PM2.5 concentration long-time prediction method for cluster site

Abstract

The invention discloses a PM2.5 concentration long-time prediction method for cluster stations, which realizes long-time prediction of multiple monitoring stations in a region based on a deep learning technology, and has the advantages of wide prediction range and high prediction precision. A layered MIC selector is designed, redundant site data is eliminated, the calculation cost is reduced, and the prediction performance is improved; annular grid position codes are designed, so that additional space information can be provided with low calculation cost; a layered attention feature matching layer is designed, model attention is guided to atmospheric features more related to current prediction, and dynamic space-time correlation between the features can be extracted; an attention circulation module is designed, dynamic feedback adjustment is performed through an attention mechanism, and the regional prediction robustness and the long-time dependence extraction capability of the model can be enhanced.

Description

Technical Field

[0001] The present invention belongs to the technical field of air quality forecasting, and in particular relates to a PM2.5 2.5 Long-term concentration prediction method. Background Art

[0002] With the rapid development of industrialization and informatization of human society, air pollution has become one of the most serious environmental problems of this century. 2.5 ) is one of the main pollutants in the air. It can be inhaled into the lungs and is therefore also called respirable particulate matter. 2.5 The large specific surface area makes it easy to adsorb pathogens and heavy metal ions, and it is exposed to high concentrations of PM 2.5 The environment will cause great harm to human health. Air pollution has a wide range of adverse effects on human health and the natural environment, which has aroused continuous attention from all walks of life. Long-term and accurate air quality forecasts are of great significance to both the government and citizens.

[0003] Data-driven methods (statistical methods) are a type of method commonly used in the field of air quality prediction in recent years. Statistical methods are based on mathematical statistics, probability theory and random processes. They use the time series correlation and spatial correlation between atmospheric data to construct statistical models to give future concentration forecasts of air pollutants. The concentration monitoring values ​​of various air pollutants and meteorological elements are often used as input data for statistical models. Statistical models do not model the physical and chemical processes of the atmosphere through professional knowledge, but only use historical data to model the regression prediction function of air pollutants. As a data-driven statistical model, it aims to construct an approximate mathematical expression between various influencing factors and pollutant concentrations, and estimate the actual model parameters based on sample data to achieve the best prediction effect. Statistical models are less sensitive to model parameters, so they are more adaptable to the accuracy of sample data. At the same time, although statistical models consume a lot of computing resources during training, the cost of a single prediction is low and the efficiency is high after training.

[0004] As a representative data-driven method, deep learning has attracted extensive attention from researchers in various fields in recent years due to its powerful feature processing capabilities and good model effects. Specifically in the field of air quality prediction, with the increasing attention of countries to atmospheric issues, more and more ground monitoring stations have been put into use, and the time span and richness of monitoring data have been improved, which provides a research basis for deep learning models with large data volume requirements. In current research work, the difficulty of air quality forecasting technology lies in the difficulty of achieving both wide-area prediction range and long-term prediction accuracy at the same time. Summary of the invention

[0005] To solve the above problems, the present invention provides a PM for cluster sites. 2.5 The long-term concentration prediction method can be used to predict the PM concentration at multiple monitoring stations in the region. 2.5 The model can make accurate long-term predictions of concentration with high prediction accuracy and good model robustness.

[0006] A PM for cluster sites 2.5 The long-term concentration prediction method uses a hierarchical attention recurrent neural network model to predict the PM concentration of all monitoring stations in the region at future times. 2.5 Each monitoring station in the cluster is used as the central monitoring station in turn, and the information of other surrounding monitoring stations is used as input to predict the PM concentration of the central monitoring station at the future time. 2.5 concentration, looping through all monitoring sites in the area to complete cluster prediction; wherein, the hierarchical attention recurrent neural network large model includes a hierarchical MIC selector, a hierarchical attention feature matching layer, an attention recurrent module and a fully connected interpretation layer; at the same time, the construction method of the hierarchical attention recurrent neural network large model is as follows:

[0007] S1: All monitoring stations in the cluster are sequentially used as central monitoring stations to perform fusion data acquisition operations to obtain 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 layers of concentric circle areas with the central monitoring station as the center;

[0009] S12: Use the layered MIC selector to obtain PM of the central monitoring site in each layer of concentric circle area. 2.5 The monitoring station with the largest concentration MIC measurement is used as the surrounding monitoring station, and the location codes of the surrounding monitoring stations in each layer are obtained at the same time;

[0010] S13: A hierarchical attention feature matching layer is used to obtain the matching degree between the feature vectors of the surrounding monitoring sites at each layer and the feature vector of the central monitoring site at different times, where the feature vector includes the position code, PM 2.5 Concentration, PM 10 concentration, SO2 concentration, temperature, humidity, wind speed, wind direction, visibility;

[0011] S14: taking the matching degree corresponding to each layer of surrounding monitoring stations at different times as the attention weight, weighting the feature vectors of each layer of surrounding monitoring stations at different times respectively, obtaining the matched surrounding monitoring station data of each layer of surrounding monitoring stations corresponding to the central monitoring station, and splicing 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: using the attention loop module to perform feature extraction and feature fusion on the cluster site data of the central monitoring site to obtain fused data corresponding to the central monitoring site;

[0013] S2: Input the fusion data corresponding to each monitoring station into the fully connected interpretation layer in turn, loop through all monitoring stations in the entire area, and obtain the PM of each monitoring station at the future time. 2.5 Concentration prediction value;

[0014] S3: Based on the PM of each monitoring station at the future time 2.5 Concentration prediction and PM 2.5 The actual value of concentration is used to construct the root mean square error loss function;

[0015] S4: Determine whether the root mean square error loss function value is less than the set value. If so, the layered attention recurrent neural network model at this time is the final layered attention recurrent neural network model. If not, use the Adam back-propagation algorithm to adjust the parameters of the layered attention feature matching layer, the attention recurrent module, and the fully connected interpretation layer, and then use the adjusted layered attention feature matching layer, the attention recurrent module, and the fully connected interpretation layer to re-execute steps S1 to S4.

[0016] Furthermore, the intervals between adjacent concentric circles in the multi-layer concentric circle area in step S11 are adjusted according to actual conditions of the cluster sites.

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

[0018] According to the set range resolution akm and azimuth resolution Divide the multi-layer concentric circle area into grids and obtain grids, where r is the radius of the maximum relevant range of the central monitoring station, in km;

[0019] Each grid is distinguished by a binary integer pair (i, j) consisting of a distance number i and an azimuth number j, where the grid (i, j) represents a distance range of (ia, (i+1)a] and an azimuth range of area;

[0020] The location codes of the surrounding monitoring stations are determined according to the binary integer pairs (i, j) of the grids in which the surrounding monitoring stations fall. If the azimuth angle sequence number j is an even number, the location codes of the surrounding monitoring stations are:

[0021]

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

[0023] If the azimuth angle number j is an odd number, the location code of the surrounding monitoring stations is:

[0024]

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

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

[0027] Get the feature vector of the current surrounding monitoring sites at the current time t At any time t * The characteristic vector of the central monitoring station The correlation matching value a t :

[0028]

[0029] in, W T All are set weights;

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

[0031]

[0032] Among them, the time variable τ=1,2,…,T.

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

[0034]

[0035] in, For any time t * When the current surrounding monitoring site s corresponds to the central monitoring site s * The data of surrounding monitoring stations have been matched.

[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 fusion data corresponding to the central monitoring station is as follows:

[0038] The cluster site data of the corresponding central monitoring site obtained by splicing the matched surrounding monitoring site data of each layer with the feature vector of the central monitoring site is input into the bidirectional GRU for feature extraction to obtain the shallow feature u t ;

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

[0040]

[0041] Among them, w y-1 is the fusion data corresponding to the central monitoring station at the previous time t-1, are the set weights in the attention mechanism unit, b t is the attention score at the current time t, β t is the weight at the current moment t, and T is the set time step length;

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

[0043] The deep feature f t The fusion normalization unit in the input fusion unit makes the deep feature f t With shallow features u t Sum and normalize to get the fusion data w corresponding to the central monitoring station at the current time t t .

[0044] Beneficial effects:

[0045] 1. The present invention provides a PM for cluster sites 2.5 The proposed long-term concentration prediction method, hierarchical MIC selector and annular grid position encoding, take into account the temporal and spatial correlations between the atmospheric characteristics of stations at each distance level, hierarchically select the surrounding stations with rich input information and the strongest spatiotemporal correlation from the central monitoring station and inject additional station spatial information, thereby eliminating the PM 2.5 Redundant station data that is independent of concentration can reduce computing costs and improve prediction performance, providing an effective solution to the long-term forecasting problem of multi-station air quality and promoting technological development in related fields.

[0046] 2. The present invention provides a PM for cluster sites2.5 The long-term concentration prediction method proposes a hierarchical attention recurrent network, which uses the attention mechanism to dynamically adjust the input weights in time and space, so that the model focuses on the atmospheric features that have a greater impact on the prediction results; at the same time, deep spatiotemporal features are extracted using multi-layer GRU units, and dynamic feedback adjustments are made through the attention mechanism to enhance the model's regional prediction robustness and long-term dependency extraction capabilities.

[0047] 3. The present invention provides a PM for cluster sites 2.5 The long-term concentration prediction method is based on deep learning technology and does not require any expert prior knowledge or atmospheric physical and chemical modeling. It only needs to provide the data required for model training to predict the PM concentration at multiple monitoring stations in the region. 2.5 The concentration can be accurately predicted in the long term; at the same time, the model described in the present invention has good robustness of monitoring sites, and the prediction performance of the model does not change dramatically with the change of monitoring sites in the region, and is applicable to any monitoring site cluster. BRIEF DESCRIPTION OF THE DRAWINGS

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

[0049] Figure 2 Schematic diagram of the hierarchical MIC selector and ring grid position coding proposed in the present 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 the present invention. DETAILED DESCRIPTION

[0052] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0053] A PM for cluster sites 2.5 The long-term concentration prediction method uses a hierarchical attention recurrent neural network model to predict the PM concentration of all monitoring stations in the region at future times. 2.5 Each monitoring station in the cluster is used as the central monitoring station in turn, and the information of other surrounding monitoring stations is used as input to predict the PM concentration of the central monitoring station at the future time. 2.5 concentration, looping through all monitoring sites in the area to complete cluster prediction; among them, 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;

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

[0055] S0: Construct a regional multi-site cluster site prediction environment. Divide the desired prediction area and record the spatial locations of all monitoring sites in the prediction area; further, perform data preprocessing operations on the monitoring data of all sites in the area.

[0056] S1: All monitoring stations in the cluster are sequentially used as central monitoring stations to perform fusion data acquisition operations to obtain 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 layers of concentric circle areas with the central monitoring station as the center;

[0058] S12: Use the layered MIC selector to obtain PM of the central monitoring site in each layer of concentric circle area. 2.5 The monitoring station with the largest concentration MIC measurement is used as the surrounding monitoring station, and the location codes of the surrounding monitoring stations in each layer are obtained at the same time;

[0059] like Figure 2 As shown in the left sub-figure, within the range of the maximum spatial radius r of a certain monitoring station, the hierarchical MIC selector uses the spatial distance from the central monitoring station as the discrimination index, divides all surrounding stations in the range into m different distance levels in turn, and completes data screening from each level in turn to obtain the screened cluster station data. Furthermore, the hierarchical MIC selector sequentially screens the PM of the central monitoring station in each distance level. 2.5 The surrounding monitoring stations with the largest concentration MIC measurements are: Figure 2 The red triangle in the figure represents the central monitoring station, and the red circle represents the surrounding monitoring station with the largest metric. It should be noted that the hierarchical MIC selector screens the surrounding monitoring stations for all monitoring stations in the prediction area in parallel.

[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 range resolution akm and azimuth resolution Divide the multi-layer concentric circle area into grids and obtain grids, where r is the radius of the maximum relevant range of the central monitoring station, in km;

[0062] Each grid is distinguished by a binary integer pair (i, j) consisting of a distance number i and an azimuth number j, where the grid (i, j) represents a distance range of (ia, (i+1)a] and an azimuth range of area;

[0063] The location codes of the surrounding monitoring stations are determined according to the binary integer pairs (i, j) of the grids in which the surrounding monitoring stations fall. If the azimuth angle sequence number j is an even number, the location codes of the surrounding monitoring stations are:

[0064]

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

[0066] If the azimuth angle number j is an odd number, the location code of the surrounding monitoring stations is:

[0067]

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

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

[0070] S13: A hierarchical attention feature matching layer is used to obtain the matching degree between the feature vectors of the surrounding monitoring sites at each layer and the feature vector of the central monitoring site at different times, where the feature vector includes the position code, 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 of any layer and the feature vector of the central monitoring station is as follows:

[0072] S13a: Obtain the feature vector of the current surrounding monitoring sites at the current time t At any time t * The characteristic vector of the central monitoring station The correlation matching value a t :

[0073]

[0074] in, W T All are set weights;

[0075] It should be noted that the central monitoring site usually adopts s * It indicates that the total number of moments of the surrounding monitoring stations is equal to the total number of moments of the central monitoring station, but when matching, the two matching moments can be different; for example, if there are 10 moments in total, the second moment of the surrounding monitoring station can be used to match the third moment of the central monitoring station; but if you want to obtain the output of the third moment of the matched surrounding monitoring station, you need to match the features of the surrounding monitoring stations at moments 1 to 10 with the features of the central monitoring station at the third moment.

[0076] S13b: Match the value a according to the correlation within the set time period T before and after the current time t t Calculate the matching degree α between the feature vector of the surrounding monitoring station 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 in the time period T, and t is a certain moment in the time period T.

[0079] It can be seen that step S13 aims to construct a hierarchical attention feature matching neural network layer according to the distance level number m determined in step 2; Figure 3 As shown in the figure, the hierarchical attention feature matching layer uses the attention mechanism to convert the feature vectors of the surrounding monitoring sites with the largest metric values ​​at each level Characteristics of the monitoring sites Independently perform correlation matching and adjust the input weights of the surrounding monitoring station data with the largest metric values W T , and finally use the attention weight α t Quantitatively indicates the degree of matching.

[0080] S14: taking the matching degree corresponding to each layer of surrounding monitoring stations at different times as the attention weight, weighting the feature vectors of each layer of surrounding monitoring stations at different times respectively, obtaining the matched surrounding monitoring station data of each layer of surrounding monitoring stations corresponding to the central monitoring station, and splicing 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;

[0081] The method for obtaining the matched surrounding monitoring site data of the surrounding monitoring sites corresponding to the central monitoring site of any layer is as follows:

[0082]

[0083] in, For any time t * When the current surrounding monitoring site s corresponds to the central monitoring site s * The data of surrounding monitoring stations have been matched.

[0084] S15: using the attention loop module to perform feature extraction and feature fusion on the cluster site data of the central monitoring site to obtain fused data corresponding to the central monitoring site;

[0085] It should be noted that if Figure 1 As shown in the figure, the hierarchical MIC selector and the ring grid position encoding are the data preprocessing parts of the hierarchical attention recurrent network, while the main body of the network is obtained by splicing a layer of hierarchical attention feature matching layer and multiple attention recurrent module sequences. The hierarchical attention recurrent neural network model uses a fully connected layer as the interpretation output layer. Figure 4 As shown, the attention cycle 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 fusion data corresponding to the central monitoring station is as follows:

[0087] The cluster site data corresponding to the central monitoring site obtained by splicing the matched surrounding monitoring site data of each layer with the feature vector of the central monitoring site is input into the bidirectional GRU for feature extraction to obtain the shallow feature u t ;

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

[0089]

[0090] Among them, w t-1 is the fusion data corresponding to the central monitoring station at the previous time t-1, are the set weights in the attention mechanism unit, b t is the attention score at the current time t, β t is the weight at the current moment t, and T is the set time step length;

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

[0092] The deep feature ft The fusion normalization unit in the input fusion unit makes the deep feature f t With shallow features u t Sum and normalize to get the fusion data w corresponding to the central monitoring station at the current time t t .

[0093] S2: Input the fusion data corresponding to each monitoring station into the fully connected interpretation layer in turn, loop through all monitoring stations in the entire area, and obtain the PM of each monitoring station at the future time. 2.5 Concentration prediction value;

[0094] S3: Based on the PM of each monitoring station at the future time 2.5 Concentration prediction and PM 2.5 The actual value of concentration is used to construct the root mean square error loss function;

[0095] S4: Determine whether the root mean square error loss function value is less than the set value. If so, the layered attention recurrent neural network model at this time is the final layered attention recurrent neural network model. If not, use the Adam back-propagation algorithm to adjust the parameters of the layered attention feature matching layer, the attention recurrent module, and the fully connected interpretation layer, and then use the adjusted layered attention feature matching layer, the attention recurrent module, and the fully connected interpretation layer to re-execute steps S1 to S4.

[0096] It can be seen that the present invention completes neural network training for the hierarchical attention recurrent neural network based on the cluster site data and the hierarchical attention recurrent neural network large model, divides the reprocessed cluster site data into a training set and a verification test set, selects the root mean square error (RMSE) as the loss function, and uses the Adam back propagation algorithm to train the hierarchical attention recurrent neural network large model and update the model parameters. Further, the model hyperparameters are adjusted using the verification test set data, and the prediction performance of the model is tested until the hierarchical attention recurrent network reaches the optimal prediction performance, the training is terminated, and the neural network model file and process file are saved.

[0097] In summary, the present invention provides a PM for cluster sites. 2.5The method for long-term concentration prediction first constructs a hierarchical MIC (Maximal Information Coefficient, MIC) selector, divides the monitoring sites into different distance levels with spatial distance as the discrimination index, and completes data screening from each level in turn; then the obtained screened cluster site data is reprocessed, and a position encoding vector is embedded in the monitoring site data in the form of a ring grid; then, according to the determined number of distance levels, a hierarchical attention feature matching neural network layer is constructed, and the hierarchical attention feature matching layer uses the attention mechanism to independently adjust the input weight of each monitoring site data; at the same time, the present invention constructs an attention recurrent network module based on GRU and feedback attention mechanism, and uses the attention mechanism based on the attention recurrent module to feedback adjust the extracted features, and finally obtains a hierarchical attention recurrent neural network large model with optimal prediction performance, which provides an effective solution to the long-term forecasting problem of multi-site air quality and can promote the technical development of related fields.

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

Claims

1. A PM for cluster sites 2.5 The method for long-term concentration prediction is characterized by: A hierarchical attention recurrent neural network model is used to predict the PM of all monitoring stations in the area at future times. 2.5 Each monitoring station in the cluster is used as the central monitoring station in turn, and the information of other surrounding monitoring stations is used as input to predict the PM concentration of the central monitoring station at the future time. 2.5 concentration, looping through all monitoring sites in the area to complete cluster prediction; wherein, the hierarchical attention recurrent neural network large model includes a hierarchical MIC selector, a hierarchical attention feature matching layer, an attention recurrent module and a fully connected interpretation layer; at the same time, the construction method of the hierarchical attention recurrent neural network large model is as follows: S1: All monitoring stations in the cluster are sequentially used as central monitoring stations to perform fusion data acquisition operations to obtain 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 layers of concentric circle areas with the central monitoring station as the center; S12: Use the layered MIC selector to obtain PM of the central monitoring site in each layer of concentric circle area. 2.5 The monitoring station with the largest concentration MIC measurement is used as the surrounding monitoring station, and the location codes of the surrounding monitoring stations in each layer are obtained at the same time; S13: A hierarchical attention feature matching layer is used to obtain the matching degree between the feature vectors of the surrounding monitoring sites at each layer and the feature vector of the central monitoring site at different times, where the feature vector includes the position code, PM 2.5 Concentration, PM 10 concentration, SO2 concentration, temperature, humidity, wind speed, wind direction, visibility; S14: taking the matching degree corresponding to each layer of surrounding monitoring stations at different times as the attention weight, weighting the feature vectors of each layer of surrounding monitoring stations at different times respectively, obtaining the matched surrounding monitoring station data of each layer of surrounding monitoring stations corresponding to the central monitoring station, and splicing 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: using the attention loop module to perform feature extraction and feature fusion on the cluster site data of the central monitoring site to obtain fused data corresponding to the central monitoring site; S2: Input the fusion data corresponding to each monitoring station into the fully connected interpretation layer in turn, loop through all monitoring stations in the entire area, and obtain the PM of each monitoring station at the future time. 2.5 Concentration prediction value; S3: Based on the PM of each monitoring station at the future time 2.5 Concentration prediction and PM 2.5 The actual value of concentration is used to construct the root mean square error loss function; S4: Determine whether the root mean square error loss function value is less than the set value. If so, the layered attention recurrent neural network model at this time is the final layered attention recurrent neural network model. If not, use the Adam back-propagation algorithm to adjust the parameters of the layered attention feature matching layer, the attention recurrent module, and the fully connected interpretation layer, and then use the adjusted layered attention feature matching layer, the attention recurrent module, and the fully connected interpretation layer to re-execute steps S1 to S4.

2. A PM for cluster sites as claimed in claim 1 2.5 The method for long-term concentration prediction is characterized by: The intervals between adjacent concentric circles in the multi-layer concentric circle area in step S11 are adjusted according to the actual situation of the cluster site.

3. A PM for cluster sites as claimed in claim 1 2.5 The method for long-term concentration prediction is characterized by: In step S12, the method for obtaining the location code of any surrounding monitoring site of the central monitoring site is as follows: According to the set range resolution akm and azimuth resolution Divide the multi-layer concentric circle area into grids and obtain grids, where r is the radius of the maximum relevant range of the central monitoring station, in km; Each grid is distinguished by a binary integer pair (i, j) consisting of a distance number i and an azimuth number j, where the grid (i, j) represents a distance range of (ia, (i+1)a] and an azimuth range of area; The location codes of the surrounding monitoring stations are determined according to the binary integer pairs (i, j) of the grids in which the surrounding monitoring stations fall. If the azimuth angle sequence number j is an even number, the location codes of the surrounding monitoring stations are: At this time, j = 2k, k is a natural number, and C is a set constant; If the azimuth angle number j is an odd number, the location code of the surrounding monitoring stations is: At this time, j=2k+1.

4. A PM for cluster sites as claimed in claim 1 2.5 The method for long-term concentration prediction is characterized by: In step S13, the method for obtaining the matching degree between the feature vectors of the surrounding monitoring sites of any layer and the feature vector of the central monitoring site is as follows: Get the feature vector of the current surrounding monitoring sites at the current time t At any time t * The characteristic vector of the central monitoring station The correlation matching value a t : in, W T All are set weights; According to the correlation matching value a within the set time period T before and after the current time t t Calculate the matching degree α between the feature vector of the surrounding monitoring station and the feature vector of the central monitoring station at the current time t t : Among them, the time variable τ=1,2,…,T.

5. A method for long-term prediction of PM2.5 concentration for cluster sites as claimed in claim 4, characterized in that: In step S14, the method for obtaining the matched surrounding monitoring site data of the surrounding monitoring sites corresponding to the central monitoring site of any layer is as follows: in, For any time t * When the current surrounding monitoring site s corresponds to the central monitoring site s * The data of surrounding monitoring stations have been matched.

6. A PM for cluster sites as claimed in claim 1 2.5 The method for long-term concentration prediction 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 fusion data corresponding to the central monitoring station is as follows: The cluster site data of the corresponding central monitoring site obtained by splicing the matched surrounding monitoring site data of each layer with the feature vector of the central monitoring site is input into the bidirectional GRU for feature extraction to obtain the shallow feature u t ; The shallow feature u t Input the attention mechanism unit for feature extraction to obtain intermediate features Among them, the intermediate features The extraction method is as follows: Among them, w t-1 is the fusion data corresponding to the central monitoring station at the previous time t-1, are the set weights in the attention mechanism unit, b t is the attention score at the current time t, β t is the weight at the current moment t, and T is the set time step length; The intermediate features Input the bidirectional GRU in the fusion unit for feature extraction to obtain the deep feature f t ; The deep feature f t The fusion normalization unit in the input fusion unit makes the deep feature f t With shallow features u t Sum and normalize to get the fusion data w corresponding to the central monitoring station at the current time t t .

Images