A PM for a cluster site 2.5 Concentration long-term prediction method

Abstract

This invention discloses a PM for cluster sites. 2.5 The long-term concentration prediction method, based on deep learning technology, achieves long-term prediction for multiple monitoring stations in a region, with the advantages of wide prediction range and high prediction accuracy. A hierarchical MIC selector is designed to remove redundant station data, which helps reduce computational costs and improve prediction performance. A ring-shaped grid position encoding is designed to provide additional spatial information at low computational cost. A hierarchical attention feature matching layer is designed to direct the model's attention to atmospheric features that are more relevant to the current prediction, which is beneficial for extracting dynamic spatiotemporal correlations between features. An attention loop module is designed to dynamically adjust the model through the attention mechanism, which helps enhance the model's regional prediction robustness and long-term dependency extraction capability.

Description

Technical Field

[0001] This invention belongs to the field of air quality forecasting technology, and particularly relates to a PM2.5 meter for cluster monitoring stations. 2.5 Long-term concentration prediction method. Background Technology

[0002] With the rapid development of industrialization and informatization in human society, air pollution has become one of the most serious environmental problems of this century. Fine particulate matter (PM2.5), with a diameter of less than 2.5 micrometers, is a major contributor to air pollution. 2.5 PM2.5 is one of the main air pollutants; it can be inhaled into the lungs and is therefore also called respirable particulate matter. 2.5 Its large specific surface area makes it easy to adsorb pathogens and heavy metal ions, and it is also effective when exposed to high concentrations of PM2.5. 2.5 Environmental pollution poses a significant threat to human health. Air pollution has widespread adverse effects on both human health and the natural environment, thus drawing sustained attention from all sectors of society. Accurate, long-term air quality forecasts are of vital importance to both governments and citizens.

[0003] Data-driven methods (statistical methods) have become a commonly used approach in air quality forecasting in recent years. Based on mathematical statistics, probability theory, and stochastic processes, statistical methods utilize the temporal and spatial correlations between atmospheric data to construct statistical models that predict future concentrations of air pollutants. Monitoring values ​​of various air pollutant concentrations and meteorological elements are commonly used as input data for these models. Statistical models do not model atmospheric physicochemical processes using specialized knowledge; instead, they model regression prediction functions for air pollutants based solely on historical data. As data-driven statistical models, their aim is to construct approximate mathematical expressions of the relationship between various influencing factors and pollutant concentrations, and to estimate actual model parameters based on sample data to achieve optimal prediction results. Statistical models have low sensitivity to model parameters, thus exhibiting greater adaptability to the accuracy of sample data. Furthermore, although statistical models consume significant computational resources during training, the cost per prediction after training is low and the efficiency is high.

[0004] Deep learning, as a representative data-driven method, has attracted widespread attention from researchers in various fields in recent years due to its powerful feature processing capabilities and good model performance. Specifically in the field of air quality forecasting, with increasing global attention to atmospheric issues, more and more ground-based monitoring stations have been put into use, leading to an increase in the time span and richness of monitoring data. This provides a research foundation for deep learning models that require large data volumes. In current research, the challenge of air quality forecasting technology lies in simultaneously achieving wide-area forecasting range and long-term forecasting accuracy. Summary of the Invention

[0005] To address the above problems, this invention provides a PM for cluster sites. 2.5 Long-term concentration prediction methods can predict PM2.5 concentrations at multiple monitoring stations within a region. 2.5 It can make accurate long-term concentration predictions with high accuracy and good model robustness.

[0006] A PM for cluster sites 2.5 The long-term concentration prediction method utilizes a hierarchical attention recurrent neural network model to predict the PM2.5 concentration at future times for all monitoring stations within the region. 2.5 The concentration is determined by a method where each monitoring station in the cluster serves as the central monitoring station in turn, supplemented by information from other surrounding monitoring stations, to predict the PM2.5 concentration at the central monitoring station at future times. 2.5 Concentration is determined by iterating through all monitoring stations within the region to complete cluster prediction; the hierarchical attention recurrent neural network model includes a hierarchical MIC selector, a hierarchical attention feature matching layer, an attention recurrent module, and a fully connected interpretation layer; the construction method of the hierarchical attention recurrent neural network model is as follows:

[0007] S1: Sequentially treat all monitoring stations in the cluster as the central monitoring station and perform the fusion data acquisition operation to obtain the fusion data corresponding to each monitoring station. The fusion data acquisition operation is as follows:

[0008] S11: Divide the area within the maximum monitoring range of the central monitoring station into multiple concentric circle areas, with the central monitoring station as the center.

[0009] S12: Obtain PM data for the central monitoring station from each concentric circle area using a layered MIC selector. 2.5 The monitoring station with the highest concentration MIC is used as the surrounding monitoring station, and the location codes of the surrounding monitoring stations at each layer are also obtained.

[0010] S13: A hierarchical attention feature matching layer is used to obtain the matching degree between the feature vectors of the surrounding monitoring stations and the feature vector of the central monitoring station at different time points. The feature vectors include location encoding, PM... 2.5 Concentration, PM 10 Concentration, SO2 concentration, temperature, humidity, wind speed, wind direction, visibility;

[0011] S14: Use the matching degree of each layer's surrounding monitoring stations at different times as the attention weight, and apply the attention weight to the feature vectors of each layer's surrounding monitoring stations at different times to obtain the matched surrounding monitoring station data of the central monitoring station corresponding to each layer's surrounding monitoring stations. Concatenate the matched surrounding monitoring station data of each layer with the feature vector of the central monitoring station to obtain the cluster station data corresponding to the central monitoring station.

[0012] S15: The attention loop module is used to extract and fuse features from the cluster site data of the central monitoring site to obtain the fused data corresponding to the central monitoring site.

[0013] S2: Input the fused data from each monitoring station into the fully connected interpretation layer sequentially, and iterate through all monitoring stations in the entire area to obtain the PM2.5 at each monitoring station at future time. 2.5 Concentration prediction;

[0014] S3: Based on the PM levels at each monitoring station at future times. 2.5 Concentration prediction and PM 2.5 The root mean square error loss function is constructed from the actual concentration values;

[0015] S4: Determine whether the root mean square error loss function value is less than the set value. If yes, the hierarchical attention recurrent neural network model at this time is the final hierarchical attention recurrent neural network model. If no, use the Adam backpropagation algorithm to adjust the parameters of the hierarchical attention feature matching layer, attention recurrent module and fully connected interpretation layer, and then use the adjusted hierarchical attention feature matching layer, attention recurrent module and fully connected interpretation layer to re-execute steps S1 to S4.

[0016] Furthermore, the spacing between adjacent concentric circles in the multi-layer concentric circle region in step S11 is adjusted according to the actual situation of the cluster site.

[0017] Furthermore, in step S12, the method for obtaining the location code of any surrounding monitoring station of the central monitoring station is as follows:

[0018] According to the set distance resolution akm and azimuth resolution The multi-layered concentric circular regions are divided into meshes to obtain... There are 1 grid, where r is the radius of the maximum relevance range of the central monitoring station, in km;

[0019] Each grid is distinguished by a binary integer pair (i,j) consisting of a distance index i and an azimuth index j. Here, grid (i,j) represents a distance range of (ia, (i+1)a] from the central monitoring station and an azimuth range of... The area;

[0020] The location codes of surrounding monitoring stations are determined based on the binary integer pairs (i,j) of the grids in which they fall. If the azimuth index j is even, then the location codes of the surrounding monitoring stations are:

[0021]

[0022] At this point, j = 2k, where k is a natural number and C is a set constant;

[0023] If the azimuth index j is odd, then the location codes of the surrounding monitoring stations are:

[0024]

[0025] At this point, j = 2k + 1.

[0026] Furthermore, in step S13, the method for obtaining the matching degree between the feature vectors of the surrounding monitoring stations in any layer and the feature vector of the central monitoring station is as follows:

[0027] Obtain the feature vector of the surrounding monitoring stations at the current time t. With any time t * Feature vectors of central monitoring stations The relevance matching value a t :

[0028]

[0029] in, W T All of these are weighted;

[0030] Based on the correlation matching value a within a set time period T before and after the current time t. t Calculate the matching degree α between the feature vectors of the surrounding monitoring stations and the feature vector of the central monitoring station at the current time t. t :

[0031]

[0032] Wherein, the time variable τ = 1, 2, ..., T.

[0033] Furthermore, in step S14, the method for obtaining the matched surrounding monitoring station data of the central monitoring station corresponding to the surrounding monitoring stations at any layer is as follows:

[0034]

[0035] in, For any time t * At present, the surrounding monitoring stations s correspond to the central monitoring station s * The data has been matched with data from surrounding monitoring stations.

[0036] Further, in step S15, the attention loop module includes a bidirectional GRU, an attention mechanism unit, and a fusion unit, and the fusion unit includes a bidirectional GRU and a fusion normalization unit.

[0037] The method for obtaining the fused data corresponding to the central monitoring stations is as follows:

[0038] The cluster site data corresponding to the central monitoring station, obtained by concatenating the matched surrounding monitoring station data of each layer with the feature vector of the central monitoring station, is input into a bidirectional GRU for feature extraction, resulting in shallow feature u. t ;

[0039] shallow features u t The input attention mechanism unit performs feature extraction to obtain intermediate features. Among them, intermediate features The extraction method is as follows:

[0040]

[0041] Among them, w y-1 For the fused data of the central monitoring station at the previous time t-1, All of these are set weights in the attention mechanism unit, b t Let β be the attention score at time t. t The weight is the weight at the current time t, where T is the set time step length;

[0042] intermediate features The bidirectional GRU in the input fusion unit is used for feature extraction to obtain deep features f. t ;

[0043] deep features f t The fusion normalization unit in the input fusion unit makes the deep features f t With shallow features u t Summation and normalization are performed to obtain the fused data w corresponding to the central monitoring station at the current time t. t .

[0044] Beneficial effects:

[0045] 1. This invention provides a PM for cluster sites. 2.5 The proposed long-term concentration prediction method employs a hierarchical MIC selector and a ring-grid location coding, simultaneously considering the temporal and spatial correlations between atmospheric characteristics of stations at different distance levels. This method hierarchically selects surrounding stations near the central monitoring station that possess rich input information and the strongest spatiotemporal correlation, injecting additional station spatial information, thereby eliminating stations with PM2.5 concentrations closely correlated with the central monitoring station. 2.5 The use of concentration-independent redundant site data reduces computational costs and improves prediction performance, providing an effective solution to the problem of long-term forecasting of air quality at multiple sites and promoting technological development in related fields.

[0046] 2. This invention provides a PM for cluster sites.2.5 The proposed long-term concentration prediction method utilizes a hierarchical attention recurrent network, which dynamically adjusts the input weights spatiotemporally using an attention mechanism, thereby enabling the model to focus on atmospheric features that have a greater impact on the prediction results. Simultaneously, deep spatiotemporal features are extracted using multi-layer GRU units and dynamically adjusted through an attention mechanism to enhance the model's regional prediction robustness and long-term dependency extraction capabilities.

[0047] 3. This invention provides a PM for cluster sites. 2.5 The long-term concentration prediction method, based on deep learning technology, requires no prior expert knowledge or atmospheric physicochemical modeling. It only needs to provide the data required for model training to predict PM2.5 concentrations at multiple monitoring stations within a region. 2.5 The model provides accurate long-term concentration predictions. Furthermore, it exhibits good robustness to monitoring sites, with its predictive performance remaining consistent regardless of changes in monitoring sites within the region, making it suitable for any cluster of monitoring sites. Attached Figure Description

[0048] Figure 1 This is a diagram of the overall network structure of the present invention.

[0049] Figure 2 This is a schematic diagram of the layered MIC selector and annular grid position encoding proposed in this invention.

[0050] Figure 3 This is a structural diagram of the hierarchical attention feature matching layer proposed in this invention.

[0051] Figure 4 This is a structural diagram of the attention loop module proposed in this invention. Detailed Implementation

[0052] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings.

[0053] A PM for cluster sites 2.5 The long-term concentration prediction method utilizes a hierarchical attention recurrent neural network model to predict the PM2.5 concentration at future times for all monitoring stations within the region. 2.5 The concentration is determined by a method where each monitoring station in the cluster serves as the central monitoring station in turn, supplemented by information from other surrounding monitoring stations, to predict the PM2.5 concentration at the central monitoring station at future times. 2.5 Concentration, iterating through all monitoring stations in the region to complete cluster prediction; among them, the hierarchical attention recurrent neural network large model includes a hierarchical MIC picker, a hierarchical attention feature matching layer, an attention recurrent module and a fully connected interpretation layer;

[0054] At the same time, such as Figure 1 As shown, the method for constructing the hierarchical attention recurrent neural network large model is as follows:

[0055] S0: Construct a cluster site prediction environment with multiple sites in the region. Divide the desired prediction area and record the spatial location of all monitoring sites within the prediction area; further, perform data preprocessing operations on the monitoring data of all sites in the area.

[0056] S1: Sequentially treat all monitoring stations in the cluster as the central monitoring station and perform the fusion data acquisition operation to obtain the fusion data corresponding to each monitoring station. The fusion data acquisition operation is as follows:

[0057] S11: Divide the area within the maximum monitoring range of the central monitoring station into multiple concentric circle areas, with the central monitoring station as the center.

[0058] S12: Obtain PM data for the central monitoring station from each concentric circle area using a layered MIC selector. 2.5 The monitoring station with the highest concentration MIC is used as the surrounding monitoring station, and the location codes of the surrounding monitoring stations at each layer are also obtained.

[0059] like Figure 2 As shown in the left sub-figure, within a range of a maximum spatial radius *r* for a given monitoring station, the hierarchical MIC selector uses the spatial distance from the central monitoring station as the criterion to sequentially divide all surrounding stations within the range into *m* different distance levels. Data is then filtered sequentially from each level to obtain the filtered cluster of station data. Further, the hierarchical MIC selector sequentially filters out PMs relevant to the central monitoring station from each distance level. 2.5 The surrounding monitoring site with the highest concentration MIC measurement, among which, Figure 2 The red triangle in the diagram represents the central monitoring station, and the red circle represents the surrounding monitoring station with the highest metric. It should be noted that the hierarchical MIC selector performs the screening of surrounding monitoring stations in parallel for all monitoring stations in the prediction area.

[0060] The method for obtaining the location code of any surrounding monitoring station of the central monitoring station is as follows:

[0061] According to the set distance resolution akm and azimuth resolution The multi-layered concentric circular regions are divided into meshes to obtain... There are 1 grid, where r is the radius of the maximum relevance range of the central monitoring station, in km;

[0062] Each grid is distinguished by a binary integer pair (i,j) consisting of a distance index i and an azimuth index j. Here, grid (i,j) represents a distance range of (ia, (i+1)a] from the central monitoring station and an azimuth range of... The area;

[0063] The location codes of surrounding monitoring stations are determined based on the binary integer pairs (i,j) of the grids in which they fall. If the azimuth index j is even, then the location codes of the surrounding monitoring stations are:

[0064]

[0065] At this point, j = 2k, where k is a natural number and C is a set constant;

[0066] If the azimuth index j is odd, then the location codes of the surrounding monitoring stations are:

[0067]

[0068] At this point, j = 2k + 1.

[0069] like Figure 2 As shown in the right sub-figure, the area with a maximum spatial radius of rkm of a certain monitoring station is divided into grids in the form of a ring grid; all monitoring stations near the central monitoring station fall into the divided grids, and the monitoring data of each surrounding station is embedded with a location encoding vector according to the grid number in which it falls.

[0070] S13: A hierarchical attention feature matching layer is used to obtain the matching degree between the feature vectors of the surrounding monitoring stations and the feature vector of the central monitoring station at different time points. The feature vectors include location encoding, PM... 2.5 Concentration, PM 10 Concentration, SO2 concentration, temperature, humidity, wind speed, wind direction, visibility;

[0071] The method for obtaining the matching degree between the feature vectors of the surrounding monitoring stations and the feature vector of the central monitoring station at any layer is as follows:

[0072] S13a: Obtain the feature vectors of the surrounding monitoring stations at the current time t. With any time t * Feature vectors of central monitoring stations The relevance matching value a t :

[0073]

[0074] in, W T All of these are weighted;

[0075] It should be noted that the central monitoring station typically uses s * The total number of times for the surrounding monitoring stations is equal to the total number of times for the central monitoring station. However, the two matching times can be different. For example, if there are 10 times in total, the second time of the surrounding monitoring stations can be used to match the third time of the central monitoring station. However, if you want to obtain the output of the third time of the surrounding monitoring stations after matching, you need to match the features of the surrounding monitoring stations from time 1 to time 10 with the features of the central monitoring station at time 3.

[0076] S13b: Based on the correlation matching value a within a set time period T before and after the current time t. t Calculate the matching degree α between the feature vectors of the surrounding monitoring stations and the feature vector of the central monitoring station at the current time t. t :

[0077]

[0078] Wherein, the time variable t = 1, 2, ..., T. It should be noted that t is a specific moment within the time period T.

[0079] Therefore, step S13 aims to construct a hierarchical attention feature matching neural network layer based on the distance level m determined in step 2; as shown... Figure 3 As shown, the hierarchical attention feature matching layer utilizes an attention mechanism to match the feature vectors of surrounding monitoring stations with the highest metric values ​​at each level. Characteristics of the central monitoring station Independently perform correlation matching and adjust the input weights of data from surrounding monitoring stations with the highest values ​​for each metric. W T Finally, attention weight α is used. t It quantitatively indicates the degree of matching.

[0080] S14: Use the matching degree of each layer's surrounding monitoring stations at different times as the attention weight, and weight the feature vectors of each layer's surrounding monitoring stations at different times to obtain the matched surrounding monitoring station data of the central monitoring station corresponding to each layer's surrounding monitoring stations. Concatenate the matched surrounding monitoring station data of each layer with the feature vector of the central monitoring station to obtain the cluster station data of the corresponding central monitoring station.

[0081] The method for obtaining the matched surrounding monitoring station data of the central monitoring station from the surrounding monitoring stations at any layer is as follows:

[0082]

[0083] in, For any time t * At present, the surrounding monitoring stations s correspond to the central monitoring station s * The data has been matched with data from surrounding monitoring stations.

[0084] S15: The attention loop module is used to extract and fuse features from the cluster site data of the central monitoring site to obtain the fused data corresponding to the central monitoring site.

[0085] It should be noted that, as Figure 1 As shown, the hierarchical MIC picker and the ring grid position encoding are the data preprocessing parts of the hierarchical attention recurrent network, while the main network is obtained by concatenating a hierarchical attention feature matching layer and multiple attention recurrent module sequences. The large model of the hierarchical attention recurrent neural network uses a fully connected layer as the interpretation output layer. Further, as... Figure 4 As shown, the attention loop module includes a bidirectional GRU (Gated Recurrent Unit), an attention mechanism unit, and a fusion unit, and the fusion unit includes a bidirectional GRU and a fusion normalization unit.

[0086] The method for obtaining the fused data corresponding to the central monitoring stations is as follows:

[0087] The cluster site data corresponding to the central monitoring station, obtained by concatenating the matched surrounding monitoring station data of each layer with the feature vector of the central monitoring station, is input into a bidirectional GRU for feature extraction, resulting in shallow feature u. t ;

[0088] shallow features u t The input attention mechanism unit performs feature extraction to obtain intermediate features. Among them, intermediate features The extraction method is as follows:

[0089]

[0090] Among them, w t-1 For the fused data of the central monitoring station at the previous time t-1, All of these are set weights in the attention mechanism unit, b t Let β be the attention score at time t. t The weight is the weight at the current time t, where T is the set time step length;

[0091] intermediate features The bidirectional GRU in the input fusion unit is used for feature extraction to obtain deep features f. t ;

[0092] deep features ft The fusion normalization unit in the input fusion unit makes the deep features f t With shallow features u t Summation and normalization are performed to obtain the fused data w corresponding to the central monitoring station at the current time t. t .

[0093] S2: Input the fused data from each monitoring station into the fully connected interpretation layer sequentially, and iterate through all monitoring stations in the entire area to obtain the PM2.5 at each monitoring station at future time. 2.5 Concentration prediction;

[0094] S3: Based on the PM levels at each monitoring station at future times. 2.5 Concentration prediction and PM 2.5 The root mean square error loss function is constructed from the actual concentration values;

[0095] S4: Determine whether the root mean square error loss function value is less than the set value. If yes, the hierarchical attention recurrent neural network model at this time is the final hierarchical attention recurrent neural network model. If no, use the Adam backpropagation algorithm to adjust the parameters of the hierarchical attention feature matching layer, attention recurrent module and fully connected interpretation layer, and then use the adjusted hierarchical attention feature matching layer, attention recurrent module and fully connected interpretation layer to re-execute steps S1 to S4.

[0096] Therefore, this invention trains the hierarchical attention recurrent neural network based on cluster site data and a large-scale hierarchical attention recurrent neural network model. The reprocessed cluster site data is divided into training and validation sets. The root mean square error (RMSE) loss function is selected, and the Adam backpropagation algorithm is used to train the large-scale hierarchical attention recurrent neural network model, updating the model parameters. Furthermore, the model hyperparameters are adjusted using the validation set data, and the model's predictive performance is tested until the hierarchical attention recurrent network reaches its optimal predictive performance. Training then ends, and the neural network model file and process file are saved.

[0097] In summary, this invention provides a PM for cluster sites. 2.5The long-term concentration prediction method first constructs a hierarchical MIC (Maximal Information Coefficient) selector, using spatial distance as the discrimination index to divide monitoring stations into different distance levels, and sequentially filters data from each level. Then, the filtered cluster station data is further processed, embedding location encoding vectors into the monitoring station data in a ring grid. Next, based on the determined number of distance levels, a hierarchical attention feature matching neural network layer is constructed. This layer independently adjusts the input weights of each monitoring station's data using an attention mechanism. Simultaneously, this invention constructs an attention recurrent network module based on GRU and a feedback attention mechanism, and uses this module to adjust the extracted features using the attention mechanism. Finally, a large-scale hierarchical attention recurrent neural network model achieving optimal prediction performance is obtained, providing an effective solution for the long-term forecasting of air quality at multiple stations and promoting technological development in related fields.

[0098] Of course, the present invention may have other various embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art can make various corresponding changes and modifications according to the present invention, but these corresponding changes and modifications should all fall within the protection scope of the appended claims.

Claims

1. A PM for clustered sites 2.5 The long-term concentration prediction method is characterized by, A hierarchical attention recurrent neural network model is used to predict the PM2.5 levels at all monitoring stations within the region at future times. 2.5 The concentration is determined by a method where each monitoring station in the cluster serves as the central monitoring station in turn, supplemented by information from other surrounding monitoring stations, to predict the PM2.5 concentration at the central monitoring station at future times. 2.5 Concentration is determined by iterating through all monitoring stations within the region to complete cluster prediction; the hierarchical attention recurrent neural network model includes a hierarchical MIC selector, a hierarchical attention feature matching layer, an attention recurrent module, and a fully connected interpretation layer; the construction method of the hierarchical attention recurrent neural network model is as follows: S1: Sequentially treat all monitoring stations in the cluster as the central monitoring station and perform the fusion data acquisition operation to obtain the fusion data corresponding to each monitoring station. The fusion data acquisition operation is as follows: S11: Divide the area within the maximum monitoring range of the central monitoring station into multiple concentric circle areas, with the central monitoring station as the center. S12: Obtain PM data for the central monitoring station from each concentric circle area using a layered MIC selector. 2.5 The monitoring station with the highest concentration MIC is used as the surrounding monitoring station, and the location codes of the surrounding monitoring stations at each layer are also obtained. S13: A hierarchical attention feature matching layer is used to obtain the matching degree between the feature vectors of the surrounding monitoring stations and the feature vector of the central monitoring station at different time points. The feature vectors include location encoding, PM... 2.5 Concentration, PM 10 Concentration, SO2 concentration, temperature, humidity, wind speed, wind direction, visibility; S14: Use the matching degree of each layer's surrounding monitoring stations at different times as the attention weight, and apply the attention weight to the feature vectors of each layer's surrounding monitoring stations at different times to obtain the matched surrounding monitoring station data of the central monitoring station corresponding to each layer's surrounding monitoring stations. Concatenate the matched surrounding monitoring station data of each layer with the feature vector of the central monitoring station to obtain the cluster station data corresponding to the central monitoring station. S15: The attention loop module is used to extract and fuse features from the cluster site data of the central monitoring site to obtain the fused data corresponding to the central monitoring site. S2: Input the fused data from each monitoring station into the fully connected interpretation layer sequentially, and iterate through all monitoring stations in the entire area to obtain the PM2.5 at each monitoring station at future time. 2.5 Concentration prediction; S3: Based on the PM levels at each monitoring station at future times. 2.5 Concentration prediction and PM 2.5 The root mean square error loss function is constructed from the actual concentration values; S4: Determine whether the root mean square error loss function value is less than the set value. If yes, the hierarchical attention recurrent neural network model at this time is the final hierarchical attention recurrent neural network model. If no, use the Adam backpropagation algorithm to adjust the parameters of the hierarchical attention feature matching layer, attention recurrent module and fully connected interpretation layer, and then use the adjusted hierarchical attention feature matching layer, attention recurrent module and fully connected interpretation layer to re-execute steps S1 to S4.

2. A PM for cluster sites as described in claim 1 2.5 The long-term concentration prediction method is characterized by, The spacing between adjacent concentric circles in the multi-layered concentric circle region in step S11 is adjusted according to the actual situation of the cluster site.

3. A PM for cluster sites as described in claim 1 2.5 The long-term concentration prediction method is characterized by, In step S12, the method for obtaining the location code of any surrounding monitoring station of the central monitoring station is as follows: According to the set distance resolution akm and azimuth resolution The multi-layered concentric circular regions are divided into meshes to obtain... There are 1 grid, where r is the radius of the maximum relevance range of the central monitoring station, in km; Each grid is distinguished by a binary integer pair (i,j) consisting of a distance index i and an azimuth index j. Here, grid (i,j) represents a distance range of (ia, (i+1)a] from the central monitoring station and an azimuth range of... The area; The location codes of surrounding monitoring stations are determined based on the binary integer pairs (i,j) of the grids in which they fall. If the azimuth index j is even, then the location codes of the surrounding monitoring stations are: At this point, j = 2k, where k is a natural number and C is a set constant; If the azimuth index j is odd, then the location codes of the surrounding monitoring stations are: At this point, j = 2k + 1.

4. A PM for cluster sites as described in claim 1 2.5 The long-term concentration prediction method is characterized by, In step S13, the method for obtaining the matching degree between the feature vectors of the surrounding monitoring stations in any layer and the feature vector of the central monitoring station is as follows: Obtain the feature vector of the surrounding monitoring stations at the current time t. With any time t * Feature vectors of central monitoring stations The relevance matching value a t : in, W T All of these are weighted; Based on the correlation matching value a within a set time period T before and after the current time t. t Calculate the matching degree α between the feature vectors of the surrounding monitoring stations and the feature vector of the central monitoring station at the current time t. t : Wherein, the time variable τ = 1, 2, ..., T.

5. The method for long-term prediction of PM2.5 concentration for clustered sites as described in claim 4, characterized in that, In step S14, the method for obtaining the matched surrounding monitoring station data of the central monitoring station for any layer of surrounding monitoring stations is as follows: in, For any time t * At present, the surrounding monitoring stations s correspond to the central monitoring station s * The data has been matched with data from surrounding monitoring stations.

6. A PM for cluster sites as described in claim 1 2.5 The long-term concentration prediction method is characterized by, In step S15, the attention loop module includes a bidirectional GRU, an attention mechanism unit, and a fusion unit, and the fusion unit includes a bidirectional GRU and a fusion normalization unit. The method for obtaining the fused data corresponding to the central monitoring stations is as follows: The cluster site data corresponding to the central monitoring station, obtained by concatenating the matched surrounding monitoring station data of each layer with the feature vector of the central monitoring station, is input into a bidirectional GRU for feature extraction, resulting in shallow feature u. t ; shallow features u t The input attention mechanism unit performs feature extraction to obtain intermediate features. Among them, intermediate features The extraction method is as follows: Among them, w t-1 For the fused data of the central monitoring station at the previous time t-1, All of these are set weights in the attention mechanism unit, b t Let β be the attention score at time t. t The weight is the weight at the current time t, where T is the set time step length; intermediate features The bidirectional GRU in the input fusion unit is used for feature extraction to obtain deep features f. t ; deep features f t The fusion normalization unit in the input fusion unit makes the deep features f t With shallow features u t Summation and normalization are performed to obtain the fused data w corresponding to the central monitoring station at the current time t. t .

Images