Atmosphere monitoring point optimization identification method, device and equipment and storage medium

Abstract

The application provides an atmospheric monitoring point distribution optimization identification method, device and equipment and a storage medium. It relates to the technical field of environmental monitoring. The method comprises the following steps: calculating Voronoi area and neighborhood density as node features based on the latitude and longitude of the site, combining with the AQI correlation to construct a graph structure; training a PollutionGNN model by using a double-layer GAT network with residual connection, establishing a three-objective optimization model of site number, coverage area and prediction error, and realizing the Pareto optimal solution set of monitoring efficiency by using an NSGA-III algorithm. The application solves the problem that the traditional point distribution method is difficult to balance the cost, coverage and precision, provides a theoretical basis for cross-regional application through the influence mechanism of geographical environment and pollution source structure on algorithm performance, finally realizes the upgrading of the monitoring network from experience layout to intelligent optimization, and provides an efficient and reusable solution for different cities.

Description

Technical Field

[0001] This application relates to the field of environmental monitoring technology, and in particular to a method, apparatus, equipment and storage medium for optimizing the identification of atmospheric monitoring sites. Background Technology

[0002] Traditional atmospheric monitoring station layouts often rely on expert judgment or single-objective optimization methods. Such strategies are inadequate when dealing with the nonlinear diffusion characteristics of air pollution and the differences between regions. Air pollution is influenced by a variety of complex factors, including meteorological conditions, topography, and the distribution of pollution sources. The interactions between these factors result in highly nonlinear and spatiotemporally dynamic pollutant diffusion. Therefore, traditional monitoring station layouts have significant limitations in capturing these complex interactions, particularly in spatiotemporal dynamic fusion and multi-objective collaborative optimization.

[0003] Furthermore, traditional atmospheric monitoring station layouts struggle to find a balance between monitoring costs, coverage, and prediction accuracy. Increasing the number of monitoring stations often necessitates improving coverage and accuracy, significantly raising costs. Simultaneously, because individual models often fail to fully consider the specific characteristics of different regions during construction, they lack versatility and cannot be directly applied to atmospheric monitoring in other areas. These limitations present numerous challenges to the practical application of traditional atmospheric monitoring station layouts, making it difficult to meet the demands of modern atmospheric environmental monitoring. Therefore, it is necessary to develop more scientific and rational methods for atmospheric monitoring station layout to adapt to the complex and ever-changing atmospheric pollution conditions. Summary of the Invention

[0004] This application provides a method, device, equipment, and storage medium for optimizing and identifying atmospheric monitoring site locations. It can improve efficiency, increase prediction accuracy, better capture spatial relationships between sites and pollutant diffusion characteristics, and enhance the model's versatility for different scenarios.

[0005] Firstly, this application provides a method for optimizing and identifying atmospheric monitoring site locations, including:

[0006] Obtain the node characteristics of the monitoring stations; wherein, the node characteristics of the monitoring stations include normalized latitude and longitude, Voronoi area and spatial neighborhood density;

[0007] Construct edge features; wherein the edge features include normalized inter-site distances and AQI correlation coefficients;

[0008] Graph structure data is constructed based on node features, edge features, and the connection relationships between sites.

[0009] A pollution propagation model is constructed and trained using the graph structure data. The pollution propagation model employs a two-layer graph attention network with residual connections, and fuses node features and edge features through a dynamic masking mechanism to model the pollution propagation process, outputting the feature vector prediction value of each monitoring station.

[0010] A three-objective constrained optimization model is established with the three objectives of minimizing the number of monitoring stations, maximizing the coverage area, and minimizing the prediction error. The Pareto optimal solution set is obtained by solving the three-objective constrained optimization model using the NSGA-III algorithm. Based on the Pareto optimal solution set, an optimized layout scheme for monitoring stations is output. The prediction error is the prediction error of the training pollution propagation model.

[0011] In one possible design, the connections between sites are determined as follows:

[0012] Set distance thresholds and correlation coefficient thresholds;

[0013] Using the Pearson correlation coefficient as the AQI correlation coefficient, all station pairs are traversed. When the distance between the stations is less than the distance threshold and the absolute value of the AQI correlation coefficient is greater than the correlation coefficient threshold, it is determined that there is a valid connection between the two stations.

[0014] For sites whose number of candidate connections is less than a set number after filtering, neighboring sites of the site are added as connecting sites to ensure that the number of connecting sites of the site is not less than the set number.

[0015] In one possible design, the pollution propagation model is based on the PollutionGNN model of a graph attention network, comprising two attention layers and residual connections; wherein, in the attention layers, the attention coefficients between stations are calculated using the following formula:

[0016]

[0017] In the formula, N i Let a be the set of neighbors of site i. T Let α be the attention weight vector. ij h is the normalized attention coefficient of site i to site j. i For the characteristics of site i, h j For the characteristics of site j, h k Let be the features of neighboring site k, exp be the exponential function, and LeakyReLU be the activation function.

[0018] In one possible design, residual connections are introduced into the two attention layers for quality control. Shallow features are added to deep features using the following formula to form an identity mapping path:

[0019] hl =f(h l-1 ,E)+h l-1 ,

[0020] In the formula, f is the feature transformation function of the current layer, E is the edge feature matrix, and h l h is the output feature of the current layer. l-1 This is the output feature of the previous layer.

[0021] In one possible design, when training the pollution propagation model using the graph structure data, some sites are randomly masked using a dynamic mask; wherein the dynamic mask is represented as:

[0022] mask r atio = 0.3 + 0.4 × (1 - e) -epoch / 100 )

[0023] In the formula, mask r atio represents the site mask ratio, e represents the natural constant, and epoch represents the training epoch.

[0024] In one possible design, the three-objective constrained optimization model is expressed as:

[0025] minF(x)=[f num (x),f area (x),f error (x)] T

[0026]

[0027] x i ∈{0,1}

[0028] In the formula, f num (x) represents the number of stations, f area (x) represents the coverage area, f error (x) represents the prediction error, T is the matrix transpose, st is the constraint condition, and x i Let N be the number of measurement points, minF(x) be the minimum of the three objective functions, and N be the number of stations calculated. i ∈{0,1} indicates whether to select the i-th station;

[0029] The number of stations is determined using the following formula:

[0030]

[0031] The coverage area is determined using the following formula:

[0032]

[0033] In the formula, A i Let i be the Voronoi area;

[0034] The prediction error is determined using the following formula:

[0035]

[0036] In the formula, ∈ max To the maximum allowable prediction error, y is the predicted value of the pollution transmission model, n is the total number of samples, and MSE is the mean squared error.

[0037] In one possible design, the Pareto optimal solution set is obtained by solving the three-objective constrained optimization model using the NSGA-III algorithm, including:

[0038] Population initialization: Randomly generate an initial population P0 and calculate the objective function value for each individual.

[0039] Reference point generation: Generate a set of reference points based on the number and distribution of the objective function;

[0040] Weight vector generation: Generate a set of adaptive weight vectors based on the reference point;

[0041] Non-dominated ranking and association: Perform non-dominated ranking on the population and associate each individual with the nearest reference point;

[0042] Selection, crossover, and mutation: Generate offspring population Q0 using selection, crossover, and mutation operators;

[0043] Population update: The initial population P0 and the offspring population Q0 are merged to form a new population R0. Then, R0 is sorted and associated in a non-dominated manner to select the next generation parent population P1.

[0044] Secondly, this application provides an atmospheric monitoring site optimization identification device, the device comprising:

[0045] The node feature acquisition module is configured to acquire the node features of the monitoring station; wherein, the node features of the monitoring station include normalized latitude and longitude, Voronoi area and spatial neighborhood density.

[0046] An edge feature construction module is configured to construct edge features; wherein the edge features include normalized inter-site distances and AQI correlation coefficients;

[0047] The graph structure construction module is configured to construct graph structure data based on node features, edge features, and the connection relationships between sites.

[0048] The model training module is configured to construct a pollution propagation model and train the pollution propagation model using the graph structure data; wherein, the pollution propagation model adopts a two-layer graph attention network with residual connections, and fuses node features and edge features through a dynamic masking mechanism to model the pollution propagation process, and outputs the feature vector prediction value of each monitoring station.

[0049] The optimization solution module is configured to establish a three-objective constrained optimization model with the three objectives of minimizing the number of stations, maximizing the coverage area, and minimizing the prediction error. The module then uses the NSGA-III algorithm to solve the three-objective constrained optimization model to obtain a Pareto optimal solution set. Based on the Pareto optimal solution set, the module outputs an optimized layout scheme for monitoring stations. The prediction error is the prediction error of the training pollution propagation model.

[0050] Thirdly, embodiments of this application provide an electronic device, including: at least one processor and a memory; the memory stores computer execution instructions; the at least one processor executes the computer execution instructions stored in the memory, causing the at least one processor to execute the atmospheric monitoring site optimization and identification method as described in the first aspect and various possible designs of the first aspect.

[0051] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions. When a processor executes the computer-executable instructions, it implements the atmospheric monitoring site optimization and identification method described in the first aspect and various possible designs of the first aspect.

[0052] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the atmospheric monitoring site optimization and identification method described in the first aspect and various possible designs of the first aspect.

[0053] The atmospheric monitoring site optimization identification method, apparatus, equipment, and storage medium provided in this application have at least the following beneficial effects:

[0054] This application, through a scientifically sound design and optimization scheme, can significantly reduce the overall cost of the monitoring process. Specifically, the number of monitoring stations is significantly reduced after optimization, which not only simplifies the complexity of the monitoring system but also greatly reduces the burden of construction and maintenance. More importantly, while minimizing the number of stations, it still ensures a sufficiently wide monitoring range, covering key areas and thus not affecting the monitoring results. At the same time, due to the reduction in the number of stations, the corresponding maintenance and upkeep costs also decrease, saving substantial expenses for long-term operation.

[0055] In terms of accuracy, this application also achieves significant improvements. By optimizing the site layout, retaining key sites, and utilizing these sites for data collection and analysis, the conditions in uncovered areas can be predicted more accurately. Specifically, in terms of prediction error (MSE), compared to traditional methods, this application exhibits smaller prediction errors and more accurate results, thereby greatly improving the reliability and practicality of the monitoring system. Attached Figure Description

[0056] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0057] Figure 1 A flowchart illustrating an atmospheric monitoring site optimization and identification method provided in this application embodiment;

[0058] Figure 2 A schematic diagram of the graph structure data provided in the embodiments of this application;

[0059] Figure 3 A flowchart illustrating the solution process of the three-objective constrained optimization model provided in this application embodiment;

[0060] Figure 4 The performance comparison diagram of the atmospheric monitoring site optimization and identification method provided in the embodiments of this application and the existing methods is shown in the figure; wherein, (a) comparison of the number of stations; (b) comparison of coverage area; (c) comparison of prediction error; (d) comparison of calculation time;

[0061] Figure 5 This is a structural diagram of an atmospheric monitoring site optimization and identification device provided in an embodiment of this application.

[0062] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0063] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0064] The collection, storage, use, processing, transmission, provision, and disclosure of financial data or user data involved in the technical solution of this application all comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0065] It should be noted that in the embodiments of this application, certain software, components, models and other existing solutions in the industry may be mentioned. These should be regarded as exemplary and are only intended to illustrate the feasibility of implementing the technical solution of this application. However, it does not mean that the applicant has used or necessarily used the solution.

[0066] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0067] To accurately identify students' confused expressions in classroom teaching scenarios, this application provides a method for optimizing and identifying atmospheric monitoring site locations, such as... Figure 1 As shown, the method includes the following steps S10-S40.

[0068] S10: Obtain the node characteristics of the monitoring stations; where the node characteristics of the monitoring stations include normalized latitude and longitude, Voronoi area and spatial neighborhood density.

[0069] In this embodiment, the node characteristics of the monitoring station can be obtained through the following steps S101-S102.

[0070] S101: Data Acquisition.

[0071] Acquire AQI, PM2.5, and other pollutant data, along with latitude and longitude information, from monitoring stations. Each dataset records daily monitoring data from each station, along with the station's latitude and longitude information for spatial feature analysis.

[0072] First, the CSV file is read by traversal. The wide-format data is converted to long-format data using the melt() method. Pollutant types are automatically labeled and merged into a unified dataset. Finally, it is reshaped into wide-format data indexed by site and date.

[0073] In terms of data quality control, Dropna was used to delete invalid rows that were missing all indicators (accounting for 0.13%), and time linear interpolation was used to fill missing values ​​for individual indicators. After processing, the missing rate of all indicators was less than 0.1%. Spatiotemporal features were further constructed.

[0074] S102: Feature processing.

[0075] To adapt the data for Graph Neural Network (GNN) modeling and analysis, and to construct the graph data structure, this embodiment selects latitude and longitude, Voronoi area, and proximity density as node features. Additionally, it selects distance between stations and AQI coefficient as edge features.

[0076] To avoid scale distortion of latitude and longitude in subsequent area calculations, this embodiment uses UTM projection (Zone 50N) to convert WGS84 latitude and longitude coordinates into planar coordinates: (x,y) = UTM(lon,lat). This conversion preserves the shape and area accuracy of the local region and avoids calculation errors caused by differences in latitude and longitude scales.

[0077] The space of each monitoring station is quantified using a Voronoi diagram, which divides the space into multiple polygonal regions, such that the distance from any point within a region to its own station is less than the distance to other stations.

[0078] Next, the Voronoi area of ​​each station is calculated, and a Voronoi diagram is generated using the Fortune algorithm for each original station p. i The area A of its Voronoi polygon is calculated using Green's formula:

[0079]

[0080] In the formula, (x i ,y i () represents the coordinates of the polygon vertices, arranged in counter-clockwise order. The calculation results are converted from square meters to square kilometers.

[0081] The calculated Voronoi area A i Merge into the site data frame, and fill a small number of outliers with the mean of the same area:

[0082]

[0083] In the formula, N(i) represents the set of sites that are geographically adjacent to site i.

[0084] In addition, it is necessary to calculate the spatial neighborhood density, convert the latitude and longitude coordinates of the stations to radians, and use the Haversine formula to calculate the spherical distance:

[0085]

[0086] In the formula, R is the Earth's radius (6371 km), φ1 and φ2 represent the latitudes of the two different points being measured, and Δλ represents the difference in longitude between the two points being measured.

[0087] Then, a spatial index is built using the BallTree data structure, which is suitable for nearest neighbor search in high-dimensional space with a time complexity of O(log n). For each station, the number of neighboring stations within a 10-kilometer radius is queried.

[0088] This neighborhood density feature can be used to identify high-density areas to avoid resource waste, and can also be used to discover low-density areas to guide the layout of new sites. It can also be used for multi-objective optimization as a constraint on network optimization to balance coverage and monitoring costs.

[0089] To eliminate the impact of differences in the dimensions of different features on the model, the Z-score normalization method is used to normalize each feature:

[0090]

[0091] In the formula, μ is the mean of the feature, σ is the standard deviation of the feature, and z i These are the original node features. These are the normalized node features.

[0092] After standardization, the mean of each feature is 0 and the standard deviation is 1.

[0093] They have the same scale. Finally, a node feature matrix is ​​constructed containing four features: 'Voronoi area', 'neighbor density', 'longitude', and 'latitude'.

[0094] S20: Construct edge features; where edge features include normalized inter-site distances and AQI correlation coefficients.

[0095] In some embodiments, the distance between stations is calculated using the Haversine formula.

[0096] In some embodiments, the Pearson correlation coefficient is used to construct an AQI correlation feature between stations to describe the spatial coordination of air quality changes, and this feature is set as the edge weight index of the graph neural network (GNN). The specific processing flow is as follows:

[0097] First, the raw monitoring data is converted into a wide-format matrix, with date as the index, site code as the column, and daily average AQI as the statistical value. The data is then structured using pivot table operations to form a matrix of dimension D×N (D is the time span, N is the total number of sites), with each row corresponding to one day and each column corresponding to the daily average AQI value of one site.

[0098] The Pearson correlation coefficient was used to measure the linear correlation of AQI between stations. A minimum effective data point threshold of 30 days was set for calculation; that is, the correlation coefficient was only calculated when there were at least 30 days of synchronized effective data between two stations. Correlation values ​​below this threshold were assigned a value of 0 to address data sparsity. The final result was a D×N correlation coefficient matrix, with matrix elements ranging from -1 to 1. Larger absolute values ​​indicate a stronger co-current trend in AQI between the two stations (positive values ​​indicate co-current trends, negative values ​​indicate counter-current trends).

[0099] S30: Construct graph structure data based on node features, edge features, and connections between sites.

[0100] In some embodiments, a hybrid edge generation strategy is proposed to address the limitations of traditional single connection criteria. The connection condition is d. ij ≤50km (regional pollution transport threshold), |ρ AQI If (i,j)|≥0.6 (significant correlation level), iterate through all site pairs. Only when the distance between sites is less than the distance threshold and the absolute value of the AQI correlation coefficient is greater than the correlation coefficient threshold is a valid connection considered to exist between the two sites. To avoid the number of edges being too small due to threshold filtering, for sites with less than K=5 candidate connections after filtering, their K nearest neighbors are added as connections.

[0101] Edge feature standardization and enhancement. After storing the selected edges and their corresponding features, the edge features are standardized by using StandardScaler to normalize the distance and correlation coefficient, and using a Gaussian kernel function to transform the distance, taking the absolute value of the correlation coefficient.

[0102] The processed node features, edge connectivity information, and edge features are used to construct a graph data structure according to the format requirements of the PyTorchGeometric library. The constructed graph data structure is as follows: Figure 2 As shown.

[0103] S40: Construct a pollution propagation model and train it using graph-structured data. The pollution propagation model employs a two-layer graph attention network with residual connections, and uses a dynamic masking mechanism to fuse node features and edge features to model the pollution propagation process, outputting the predicted feature vector value for each monitoring station.

[0104] In some embodiments, a PollutionGNN model based on graph attention network is constructed as a pollution propagation model, which models the pollution propagation process by fusing node and edge features.

[0105] The core of graph attention networks is the introduction of an attention mechanism, allowing sites to dynamically adjust connection weights based on the importance of their neighbors. This is achieved by calculating the attention coefficients between nodes:

[0106]

[0107] In the formula, N i Let α be the set of neighbors of station i. ij h is the normalized attention coefficient of site i to site j. i For the characteristics of site i, h j For the characteristics of site j, h k Let be the features of neighboring site k, exp be the exponential function, LeakyReLU be the activation function, and a be the feature of neighboring site k. T The attention weight vector assigns different weights to different neighbors, thereby capturing the dependencies between nodes more flexibly and improving the model's attention to key neighbors.

[0108] A multi-head attention and hierarchical feature extraction module is adopted: the first layer uses 4-head attention (heads=4) and the second layer uses 2-head attention. By computing multiple attention distributions in parallel, the model can capture complex relationships between sites from different subspaces. The final features retain the independent information of each head in a concatenation (non-average) manner. The output dimension of the first layer is 32×4=128 and the second layer is 64×2=128 to enhance the model's ability to express spatial heterogeneity. Batch normalization layers (bn1, bn2) are added after each convolution.

[0109] Using residual connections to mitigate gradient vanishing: A residual connection (x = x2 + x1) is introduced between two GAT layers to form an identity mapping path: h l =f(h l-1 ,E)+h l-1 , where f() is the feature transformation function of the current layer, and E is the edge feature matrix.

[0110] Dynamic node masking enhances generalization: During training, some nodes are randomly masked using dynamic masking to simulate missing data in real-world scenarios, enhancing the model's robustness to incomplete data. Simultaneously, a random deactivation mechanism suppresses overfitting. The dynamic mask is represented as follows:

[0111] mask r atio = 0.3 + 0.4 × (1 - e) -epoch / 100 )

[0112] In the formula, mask r atio represents the site mask ratio, e represents the natural constant, and epoch represents the training epoch.

[0113] During training, the KMeans clustering algorithm was used to spatially stratify the latitude and longitude of the stations. Based on the latitude and longitude coordinates (LONGITUDE, LATITUDE), N stations were divided into M geographical clusters (n_clusters = M). Stratified sampling (stratify = cluster_labels) was used to divide the training set (80%) and the validation set (20%).

[0114] Gradient clipping is used to limit the gradient norm (clip_grad_norm_, threshold 1.0) to prevent gradient explosion during backpropagation. Early stopping is employed by adding weight decay (weight_decay = 1e) to the Adam optimizer. -4 To suppress model parameter overfitting, a patience value (patience = 10) and a validation loss floating threshold of 0.1% (best_loss × 0.999) are set to avoid premature termination of training due to small fluctuations, thus balancing model accuracy and training efficiency.

[0115] In some embodiments, the data processing flow of the pollution propagation model includes the following steps S401-S403.

[0116] S401: Obtain the input data for the model.

[0117] The model's input data includes:

[0118] Node feature h i Attributes of a single site, including normalized latitude and longitude values, Voronoi area, pollution concentration, etc., with the dimension denoted as d. in The features of all nodes constitute matrix H.

[0119] Edge feature E: Inter-site association attributes, including distance normalization value, AQI correlation coefficient, etc., which are stored in matrix form for edge information of site pairs.

[0120] Neighbor set N i The neighbor site index of site i can be filtered through a hybrid edge generation strategy to reflect effective connections of spatial / contamination associations.

[0121] Attention weight vector a T Learnable parameters, 2D in dimension in Because the calculation requires concatenating the features h of two nodes. i and h j Therefore, the dimension is doubled, which is used to assign dynamic weights to different neighbors.

[0122] Step S401 provides input data for subsequent attention mechanisms and feature transformations. Graph structure data is the basic carrier for the operation of graph attention networks.

[0123] S402: Multi-head attention and feature extraction, including:

[0124] S4021: Multi-head parallel computing.

[0125] The first attention layer uses 4 attention heads (4 heads), and the second attention layer uses 2 attention heads. Each head independently performs the attention coefficient calculation and feature weighting aggregation process: for the m-th attention head, the output feature of node i is:

[0126]

[0127] In the formula, W ( m) is the learnable weight matrix for the m-th head, and σ is the activation function, which realizes the feature transfer from d. in Via Dimensional transformation; It is the attention coefficient calculated for the m-th head.

[0128] S4022: Feature Concatenation and Fusion: The first layer outputs 4 features, each with a dimension of 32, resulting in a concatenated dimension of 128; the second layer outputs 2 features, each with a dimension of 64, resulting in a concatenated dimension of 128. Concatenation, rather than averaging, preserves the subspace correlation information captured by different features. For example, some features focus on nearby sites, while others focus on sites with high pollution correlation coefficients, enhancing the model's representation of spatial heterogeneity.

[0129] S4023: Batch Standardization: After multi-head attention calculation at each layer, batch standardization is performed on the features to accelerate training convergence and alleviate internal covariate bias.

[0130] Step S402 uses multi-head parallelism and splicing fusion to allow the model to capture site correlations from multiple subspace dimensions, such as spatial distance and pollution correlation, to solve the problem of the limitation of a single attention head perspective. Batch standardization ensures training stability and lays the foundation for subsequent residual connections and model convergence.

[0131] S403: Residual join operation, formula is:

[0132] h l =f(h l-1 ,E)+h l-1

[0133] In the formula, f is the feature transformation function of the current layer, E is the edge feature matrix, and h l h is the output feature of the current layer. l-1 This represents the output features of the previous layer. `f` includes attention coefficient calculation, multi-head feature aggregation, and batch normalization, outputting the transformed features.

[0134] S404: Output node feature matrix. This node feature matrix integrates the node's own attributes (latitude and longitude, pollution concentration), the dynamically weighted features of neighboring nodes (aggregated through attention coefficients), multi-subspace association information (multi-head attention stitching results), and the gradient stability and feature enhancement effects guaranteed by residual connections.

[0135] The output node features can be directly used for pollution propagation process modeling, such as predicting future pollution concentrations at stations and simulating pollution diffusion paths, or as the core input for multi-objective optimization in step S50. The model can predict the errors of different site deployment schemes, supporting the three-objective decision-making of cost (number of stations), coverage area, and accuracy (prediction error) of atmospheric monitoring site deployment.

[0136] S50: A three-objective constrained optimization model is established with the three objectives of minimizing the number of monitoring stations, maximizing the coverage area, and minimizing the prediction error. The Pareto optimal solution set is obtained by solving the three-objective constrained optimization model using the NSGA-III algorithm. The optimized layout scheme of monitoring stations is output based on the Pareto optimal solution set. The prediction error is the prediction error of the pollution propagation model.

[0137] In some embodiments, to address the conflicting objectives of balancing monitoring cost control, improving spatial coverage efficiency, and ensuring prediction accuracy in the layout of air quality monitoring networks, this invention models the monitoring site optimization problem as a three-objective constrained optimization model and provides the following mathematical definition:

[0138] minF(x)=[f num (x),f area (x),f error (x)] T

[0139]

[0140] x i ∈{0,1}

[0141] In the formula, f num (x) represents the number of stations, f area (x) represents the coverage area, f error (x) represents the prediction error, T is the matrix transpose, st is the constraint condition, and x i Let N be the number of measurement points, minF(x) be the minimum of the three objective functions, and N be the number of stations calculated. i ∈{0,1} indicates whether to select the i-th station;

[0142] Minimize the number of sites:

[0143]

[0144] After normalization, the number of stations f num (x) is mapped to the interval [0,1].

[0145] Maximize coverage:

[0146]

[0147] In the formula, A i Let be the Voronoi area of ​​site i.

[0148] The goal is maximized by minimizing the negative coverage area.

[0149] Minimize prediction error:

[0150]

[0151] In the formula, ∈ max To the maximum allowable prediction error, y is the predicted value of the pollution transmission model, n is the total number of samples, and MSE is the mean squared error.

[0152] The formula for calculating MSE is expressed as follows:

[0153]

[0154] Where y is the true value. is the predicted value, and n is the total number of samples.

[0155] Constraints: Minimum site retention rate: At least 25% of sites must be retained.

[0156] Decision variable: x i ∈{0,1} indicates whether to select the i-th station.

[0157] Geographic features: M is a pre-trained graph neural network model, ∈ max =0.5 is the maximum permissible prediction error

[0158] Finally, the NSGA-III mainstream evolutionary algorithm was used to solve the problem.

[0159] In some embodiments, the algorithm characteristics and parameter settings are as follows: NSGA-III (NSGA with reference direction).

[0160] Its core idea is to introduce a reference direction to guide the population to evolve toward a uniformly distributed Pareto front, which is applicable to high-dimensional targets.

[0161] Parameters: Reference direction is generated using the energy method (3 targets, 90 directions), population size is 100, simulated binary crossover (SBX, prob = 0.8, eta = 15), polynomial mutation (PM, eta = 20), and duplicate individuals are prohibited.

[0162] like Figure 3 As shown, the Pareto optimal solution set is obtained by solving the three-objective constrained optimization model using the NSGA-III algorithm, including:

[0163] S501. Population initialization: Randomly generate an initial population P0 and calculate the objective function value for each individual.

[0164] S502, Reference Point Generation: Generate a set of reference points based on the number and distribution of the objective function;

[0165] S503, Weight Vector Generation: Generate a set of adaptive weight vectors based on the reference point;

[0166] S504, Non-dominated ordering and association: Perform non-dominated ordering on the population and associate each individual with the nearest reference point;

[0167] S505, Selection, Crossover and Mutation: Generate offspring population Q0 by using selection, crossover and mutation operators;

[0168] S506. Population Update: Merge the initial population P0 and the offspring population Q0 to form a new population R0. Then, perform non-dominated sorting and association on R0 to select the next generation parent population P1.

[0169] S507. Repeat S503-S506 until the termination condition is met, and output the Pareto optimal solution set.

[0170] The performance of the atmospheric monitoring site optimization and identification method proposed in this application was compared with other methods under the same implementation environment. The results are as follows: Figure 4 As shown, it can be analyzed from the following four perspectives.

[0171] Number of sites: Before optimization, there were 24 sites. NSGA-II and SMS-EMOA retained 15 sites after optimization, while this algorithm retains 14 sites, all of which are significant reductions, effectively lowering costs. This algorithm is the most aggressive in reducing the number of sites, while NSGA-II and SMS-EMOA are relatively conservative, retaining slightly more sites to balance other metrics.

[0172] Coverage area: The coverage area before optimization was 23428.41km. 2 After optimization, NSGA-II (20721.00km) 2 The algorithm retains the most data, indicating its emphasis on spatial coverage; this algorithm (18988.87km)2 The significant reduction in coverage indicates a greater inclination to reduce the number of sites; SMS-EMOA (20333.19km) 2 Coverage falls somewhere in between, aiming to maintain coverage while reducing the number of sites.

[0173] Prediction Error (MSE): The prediction error before optimization was 0.1825, and it was significantly reduced after optimization. SMS-EMOA (0.0396) and our algorithm (0.0419) had the smallest errors, indicating that these two algorithms are more accurate in predicting uncovered areas using retained sites; NSGA-II (0.0802) had a relatively high error, but it was still better than before optimization.

[0174] Computation time: NSGA-II (30.8 seconds) and this algorithm (28.18 seconds) are suitable for time-sensitive scenarios; SMS-EMOA (183.03 seconds) takes significantly longer due to its high computational complexity of supervolume, but in return it achieves the lowest prediction error.

[0175] Overall conclusion: This algorithm strikes a good balance between reducing the number of sites and prediction accuracy, and its high computational efficiency makes it suitable for cost-sensitive scenarios that prioritize accuracy.

[0176] This application also provides an atmospheric monitoring site optimization and identification device, such as... Figure 5 As shown, the atmospheric monitoring site optimization and identification device includes:

[0177] The node feature acquisition module 501 is configured to acquire the node features of the monitoring station; wherein, the node features of the monitoring station include normalized latitude and longitude, Voronoi area and spatial neighborhood density.

[0178] The edge feature construction module 502 is configured to construct edge features; wherein the edge features include normalized inter-site distances and AQI correlation coefficients;

[0179] The graph structure construction module 503 is configured to construct graph structure data based on node features, edge features, and the connection relationships between sites.

[0180] The model training module 504 is configured to construct a pollution propagation model and train the pollution propagation model using the graph structure data; wherein, the pollution propagation model adopts a two-layer graph attention network with residual connections, and fuses node features and edge features through a dynamic masking mechanism to model the pollution propagation process, and outputs the feature vector prediction value of each monitoring station.

[0181] The optimization solution module 505 is configured to establish a three-objective constrained optimization model with the three objectives of minimizing the number of stations, maximizing the coverage area, and minimizing the prediction error. The module then uses the NSGA-III algorithm to solve the three-objective constrained optimization model to obtain a Pareto optimal solution set. Based on the Pareto optimal solution set, the module outputs an optimized layout scheme for monitoring stations. The prediction error is the prediction error of the training pollution propagation model.

[0182] In some embodiments, the graph structure building module is further configured to determine the connection relationships between sites in the following manner:

[0183] Set distance thresholds and correlation coefficient thresholds;

[0184] Using the Pearson correlation coefficient as the AQI correlation coefficient, all station pairs are traversed. When the distance between the stations is less than the distance threshold and the absolute value of the AQI correlation coefficient is greater than the correlation coefficient threshold, it is determined that there is a valid connection between the two stations.

[0185] For sites whose number of candidate connections is less than a set number after filtering, neighboring sites of the site are added as connecting sites to ensure that the number of connecting sites of the site is not less than the set number.

[0186] In some embodiments, the pollution propagation model is based on the PollutionGNN model of a graph attention network, comprising two attention layers and residual connections; wherein, in the attention layers, the attention coefficients between stations are calculated using the following formula:

[0187]

[0188] In the formula, N i Let a be the set of neighbors of site i. T Let α be the attention weight vector. ij h is the normalized attention coefficient of site i to site j. i For the characteristics of site i, h j For the characteristics of site j, h k Let be the features of neighboring site k, exp be the exponential function, and LeakyReLU be the activation function.

[0189] In some embodiments, the model training module is further configured to introduce residual connections in the two attention layers for quality control, and to add the shallow features and deep features using the following formula to form an identity mapping path:

[0190] h l =f(h) l-1 ,E)+h l-1 ,

[0191] In the formula, f is the feature transformation function of the current layer, E is the edge feature matrix, and h l h is the output feature of the current layer. l-1 This is the output feature of the previous layer.

[0192] In some embodiments, the model training module is further configured to randomly mask some sites using a dynamic mask when training the pollution propagation model using the graph structure data; wherein the dynamic mask is represented as:

[0193] mask r atio = 0.3 + 0.4 × (1 - e) -epoch / 100 )

[0194] In the formula, mask r atio represents the site mask ratio, e represents the natural constant, and epoch represents the training epoch.

[0195] In some embodiments, the three-objective constrained optimization model is expressed as:

[0196] minF(x)=[f num (x),f area (x),f error (x)] T

[0197]

[0198] x i ∈{0,1}

[0199] In the formula, f num (x) represents the number of stations, f area (x) represents the coverage area, f error (x) represents the prediction error, T is the matrix transpose, st is the constraint condition, and x i Let N be the number of measurement points, minF(x) be the minimum of the three objective functions, and N be the number of stations calculated. i ∈{0,1} indicates whether to select the i-th station;

[0200] The number of stations is determined using the following formula:

[0201]

[0202] The coverage area is determined using the following formula:

[0203]

[0204] In the formula, A i Let i be the Voronoi area;

[0205] The prediction error is determined using the following formula:

[0206]

[0207] In the formula, ∈ max To the maximum allowable prediction error, y is the predicted value of the pollution transmission model, n is the total number of samples, and MSE is the mean squared error.

[0208] In some embodiments, the optimization solution module is further configured to:

[0209] Population initialization: Randomly generate an initial population P0 and calculate the objective function value for each individual.

[0210] Reference point generation: Generate a set of reference points based on the number and distribution of the objective function;

[0211] Weight vector generation: Generate a set of adaptive weight vectors based on the reference point;

[0212] Non-dominated ranking and association: Perform non-dominated ranking on the population and associate each individual with the nearest reference point;

[0213] Selection, crossover, and mutation: Generate offspring population Q0 using selection, crossover, and mutation operators;

[0214] Population update: The initial population P0 and the offspring population Q0 are merged to form a new population R0. Then, R0 is sorted and associated in a non-dominated manner to select the next generation parent population P1.

[0215] Repeat the process of generating weight vectors, non-dominated sorting and association, selection, crossover and mutation, and population update until the termination condition is met, and output the Pareto optimal solution set.

[0216] This application provides an electronic device. The electronic device may include a processor and a memory, wherein the processor and the memory can communicate; exemplarily, the processor and the memory communicate via a communication bus.

[0217] The processor executes computer execution instructions stored in memory, causing the processor to perform the scheme in the above embodiments. The processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0218] The communication bus can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The system bus can be divided into address bus, data bus, control bus, etc. Transceivers are used to enable communication between database access devices and other computers (e.g., clients, read-write libraries, and read-only libraries). Memory may include random access memory (RAM) and may also include non-volatile memory.

[0219] The electronic device provided in this application embodiment can be the terminal device described in the above embodiments.

[0220] This application also provides a computer-readable storage medium storing computer instructions. When the computer instructions are executed on a computer, the computer performs the technical solution of the atmospheric monitoring site optimization and identification method described in the above embodiments.

[0221] This application also provides a computer program product, which includes a computer program stored in a computer-readable storage medium. At least one processor can read the computer program from the computer-readable storage medium. When the at least one processor executes the computer program, it can implement the technical solution of the atmospheric monitoring site optimization and identification method in the above embodiments.

[0222] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or modules, and may be electrical, mechanical, or other forms.

[0223] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to implement the solution of this embodiment according to actual needs.

[0224] Furthermore, the functional modules in the various embodiments of this application can be integrated into one processing unit, or each module can exist physically separately, or two or more modules can be integrated into one unit. The unit composed of the above modules can be implemented in hardware or in the form of hardware plus software functional units.

[0225] The integrated modules described above, implemented as software functional modules, can be stored in a computer-readable storage medium. These software functional modules, stored in a storage medium, include several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute some steps of the methods of the various embodiments of this application.

[0226] It should be understood that the aforementioned processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. A general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly manifested as being executed by a hardware processor, or executed by a combination of hardware and software modules within the processor.

[0227] The memory may include high-speed RAM, and may also include non-volatile storage (NVM), such as at least one disk storage device, and may also be a USB flash drive, external hard drive, read-only memory, disk or optical disc, etc.

[0228] Buses can be Industry Standard Architecture (ISA) buses, Peripheral Component Interconnect (PCI) buses, or Extended Industry Standard Architecture (EISA) buses, etc. Buses can be categorized into address buses, data buses, control buses, etc.

[0229] The aforementioned storage medium can be implemented from any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The storage medium can be any available medium accessible to general-purpose or special-purpose computers.

[0230] An exemplary storage medium is coupled to a processor, enabling the processor to read information from and write information to the storage medium. Alternatively, the storage medium can be an integral part of the processor. The processor and storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and storage medium can exist as discrete components in an electronic control unit or main control device.

[0231] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0232] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A method for optimizing and identifying atmospheric monitoring site locations, characterized in that, The method includes: Obtain the node characteristics of the monitoring stations; wherein, the node characteristics of the monitoring stations include normalized latitude and longitude, Voronoi area and spatial neighborhood density; Construct edge features; wherein the edge features include normalized inter-site distances and AQI correlation coefficients; Graph structure data is constructed based on node features, edge features, and the connection relationships between sites. A pollution propagation model is constructed and trained using the graph structure data. The pollution propagation model employs a two-layer graph attention network with residual connections, and fuses node features and edge features through a dynamic masking mechanism to model the pollution propagation process, outputting the feature vector prediction value of each monitoring station. A three-objective constrained optimization model is established with the three objectives of minimizing the number of monitoring stations, maximizing the coverage area, and minimizing the prediction error. The Pareto optimal solution set is obtained by solving the three-objective constrained optimization model using the NSGA-III algorithm. Based on the Pareto optimal solution set, an optimized layout scheme for monitoring stations is output. The prediction error is the prediction error of the trained pollution propagation model. The connection relationships between sites are determined in the following way: Set distance thresholds and correlation coefficient thresholds; Using the Pearson correlation coefficient as the AQI correlation coefficient, all station pairs are traversed. When the distance between the stations is less than the distance threshold and the absolute value of the AQI correlation coefficient is greater than the correlation coefficient threshold, it is determined that there is a valid connection between the two stations. For sites whose number of candidate connections is less than a set number after filtering, neighboring sites of the site are added as connecting sites to ensure that the number of connecting sites of the site is not less than the set number.

2. The atmospheric monitoring site optimization and identification method according to claim 1, characterized in that, The pollution propagation model is based on the PollutionGNN model of graph attention networks, and includes two attention layers and residual connections; wherein, in the attention layers, the attention coefficients between stations are calculated using the following formula: ； In the formula, Let i be the set of neighbors of station i. For attention weight vectors, Let be the normalized attention coefficient of site i to site j. For the characteristics of site i, For the characteristics of site j, Features of neighboring site k It is an exponential function. This is the activation function.

3. The atmospheric monitoring site optimization and identification method according to claim 2, characterized in that, Residual connections are introduced in the two attention layers for quality control. Shallow features are added to deep features using the following formula to form an identity mapping path: ； In the formula, The feature transformation function of the current layer. The edge feature matrix, Output features for the current layer. This is the output feature of the previous layer.

4. The atmospheric monitoring site optimization and identification method according to claim 1, characterized in that, When training the pollution propagation model using the graph structure data, some sites are randomly masked using a dynamic mask; wherein, the dynamic mask is represented as: ； In the formula, This indicates the site mask ratio, e represents the natural constant, and epoch represents the training epoch.

5. The atmospheric monitoring site optimization and identification method according to claim 1, characterized in that, The three-objective constrained optimization model is expressed as follows: ； In the formula, For the number of sites, For coverage area, Let T be the prediction error, and T be the matrix transpose. As constraints, The number of measurement points, Let the minimum value of the three objective functions be obtained. For the number of stations calculated, Indicates whether to select the i-th station; The number of stations is determined using the following formula: ； The coverage area is determined using the following formula: ； In the formula, Let i be the Voronoi area; The prediction error is determined using the following formula: ； In the formula, To the maximum allowable prediction error, y represents the predicted value from the pollution transmission model, y represents the actual value, and n represents the total number of samples. This represents the mean square error.

6. The atmospheric monitoring site optimization and identification method according to claim 1, characterized in that, The Pareto optimal solution set is obtained by solving the three-objective constrained optimization model using the NSGA-III algorithm, including: Population initialization: Randomly generate an initial population P0 and calculate the objective function value for each individual; Reference point generation: Generate a set of reference points based on the number and distribution of the objective function; Weight vector generation: Generate a set of adaptive weight vectors based on the reference point; Non-dominated ranking and association: Perform non-dominated ranking on the population and associate each individual with the nearest reference point; Selection, crossover, and mutation: Generate offspring population Q0 using selection, crossover, and mutation operators; Population update: The initial population P0 and the offspring population Q0 are merged to form a new population R0. Then, R0 is sorted and associated in a non-dominated manner to select the next generation parent population P1. Repeat the process of generating weight vectors, non-dominated sorting and association, selection, crossover and mutation, and population update until the termination condition is met, and output the Pareto optimal solution set.

7. An atmospheric monitoring site optimization and identification device, characterized in that, The device includes: The node feature acquisition module is configured to acquire the node features of the monitoring station; wherein, the node features of the monitoring station include normalized latitude and longitude, Voronoi area and spatial neighborhood density. An edge feature construction module is configured to construct edge features; wherein the edge features include normalized inter-site distances and AQI correlation coefficients; The graph structure construction module is configured to construct graph structure data based on node features, edge features, and inter-site connections. Inter-site connections are determined as follows: a distance threshold and a correlation coefficient threshold are set; using the Pearson correlation coefficient as the AQI correlation coefficient, all site pairs are iterated through; if the distance between sites is less than the distance threshold and the absolute value of the AQI correlation coefficient is greater than the correlation coefficient threshold, then a valid connection is determined between the two sites; for sites with fewer than a set number of candidate connections after filtering, the nearest neighbor sites of the site are added as connecting sites to ensure that the number of connected sites of the site is not less than the set number. The model training module is configured to construct a pollution propagation model and train the pollution propagation model using the graph structure data; wherein, the pollution propagation model adopts a two-layer graph attention network with residual connections, and fuses node features and edge features through a dynamic masking mechanism to model the pollution propagation process, and outputs the feature vector prediction value of each monitoring station. The optimization solution module is configured to establish a three-objective constrained optimization model with the three objectives of minimizing the number of stations, maximizing the coverage area, and minimizing the prediction error. The module then uses the NSGA-III algorithm to solve the three-objective constrained optimization model to obtain a Pareto optimal solution set. Based on the Pareto optimal solution set, the module outputs an optimized layout scheme for monitoring stations. The prediction error is the prediction error of the training pollution propagation model.

8. An electronic device, characterized in that, include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes the computer execution instructions stored in the memory to implement the atmospheric monitoring site optimization identification method as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the atmospheric monitoring site optimization and identification method as described in any one of claims 1-6.

Images