A PM2.5 chemical component vertical profile inversion system and method based on graph neural network and adaptive meta-learning

Abstract

This paper discloses a system and method for inverting vertical profiles of PM2.5 chemical components based on graph neural networks and adaptive meta-learning. By constructing a dynamic spatial graph of vertical height layers and integrating it with a meta-learning framework, this system achieves rapid migration and high-precision inversion in data-scarce areas. The system utilizes heterogeneous graph neural networks to capture component diffusion patterns and reduces the need for annotated data through MAML optimization. This system addresses the limitations of traditional models, such as their lack of generalization and reliance on large amounts of data, and is widely applicable to air quality monitoring in complex geographic environments.

Description

Technical Field

[0001] The present invention relates to the field of environmental monitoring, and in particular to a system and method for inverting the vertical profiles of PM2.5 chemical components based on graph neural networks and adaptive meta-learning. Background Art

[0002] With the increasing severity of air pollution, especially the increasingly significant impact of PM2.5 (fine particulate matter) on human health and the environment, accurate monitoring and forecasting of PM2.5 has become a key research topic in the field of environmental monitoring. In recent years, the application of deep learning technology in air quality prediction and inversion has made significant progress. Traditional models, such as convolutional neural networks (CNNs) and long short-term memory networks (LSTMs), have been widely used to predict the spatiotemporal distribution and infer concentrations of PM2.5. However, these traditional models have several limitations:

[0003] Traditional deep learning models typically focus on processing two-dimensional data and fail to effectively capture the vertical diffusion and transport of PM2.5 chemical components within the atmosphere. In particular, the vertical transmission of pollutants between layers is influenced by meteorological conditions, topography, and atmospheric structure. However, traditional models lack effective spatial modeling capabilities, resulting in low physical consistency in their inversion results.

[0004] Existing deep learning models typically rely on large amounts of labeled data for training, and much of this training data comes from specific geographic regions. However, due to variations in geographical environments (such as plateaus and coastal areas) and climatic conditions, these models have limited generalization capabilities, making them difficult to adapt to pollution prediction needs across diverse regions and environments. Furthermore, existing models often perform poorly in data-scarce scenarios.

[0005] During sudden pollution events, the temporal and spatial evolution of pollutants changes dramatically. Traditional static deep learning models often cannot respond and adjust quickly enough to adapt to this rapidly changing environment. This results in poor accuracy and responsiveness of existing models during pollution outbreaks.

[0006] Therefore, how to design a model that can better capture the vertical diffusion process of PM2.5 and improve the generalization ability and adaptability of the model is a key issue that needs to be solved in the current PM2.5 monitoring and forecasting technology. Summary of the Invention

[0007] The purpose of the present invention is to provide a PM2.5 chemical component vertical profile inversion system and method based on graph neural network and adaptive meta-learning, which solves the above-mentioned technical problems pointed out in the prior art.

[0008] The present invention provides a PM2.5 chemical component vertical profile inversion system based on graph neural network and adaptive meta-learning, including a multi-source data acquisition module, a graph structure construction module, an adaptive meta-learning module and an inversion output module;

[0009] The multi-source data acquisition module is used to collect and acquire multi-source data on k vertical height layers of the current geographical area; pre-process the multi-source data to obtain pre-processed multi-source data;

[0010] The graph structure construction module is used to construct a graph structure by taking each of the vertical height layers as a node; and calculate the edge weight between the i-th node and the j-th node in the graph structure through the meteorological variable data collected during a preset time period. ;

[0011] The adaptive meta-learning module is used to Combining multi-source data of samples collected from multiple regions to build and train a target adaptive meta-learning model;

[0012] The inversion output module is used to process the preprocessed multi-source data of the current area at the current moment through the target adaptive meta-learning model, and output the predicted PM2.5 chemical component vertical profile inversion results.

[0013] Preferably, the multi-source data includes ground-based lidar optical parameters, meteorological variable data and local component concentration data.

[0014] Preferably, the ground-based lidar optical parameters include vertical profile data of 532nm aerosol backscattering coefficient, extinction coefficient and depolarization ratio;

[0015] The meteorological variable data include the integrated ERA5 reanalysis dataset;

[0016] The local component concentration data includes real-time monitoring values of near-ground PM2.5 chemical component concentrations.

[0017] Preferably, the edge weight is calculated as follows:

[0018] (1); among them, is the wind speed eigenvector of the i-th node, is the wind speed eigenvector of the jth node, is the bandwidth parameter, which controls the weight decay speed.

[0019] Preferably, the adaptive meta-learning module is specifically used to construct and train an initial adaptive meta-learning model using multi-source sample data collected from multiple regions;

[0020] Based on the edge weight Updating the initial adaptive meta-learning model to obtain a first adaptive meta-learning model;

[0021] Based on the multi-source data of the current region collected over n consecutive time periods, the parameters of the first adaptive meta-learning model are updated through the Adam optimizer to obtain a target adaptive meta-learning model.

[0022] Preferably, the adaptive meta-learning module is further used to collect and obtain sample multi-source data from multiple regions, wherein the sample multi-source data includes multi-source data from various regions monitored continuously for one year;

[0023] Preprocessing the sample multi-source data of each region and splitting them into support sets and query sets;

[0024] Constructing a first initial adaptive meta-learning model; using the support set and query set corresponding to each region to train the first initial adaptive meta-learning model and calculate the loss function, the model converges when the loss function is less than or equal to the minimum loss function threshold, and obtaining the initial parameters θ of the first initial adaptive meta-learning model corresponding to each region;

[0025] Based on the initial parameters θ, the initial parameters θ of the first initial adaptive meta-learning model are updated through two-layer optimization to obtain the target model parameters ;

[0026] Preferably, the loss function is calculated as follows:

[0027] ;

[0028] Where, is the importance weight of each component concentration; is the predicted sample local component concentration data of the i-th sample multi-source data; is the sample local component concentration data of the i-th sample multi-source data;

[0029] The target model parameters are calculated as follows:

[0030] ;in, ; are the initial parameters of the first initial adaptive meta-learning model, is the inner learning rate, is the loss gradient, where Represents the training missions for each region, For the first initial adaptive meta-learning model to target the task The loss dependence ratio.

[0031] Preferably, the adaptive meta-learning module is further used in the specific implementation process to calculate the edge weights of the graph structure. Get the k-th layer edge weight of the k-th vertical height layer ;

[0032] Based on the k-th layer edge weight The neighbor node set of the i-th vertical height layer is iteratively updated and calculated by the GraphSAGE algorithm to obtain the embedding value of the i-th node in the k-th vertical height layer;

[0033] The initial adaptive meta-learning model is updated based on the node embedding value to obtain a first adaptive meta-learning model.

[0034] Preferably, the node embedding value is calculated as follows:

[0035] Where, is the edge weight of the kth layer; is the set of neighbor nodes; is the embedding of node i in the k-1th layer, Indicates the jth node from the set of neighbor nodes; is an aggregate function, is the connection function, is a non-linear activation function.

[0036] Accordingly, the present invention also proposes a method for inverting vertical profiles of PM2.5 chemical components based on graph neural networks and adaptive meta-learning, which includes the following steps:

[0037] Collecting and obtaining multi-source data on k vertical height layers of the current geographical area; preprocessing the multi-source data to obtain preprocessed multi-source data;

[0038] Each of the vertical height layers is used as a node to construct a graph structure; and the edge weight between the i-th node and the j-th node in the graph structure is calculated using the meteorological variable data collected during a preset time period. ;

[0039] Based on the edge weight Combining multi-source data of samples collected from multiple regions to build and train a target adaptive meta-learning model;

[0040] The pre-processed multi-source data of the current area at the current moment is processed by the target adaptive meta-learning model, and the predicted PM2.5 chemical component vertical profile inversion result is output.

[0041] Compared with the prior art, the embodiments of the present invention have at least the following technical advantages:

[0042] From the analysis of the above-mentioned PM2.5 chemical component vertical profile inversion system and method based on graph neural network and adaptive meta-learning provided by the present invention, it can be seen that in specific applications, 60 layers of data with a layer of 100m in the vertical range of 0-6km are first obtained through ground-based lidar (optical parameters), meteorological monitoring stations (ERA5 data set) and ground monitoring stations (PM2.5 component concentrations), covering aerosol characteristics, meteorological conditions and near-ground component concentrations, providing multi-dimensional input for the model; then, by using 60 vertical height layers as nodes, a spatial topological network is constructed to explicitly express the correlation between different height layers (such as the vertical diffusion of aerosols is affected by the meteorological conditions of the upper and lower layers), thereby enhancing the model's ability to characterize physical processes; further, by utilizing Sample data from different geographical environments (such as cities, suburbs, and different climatic zones) are used to learn common cross-regional characteristics (such as meteorological-component correlation patterns), improve model generalization, and quickly adapt to new regions with a small amount of labeled data through meta-learning processing, solving the performance degradation problem of traditional methods caused by scarcity of regional data and reducing dependence on large amounts of labeled data from a single region; finally, the real-time pre-processed data (optical parameters, meteorological, and near-ground components) of the current region are input into the model, and the features of nodes at different altitudes are aggregated through the graph neural network (GNN) (such as the spatial correlation between aerosol optical properties and wind speed), combined with meta-learning to dynamically adjust the model parameters, and finally the 60-layer PM2.5 chemical component concentrations (such as sulfate, nitrate, etc.) are output through the fully connected layer mapping. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 Schematic diagram of the overall architecture of a PM2.5 chemical composition vertical profile inversion system based on graph neural network and adaptive meta-learning;

[0044] Figure 2 Schematic diagram of the main process of a PM2.5 chemical composition vertical profile inversion method based on graph neural network and adaptive meta-learning;

[0045] Figure 3 A schematic diagram of the CORR comparison curve simulation after fine-tuning in different regions in a PM2.5 chemical component vertical profile inversion method based on graph neural network and adaptive meta-learning;

[0046] Figure 4 A schematic diagram showing the resource usage comparison between model training and inference in a PM2.5 chemical component vertical profile inversion method based on graph neural network and adaptive meta-learning.

[0047] Reference numerals: multi-source data acquisition module 10 , graph structure construction module 20 , adaptive meta-learning module 30 , inversion output module 40 . DETAILED DESCRIPTION

[0048] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0049] The present invention will be further described in detail below through specific embodiments in conjunction with the accompanying drawings.

[0050] Example 1

[0051] like Figure 1 As shown, the first embodiment of the present invention provides a PM2.5 chemical component vertical profile inversion system based on graph neural network and adaptive meta-learning, including a multi-source data acquisition module 10, a graph structure construction module 20, an adaptive meta-learning module 30 and an inversion output module 40;

[0052] The multi-source data acquisition module 10 is used to collect and acquire multi-source data on k vertical height layers of the current geographical area; pre-process the multi-source data to obtain pre-processed multi-source data;

[0053] The graph structure construction module 20 is used to construct a graph structure by taking each of the vertical height layers as a node; and calculate the edge weight between the i-th node and the j-th node in the graph structure using the meteorological variable data collected during a preset time period. ;

[0054] The adaptive meta-learning module 30 is used to Combining multi-source data of samples collected from multiple regions to build and train a target adaptive meta-learning model;

[0055] The inversion output module 40 is used to process the pre-processed multi-source data of the current area at the current moment through the target adaptive meta-learning model, and output the inversion result of the predicted PM2.5 chemical component vertical profile.

[0056] The multi-source data includes ground-based lidar optical parameters, meteorological variable data and local component concentration data.

[0057] The optical parameters of the ground-based lidar include vertical profile data of 532nm aerosol backscattering coefficient, extinction coefficient and depolarization ratio;

[0058] The meteorological variable data include the integrated ERA5 reanalysis dataset;

[0059] The local component concentration data includes real-time monitoring values of near-ground PM2.5 chemical component concentrations.

[0060] The edge weight is calculated as follows:

[0061] (1); among them, is the wind speed eigenvector of the i-th node, is the wind speed eigenvector of the jth node, is the bandwidth parameter, which controls the weight decay speed.

[0062] The adaptive meta-learning module 30 is specifically used to construct and train an initial adaptive meta-learning model using sample multi-source data collected from multiple regions;

[0063] Based on the edge weight Updating the initial adaptive meta-learning model to obtain a first adaptive meta-learning model;

[0064] Based on the multi-source data of the current region collected over n consecutive time periods, the parameters of the first adaptive meta-learning model are updated through the Adam optimizer to obtain a target adaptive meta-learning model.

[0065] The adaptive meta-learning module 30 is further configured to collect and acquire sample multi-source data from multiple regions, wherein the sample multi-source data includes multi-source data from various regions monitored continuously for one year;

[0066] Preprocessing the sample multi-source data of each region and splitting them into support sets and query sets;

[0067] Constructing a first initial adaptive meta-learning model; using the support set and query set corresponding to each region to train the first initial adaptive meta-learning model and calculate the loss function, the model converges when the loss function is less than or equal to the minimum loss function threshold, and obtaining the initial parameters θ of the first initial adaptive meta-learning model corresponding to each region;

[0068] Based on the initial parameters θ, the initial parameters θ of the first initial adaptive meta-learning model are updated through two-layer optimization to obtain the target model parameters ;

[0069] The loss function is calculated as follows:

[0070] Where, is the importance weight of each component concentration; is the predicted sample local component concentration data of the i-th sample multi-source data; is the sample local component concentration data of the i-th sample multi-source data;

[0071] The target model parameters are calculated as follows:

[0072] ;

[0073] in, ; are the initial parameters of the first initial adaptive meta-learning model, is the inner learning rate, is the loss gradient, where Represents the training missions for each region, For the first initial adaptive meta-learning model to target the task The loss dependence ratio.

[0074] The adaptive meta-learning module 30 is also used in the specific implementation process to calculate the edge weights of the graph structure. Get the k-th layer edge weight of the k-th vertical height layer ;

[0075] Based on the k-th layer edge weight The neighbor node set of the i-th vertical height layer is iteratively updated and calculated by the GraphSAGE algorithm to obtain the embedding value of the i-th node in the k-th vertical height layer;

[0076] The initial adaptive meta-learning model is updated based on the node embedding value to obtain a first adaptive meta-learning model.

[0077] The node embedding value is calculated as follows:

[0078] Where, is the edge weight of the kth layer; is the set of neighbor nodes; is the embedding of node i in the k-1th layer, Indicates the jth node from the set of neighbor nodes; is an aggregate function, is the connection function, is a non-linear activation function.

[0079] In summary, the first embodiment of the present application provides a PM2.5 chemical component vertical profile inversion system based on graph neural network and adaptive meta-learning, which constructs a graph structure by utilizing multi-source data of multiple vertical height layers in the current area, models the spatial topological relationship of the vertical height layers according to the graph structure, and improves the physical rationality of the inversion; trains a pre-constructed lightweight model framework (i.e., an adaptive meta-learning model) with multi-source data of samples from multiple different regions, reduces dependence on labeled data, supports rapid migration of a small number of samples, and thereby realizes dynamic adaptive optimization of the model in multiple geographical environments.

[0080] Example 2

[0081] like Figure 2 As shown, the first embodiment of the present invention provides a method for inverting the vertical profiles of PM2.5 chemical components based on graph neural networks and adaptive meta-learning, including the following steps:

[0082] Step S10: collecting and acquiring multi-source data on k vertical height layers of the current geographical area; pre-processing the multi-source data to obtain pre-processed multi-source data;

[0083] The multi-source data includes ground-based lidar optical parameters, meteorological variable data and local component concentration data;

[0084] The optical parameters of the ground-based lidar include vertical profile data of 532nm aerosol backscattering coefficient, extinction coefficient and depolarization ratio;

[0085] The meteorological variable data includes the integrated ERA5 reanalysis data set (the integrated ERA5 reanalysis data set includes parameters such as temperature, humidity, wind speed, vertical velocity, etc.);

[0086] The local component concentration data includes real-time monitoring values of near-ground PM2.5 chemical component concentrations;

[0087] Specifically, the multi-source data is collected (using ground-based lidar, meteorological monitoring stations and ground monitoring stations) at intervals of one hour to collect 60 vertical layers of multi-source data at a vertical height of 6 km, with each layer being 100 meters.

[0088] There are 60 vertical height layers; each vertical height of the vertical height layer is 100m;

[0089] The preprocessing mentioned above refers to cleaning the multi-source data. For lidar data, Hampel filtering is used to remove outliers (window width is 12 hours, threshold is 3 times the median absolute deviation); for meteorological data, missing values are filled by linear interpolation.

[0090] It should be noted that the above-mentioned embodiment of the present application collects multi-source data from a vertical height of 0-6 km, wherein the vertical height of 0-6 km is divided into 60 vertical height layers with each layer being 100 meters. Each vertical height layer contains multi-source data information, namely, ground-based lidar optical parameters collected by ground-based lidar, meteorological variable data collected by meteorological monitoring stations, and local component concentration data collected by ground monitoring stations; thereby providing a data basis for subsequent PM2.5 monitoring and forecasting.

[0091] Step S20: constructing a graph structure by taking each of the vertical height layers as a node; and calculating the edge weight between the i-th node and the j-th node in the graph structure by using the meteorological variable data collected during a preset time period. ;

[0092] The edge weight is calculated as follows:

[0093] (1); among them, is the wind speed eigenvector of the i-th node, is the wind speed eigenvector of the jth node, is the bandwidth parameter, which controls the weight decay speed;

[0094] The above-mentioned embodiment of the application improves the physical rationality of the inversion by modeling the spatial topological relationship of vertical height layers through a graph structure;

[0095] Step S30: Based on the edge weight Combining multi-source data of samples collected from multiple regions to build and train a target adaptive meta-learning model;

[0096] It should be noted that the above-mentioned embodiments of the present application reduce dependence on labeled data by constructing a lightweight model framework, support rapid migration of a small number of samples, and jointly construct and train the target adaptive meta-learning model through multi-source data of samples from multiple regions to achieve dynamic adaptive optimization of the model in multiple geographical environments.

[0097] Step S40: Processing the pre-processed multi-source data of the current region at the current moment through the target adaptive meta-learning model, and outputting the inversion result of the predicted PM2.5 chemical component vertical profile;

[0098] It should be noted that the above-mentioned processing of the preprocessed multi-source data of the current region at the current moment through the target adaptive meta-learning model to output the predicted PM2.5 chemical component vertical profile inversion result is obtained by inputting the preprocessed multi-source data of the current region at the current moment into the target adaptive meta-learning model, and then, after calculation and processing by the target adaptive meta-learning model, mapping through the fully connected layer of the target adaptive meta-learning model to obtain the concentration value of each chemical component (the concentration value of each chemical component is the PM2.5 chemical component vertical profile inversion result);

[0099] The above embodiment of the present application first obtains 60 layers of data every 100m in the vertical range of 0-6km through ground-based lidar (optical parameters), meteorological monitoring stations (ERA5 data set) and ground monitoring stations (PM2.5 component concentration), covering aerosol characteristics, meteorological conditions and near-ground component concentrations, providing multi-dimensional input for the model; then, by using 60 vertical height layers as nodes, a spatial topological network is constructed to explicitly express the correlation between different height layers (such as the vertical diffusion of aerosols is affected by the meteorological conditions of the upper and lower layers), thereby enhancing the model's ability to characterize physical processes; further, by utilizing sample data from different geographical environments (such as cities, suburbs, and different climatic zones), the model can be used to generate a spatial topological network with a topological network. ), learn common cross-regional features (such as meteorological-component correlation patterns), improve model generalization, and quickly adapt to new regions with a small amount of labeled data through meta-learning processing, solving the performance degradation problem of traditional methods caused by scarcity of regional data and reducing dependence on a large amount of labeled data in a single region; finally, the real-time pre-processed data of the current region (optical parameters, meteorological, near-ground components) are input into the model, and the features of nodes at different height layers (such as the spatial correlation between aerosol optical properties and wind speed) are aggregated through the graph neural network (GNN), combined with meta-learning to dynamically adjust the model parameters, and finally the 60-layer PM2.5 chemical component concentration (such as sulfate, nitrate, etc.) is output through the fully connected layer mapping.

[0100] Specifically, in step S30, based on the edge weight Combining sample multi-source data collected from multiple regions to build and train a target adaptive meta-learning model includes the following steps:

[0101] Step S31: constructing and training an initial adaptive meta-learning model by using sample multi-source data collected from multiple regions;

[0102] It should be noted that the above-mentioned multiple regions refer to different geographical areas (such as cities and plateaus);

[0103] Step S32: Based on the edge weight Updating the initial adaptive meta-learning model to obtain a first adaptive meta-learning model;

[0104] Step S33: Based on the multi-source data of the current region collected for n consecutive time periods, the parameters of the first adaptive meta-learning model are updated through the Adam optimizer to obtain the target adaptive meta-learning model.

[0105] It should be noted that the above-mentioned embodiment of the present application jointly trains the model by utilizing heterogeneous data (different climates, pollution sources, and terrains) from multiple regions; further, through the meta-learning framework, the model is equipped with "meta-knowledge" that can quickly adapt to new tasks, laying the foundation for subsequent dynamic optimization and reducing dependence on labeled data from a single region; further, by introducing the edge weights calculated in step S20 (reflecting the similarity of wind speeds between different altitude layers), the physical correlation of vertical altitude layers (such as turbulent mixing and horizontal transport) is encoded into the model, constraining the model learning to conform to the spatial propagation laws of atmospheric dynamics, and correcting the model parameters (such as the strength of GNN inter-layer message transmission) through edge weights to improve the vertical coherence and physical rationality of the inversion results; further, by utilizing real-time multi-source data for n consecutive time periods (such as 72 hours) in the current region (that is, using the multi-source data of the current region as labeled data to fine-tune the parameters of the first adaptive meta-learning model), the first model is lightweight fine-tuned through the Adam optimizer, so that the model can quickly adapt to the local unique environment (such as local pollution source emission characteristics and boundary layer structure differences caused by terrain).

[0106] Specifically, in step S31, an initial adaptive meta-learning model is constructed and trained by using multi-source sample data collected from multiple regions, including the following steps:

[0107] Step S311: Collecting and obtaining sample multi-source data from multiple regions, wherein the sample multi-source data includes multi-source data from each region monitored continuously for one year;

[0108] Step S312: pre-processing the sample multi-source data of each region and splitting them into a support set and a query set;

[0109] Step S313: Constructing a first initial adaptive meta-learning model; using the support set and query set corresponding to each region, respectively training the first initial adaptive meta-learning model and calculating the loss function; when the loss function is less than or equal to the minimum loss function threshold, the model converges, and initial parameters θ of the first initial adaptive meta-learning model corresponding to each region are obtained;

[0110] The loss function is calculated as follows:

[0111] Where, is the importance weight of each component concentration (e.g. black carbon has a higher weight); is the predicted sample local component concentration data of the i-th sample multi-source data; is the sample local component concentration data of the i-th sample multi-source data;

[0112] Step S314: Based on the initial parameters θ, the initial parameters θ of the first initial adaptive meta-learning model are updated through double-layer optimization to obtain the target model parameters ;

[0113] The target model parameters are calculated as follows:

[0114] ;in, ; are the initial parameters of the first initial adaptive meta-learning model, is the inner learning rate, is the loss gradient, where represents the training tasks for each region (the training tasks are the training tasks for the first initial adaptive meta-learning model using each region), For the first initial adaptive meta-learning model to target the task The loss dependence ratio.

[0115] It should be noted that the above-mentioned embodiment of the present application collects multi-source data (ground-based lidar, meteorology, and ground component monitoring) from multiple regions for one year to ensure that the data covers different seasons, climate conditions (such as winter inversion and summer convection), and pollution events (such as sandstorms and haze), thereby enhancing the model's adaptability to spatiotemporal heterogeneity. The periodic characteristics of the meteorological-pollution coupling relationship (such as diurnal boundary layer changes and monsoon transmission effects) are captured through the data throughout the year, providing the model with stable physical model prior knowledge. Then, the data from each region are uniformly preprocessed (such as Hampel filtering and meteorological interpolation). Eliminate instrument errors or data format differences between regions to ensure the compatibility of cross-regional training; by splitting the support set and query set, construct multiple "pseudo-tasks" (one task for each region) to adapt to the training requirements of meta-learning frameworks (such as MAML); through iterative training of support sets and query sets in multiple regions, calculate the loss function, and gradually optimize the initial parameters θ, so that the model has preliminary cross-regional generalization capabilities; finally, through two-layer optimization, the model can not only retain common laws across regions (such as the universal dynamic mechanism of vertical diffusion), but also quickly adapt to the local characteristics of new regions (such as the emission intensity of local pollution sources).

[0116] Specifically, in step S32, based on the edge weight Updating the initial adaptive meta-learning model to obtain a first adaptive meta-learning model includes the following steps:

[0117] Step S321: According to the edge weights of the graph structure Get the k-th layer edge weight of the k-th vertical height layer ;

[0118] Step S322: Based on the k-th layer edge weight The neighbor node set of the i-th vertical height layer is iteratively updated and calculated by the GraphSAGE algorithm to obtain the embedding value of the i-th node in the k-th vertical height layer;

[0119] The node embedding value is calculated as follows:

[0120] Where, is the edge weight of the kth layer; is the set of neighbor nodes, is the embedding of node i in the k-1th layer, Indicates the jth node from the set of neighbor nodes; is an aggregate function, is the connection function, is a nonlinear activation function;

[0121] Step S323: updating the initial adaptive meta-learning model based on the node embedding value to obtain a first adaptive meta-learning model.

[0122] It should be noted that the above-mentioned embodiment of the present application extracts the edge weight corresponding to the k-th vertical height layer by using the edge weight calculated based on step S20 (reflecting the wind speed similarity between different height layers), and clarifies the interaction strength between this layer and the neighboring layer (such as the strong mixing effect between high wind speed layers and the relative independence between low wind speed layers); through the GraphSAGE algorithm, the neighbor set of the i-th node is weightedly aggregated using the edge weight of the k-th layer, and the characteristics of the neighboring layers (such as the meteorological conditions of the upper layer and the aerosol concentration of the lower layer) are integrated to simulate the physical interaction process in the vertical direction (such as the vertical diffusion of pollutants and the upward and downward transmission of meteorological elements); finally, the updated node embedding value (including vertical interaction information) is input into the initial adaptive meta-learning model, and the parameters of its graph neural network (GNN) part are adjusted to make the model more in line with the actual atmospheric dynamic process; through the embedding update guided by the edge weight, the model is constrained to learn the spatial propagation pattern that conforms to the laws of meteorological dynamics (such as pollutants diffuse upward with rising air currents and are suppressed by downdrafts and retained near the ground).

[0123] Based on the above embodiments of the present application, researchers conducted field tests:

[0124] By collecting PM2.5 chemical composition data from five typical regions in China (Beijing, Chengdu, Lhasa, Guangzhou, and Urumqi) from 2021 to 2023 as sample multi-source data;

[0125] We also used traditional models as comparison baselines: CNN-ATT-BiLSTM model and traditional WRF-Chem model;

[0126] By calculating the accuracy: CORR (correlation coefficient), RMSE (μg / m³) as the analysis results;

[0127] As shown in Table 1, the technical solution adopted in the embodiment of the present application is significantly better than the effect of the baseline technical solution;

[0128] Table 1

[0129] area Components CORR (this invention) CORR (baseline) RMSE (present invention) RMSE (baseline) Lhasa <![CDATA[SO4²⁻]]> 0.88 0.72 1.5 3.2 Chengdu <![CDATA[NO3⁻]]> 0.85 0.68 2.0 4.1 Guangzhou BC 0.91 0.75 0.4 1.3

[0130] like Figure 3 As shown, in terms of migration efficiency, compared with the baseline technical solution, the technical solution adopted in this application only requires 5% labeled data (36 hours) in the Urumqi area, and after fine-tuning for 1 hour, the CORR reaches 0.82, an improvement of 30% over the baseline.

[0131] like Figure 4 As shown in the figure, compared with the baseline technical solution, the computing resource consumption of the technical solution adopted in this application is: GPU memory usage is reduced by 45% during the training phase, and the inference speed is improved to real-time (<1 second / time).

[0132] In summary, the technical solution adopted in the above-mentioned embodiments of the present application captures the diffusion and transmission laws of vertical height layers through a dynamic graph structure, and the physical consistency of the inversion results is improved by 25%; the meta-learning framework reduces the model's training data requirements in data-scarce areas by 70%, and migration to new locations only requires 1 hour of fine-tuning; the local aggregation strategy of GraphSAGE reduces computational complexity, speeding up by 40% compared to traditional GCN.

[0133] In summary, the present invention proposes a system and method for inverting the vertical profile of PM2.5 chemical components based on graph neural networks and adaptive meta-learning. First, 60 layers of data are obtained every 100 meters in the vertical range of 0-6 km through ground-based lidar, meteorological monitoring stations and ground monitoring stations, covering aerosol characteristics, meteorological conditions and near-ground component concentrations, providing multi-dimensional input for the model. Then, a spatial topological network is constructed by using 60 vertical height layers as nodes to explicitly express the correlation between different height layers and enhance the model's ability to characterize physical processes. Furthermore, by utilizing sample data from different geographical environments, common cross-regional features are learned to improve the generalization of the model. Meta-learning processing can quickly adapt to new regions with a small amount of labeled data, solving the performance degradation problem of traditional methods caused by scarcity of regional data and reducing dependence on a large amount of labeled data in a single region. Finally, the real-time pre-processed data of the current region is input into the model, and the features of nodes at different height layers are aggregated through the graph neural network. The model parameters are dynamically adjusted in combination with meta-learning, and finally the PM2.5 chemical component concentrations of 60 layers are output through full-connection layer mapping.

[0134] During the specific operation, sample data from various geographical regions are used to train the model with each region as a meta-task. The model parameters are comprehensively trained and optimized. Heterogeneous graph neural networks are used to capture the diffusion laws of components, and MAML optimization is used to reduce the demand for labeled data, solving problems such as insufficient generalization ability of traditional models and dependence on large amounts of data.

[0135] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them. A person skilled in the art may modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A PM2.5 chemical component vertical profile inversion system based on graph neural network and adaptive meta-learning, characterized by: It includes a multi-source data acquisition module, a graph structure construction module, an adaptive meta-learning module, and an inversion output module; The multi-source data acquisition module is used to collect and acquire multi-source data on k vertical height layers of the current geographical area; pre-process the multi-source data to obtain pre-processed multi-source data; The graph structure construction module is used to construct a graph structure by taking each of the vertical height layers as a node; and calculate the edge weight between the i-th node and the j-th node in the graph structure through the meteorological variable data collected during a preset time period. ; The adaptive meta-learning module is used to Combining multi-source data of samples collected from multiple regions to build and train a target adaptive meta-learning model; The inversion output module is used to process the preprocessed multi-source data of the current area at the current moment through the target adaptive meta-learning model, and output the predicted PM2.5 chemical component vertical profile inversion results.

2. A PM2.5 chemical component vertical profile inversion system based on graph neural network and adaptive meta-learning according to claim 1, characterized in that: The multi-source data includes ground-based lidar optical parameters, meteorological variable data and local component concentration data.

3. A PM2.5 chemical component vertical profile inversion system based on graph neural network and adaptive meta-learning according to claim 2, characterized in that: The optical parameters of the ground-based lidar include vertical profile data of 532nm aerosol backscattering coefficient, extinction coefficient and depolarization ratio; The meteorological variable data include the integrated ERA5 reanalysis dataset; The local component concentration data includes real-time monitoring values of near-ground PM2.5 chemical component concentrations.

4. The PM2.5 chemical component vertical profile inversion system based on graph neural network and adaptive meta-learning according to claim 3 is characterized in that: The edge weight is calculated as follows: (1)； in, is the wind speed eigenvector of the i-th node, is the wind speed eigenvector of the jth node, is the bandwidth parameter, which controls the weight decay speed.

5. The PM2.5 chemical component vertical profile inversion system based on graph neural network and adaptive meta-learning according to claim 4 is characterized in that: The adaptive meta-learning module is specifically used to construct and train an initial adaptive meta-learning model using sample multi-source data collected from multiple regions; Based on the edge weight Updating the initial adaptive meta-learning model to obtain a first adaptive meta-learning model; Based on the multi-source data of the current region collected over n consecutive time periods, the parameters of the first adaptive meta-learning model are updated through the Adam optimizer to obtain a target adaptive meta-learning model.

6. The PM2.5 chemical component vertical profile inversion system based on graph neural network and adaptive meta-learning according to claim 5 is characterized in that: The adaptive meta-learning module is further used to collect and obtain sample multi-source data from multiple regions, wherein the sample multi-source data includes multi-source data from various regions monitored continuously for one year; Preprocessing the sample multi-source data of each region and splitting them into support sets and query sets; Constructing a first initial adaptive meta-learning model; using the support set and query set corresponding to each region to train the first initial adaptive meta-learning model and calculate the loss function, the model converges when the loss function is less than or equal to the minimum loss function threshold, and obtaining the initial parameters θ of the first initial adaptive meta-learning model corresponding to each region; Based on the initial parameters θ, the initial parameters θ of the first initial adaptive meta-learning model are updated through two-layer optimization to obtain the target model parameters .

7. The PM2.5 chemical component vertical profile inversion system based on graph neural network and adaptive meta-learning according to claim 6 is characterized in that: The loss function is calculated as follows: ； Where, is the importance weight of each component concentration; is the predicted sample local component concentration data of the i-th sample multi-source data; is the sample local component concentration data of the i-th sample multi-source data; The target model parameters are calculated as follows: ； ; are the initial parameters of the first initial adaptive meta-learning model, is the inner learning rate, is the loss gradient, where Represents the training missions for each region, For the first initial adaptive meta-learning model to target the task The loss dependence ratio.

8. The PM2.5 chemical component vertical profile inversion system based on graph neural network and adaptive meta-learning according to claim 7 is characterized in that: The adaptive meta-learning module is also used in the specific implementation process to calculate the edge weights of the graph structure. Get the k-th layer edge weight of the k-th vertical height layer ; Based on the k-th layer edge weight The neighbor node set of the i-th vertical height layer is iteratively updated and calculated by the GraphSAGE algorithm to obtain the embedding value of the i-th node in the k-th vertical height layer; The initial adaptive meta-learning model is updated based on the node embedding value to obtain a first adaptive meta-learning model.

9. The PM2.5 chemical component vertical profile inversion system based on graph neural network and adaptive meta-learning according to claim 8, characterized in that: The node embedding value is calculated as follows: ； Where, is the edge weight of the kth layer; is the set of neighbor nodes; is the embedding of node i in the k-1th layer, Indicates the jth node from the set of neighbor nodes; is an aggregate function, is the connection function, is a non-linear activation function.

10. A method for inverting vertical profiles of PM2.5 chemical components based on graph neural networks and adaptive meta-learning, characterized in that: The steps are as follows: Collecting and obtaining multi-source data on k vertical height layers of the current geographical area; preprocessing the multi-source data to obtain preprocessed multi-source data; Each of the vertical height layers is used as a node to construct a graph structure; The edge weight between the i-th node and the j-th node in the graph structure is calculated by using the meteorological variable data collected during a preset time period. ; Based on the edge weight Combining multi-source data of samples collected from multiple regions to build and train a target adaptive meta-learning model; The pre-processed multi-source data of the current area at the current moment is processed by the target adaptive meta-learning model, and the predicted PM2.5 chemical component vertical profile inversion result is output.

Images