitive feature mapping on the sampling offset parameters through a bilinear interpolation layer; The mapped features are matrix-multiplied by the weight coefficients to generate a direction-sensitive feature map.

6. The method according to claim 1, characterized in that The multi-stage recursive optimization of the spatiotemporal feature matrix by a hierarchical time-domain aggregation algorithm is performed to generate an aerosol concentration prediction sequence for the target area, including: Dividing the spatiotemporal characteristic matrix into a first periodic fluctuation component and a second periodic fluctuation component; wherein the period length of the first periodic fluctuation component is shorter than the period length of the second periodic fluctuation component; Dynamically attenuating the frequency domain energy of the first periodic fluctuation component by an adaptive weight allocator; Using a residual connection structure, the attenuated first periodic fluctuation component and the second periodic fluctuation component are superimposed to generate an optimized feature vector; Inputting the optimized feature vector into a mixed density network to generate probability distribution parameters of aerosol concentration; Performing 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 step of inputting the optimized feature vector into a mixed density network to generate probability distribution parameters of aerosol concentration includes: Dividing the optimized eigenvector into a meteorological driving component and a pollution source contribution component; Performing a nonlinear transformation on the meteorological driving component through a fully connected layer to generate mean and variance estimates; Modeling the time lag effect of the pollution source contribution component using a gated linear unit to generate a skewness coefficient; The mean, the variance and the skewness coefficient are combined into a three-parameter lognormal distribution, and probability distribution parameters of aerosol concentration are generated through the three-parameter lognormal distribution.

7. The method according to claim 1, characterized in that After generating the aerosol concentration prediction sequence of the target area, the method further includes: Obtain real-time meteorological reanalysis data streams and satellite aerosol optical thickness observations; Inputting the real-time meteorological reanalysis data stream into a pre-trained spatiotemporal feature fusion network to extract a real-time spatiotemporal feature matrix; Dynamically aligning the real-time spatiotemporal feature matrix with the historical spatiotemporal feature matrix through a sliding time window to obtain a dynamic alignment result; Based on the dynamic alignment result, the error covariance matrix of the aerosol concentration prediction sequence is iteratively updated using a Kalman filter algorithm; The aerosol concentration prediction sequence of the target area is corrected online according to the iteratively updated error covariance matrix.

8. The method according to claim 7, characterized in that The iterative updating of the error covariance matrix of the aerosol concentration prediction sequence by using the Kalman filter algorithm includes: Generate parameterized expressions of aerosol concentration observation model and state transition model; Determine the innovation covariance matrix based on real-time satellite aerosol optical thickness observations; Testing the positive definiteness of the innovation covariance matrix by Cholesky decomposition; When it is detected that the innovation covariance matrix is ​​non-positive, a regularization factor is used to diagonally load the prior estimate to obtain a target covariance matrix; The Kalman gain coefficient is updated according to the target covariance matrix to generate a corrected aerosol concentration prediction sequence.

9. The method according to claim 7, 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: Obtaining 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; generating a weight distribution ratio of the aerosol concentration observation value at the current moment according to the Kalman gain coefficient of the error covariance matrix; performing linear weighted fusion of the initial prediction value and the aerosol concentration observation value based on the weight distribution ratio to generate a corrected aerosol concentration value at the current moment; Extracting the residual component of the corrected aerosol concentration value and the spatial distribution residual of the satellite aerosol optical thickness observation value; performing feature mapping on the spatial distribution residual through the spatiotemporal feature fusion network to generate a residual spatiotemporal correlation matrix; Inputting the residual spatiotemporal correlation matrix into the hierarchical time-domain aggregation algorithm to extract the periodic fluctuation pattern of the residual component; dynamically adjusting the time attenuation factor of the Kalman gain coefficient according to the periodic fluctuation pattern; The sliding window length of the error covariance matrix is ​​updated using the adjusted time decay factor; based on the updated sliding window length, a sliding average filter is performed on the aerosol concentration prediction sequence at the next moment to generate a smoothed aerosol concentration prediction value; The smoothed aerosol concentration prediction value and the corrected aerosol concentration value are spliced ​​in time series to generate an online corrected target area aerosol concentration prediction sequence; The step of performing feature mapping on the spatial distribution residuals through the spatiotemporal feature fusion network to generate a residual spatiotemporal association matrix includes: Decomposing the spatial distribution residual into a multi-channel residual tensor of a spatial residual component and a temporal residual component; performing 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; Generate a spatial attention weight matrix according to the local direction consistency of the spatial gradient mutation feature; 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-modal feature spliced ​​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 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; Performing nonlinear projection on the three-dimensional feature encoding vector through a multi-layer perceptron to generate a residual spatiotemporal association matrix; wherein each element of the residual spatiotemporal association matrix represents the residual association strength of different spatial positions and time steps; The residual spatiotemporal correlation matrix is ​​input into the hierarchical time-domain aggregation algorithm to extract the periodic fluctuation pattern of the residual component.

10. An aerosol prediction system, characterized in that: It 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 executes the steps of any one of the methods of claims 1 to 9.

Images