Space-time multi-site combined prediction method for ozone concentration

Abstract

The invention discloses an ozone concentration space-time multi-site joint prediction method, which comprises the following steps of: constructing a site space distance map, a site function logic map and an ozone correlation map, and extracting a space dependency relationship under different perspectives based on a constructed feature map by utilizing the capturing capability of a GCN (Generic Core Network) on space correlation features; and taking the extracted feature sequence and the multi-source influence factor as an input sequence of an ozone concentration prediction model, extracting global information causing ozone concentration change under multiple perspectives, and realizing prediction of the ozone concentration of the target station in different time periods in the future. According to the method, spatial relevance, logic similarity, ozone relevance and sequential multi-view experience knowledge information among stations are fully considered, the high efficiency of extracting the spatial relevance by the GCN model and the accuracy of predicting the sequence task of the Transform model are combined, and multi-view experience knowledge information causing ozone concentration change is fully considered; and the ozone concentration prediction quality of the station is effectively improved.

Description

Technical Field

[0001] The invention belongs to the technical field of air quality detection, and in particular relates to a spatiotemporal multi-site joint prediction method for ozone concentration. Background Art

[0002] Ozone (O3) is one of the important indicators for air quality monitoring. Changes in its concentration affect people's health, crop growth, and the ecological environment. However, the increase in ozone pollution and volatile organic compounds has resulted in the ozone prevention and control situation remaining severe. Monitoring and control of ozone pollution is a long-term task with complex interrelated issues. Summary of the invention

[0003] The purpose of the present invention is to provide a method for spatiotemporal multi-site joint prediction of ozone concentration, which fully considers the multi-perspective feature information of ozone concentration changes: the spatial correlation, logical similarity, ozone correlation and temporal characteristics of multi-source influencing factors between sites, which together serve as the knowledge information source for ozone concentration prediction. The Transformer model extracts the global information of the multi-perspective feature sequence to realize the spatiotemporal multi-site joint prediction of ozone concentration, effectively improves the accuracy and interpretability of ozone prediction, and effectively improves the prediction quality of ozone concentration.

[0004] The present invention is implemented by the following technical solutions:

[0005] A spatiotemporal multi-site joint prediction method for ozone concentration is proposed, including:

[0006] Construct spatial distance maps, functional logic maps, and correlation maps for multiple ozone concentration monitoring sites;

[0007] Taking the constructed spatial distance map, functional logic map and correlation map as input, the GCN model is used to capture the spatial correlation characteristics, logical similarity characteristics and correlation characteristics of ozone concentration changes;

[0008] The ozone concentration is predicted using the spatial correlation feature sequence, logical similarity feature sequence, ozone correlation feature sequence and the measured sequence of multi-source influencing factors at the target monitoring stations as the input of the Transformer model.

[0009] In some embodiments of the present invention, constructing a spatial distance map includes:

[0010] Each monitoring station is regarded as a node of the graph, and the spatial relationship between the monitoring stations is used as the edge to construct a spatial distance graph; the inverse of the Euclidean distance between the target monitoring station and the surrounding monitoring stations is used as the edge weight; the node characteristic data includes the measured data of the air quality index, air pollutant indicators, and meteorological parameters.

[0011] In some embodiments of the present invention, constructing a functional logic diagram includes:

[0012] Each monitoring site is taken as a node of the graph, and a functional logic graph is constructed with the logical similarities between monitoring sites as edges. If the functional types of the areas to which two monitoring sites belong are the same, they are connected by edges; if the functional types of the areas to which the two monitoring sites belong are different, there is no edge connection. The node characteristic data include the measured data of air quality index, air pollutant indicators, and meteorological parameters.

[0013] In some embodiments of the present invention, constructing an ozone correlation map includes:

[0014] Each monitoring station is regarded as a node of the graph, and a correlation graph is constructed with the dynamic correlation of ozone between monitoring stations as the edge. The correlation between two monitoring stations is used as the edge weight. The node characteristic data includes the measured data of air quality index, air pollutant indicators and meteorological parameters.

[0015] In some embodiments of the present invention, the GCN model is used to capture the spatial correlation characteristics, logical similarity characteristics and correlation characteristics of ozone concentration changes, including:

[0016] Initialize the feature representation of nodes in the spatial distance graph, functional logic graph, and correlation graph;

[0017] The graph convolution operation is performed iteratively, and the feature representation of the node is updated in each iteration. The information of each node is obtained by taking the weighted sum of the information of the previous layer node and the information of the adjacent nodes.

[0018] Perform linear changes and nonlinear transformations of activation functions;

[0019] The iteration is repeated until the feature representation of the node reaches a predetermined number of iterations.

[0020] In some embodiments of the present invention, the Transformer model predicts ozone concentration, including:

[0021] Vectorize the input sequence;

[0022] Use sine and cosine functions of different frequencies to encode position information;

[0023] The positional encoding is added to the input sequence vector as the input to the encoder of the Transformer model; the encoder is formed by stacking multiple identical layers, each of which contains a self-attention mechanism and a feedforward neural network;

[0024] The output of the encoder is used as the input of the decoder, and the decoder outputs the predicted value of ozone concentration; the decoder is formed by stacking multiple identical layers, each of which contains a self-attention mechanism, an interaction layer with the encoder, and a feedforward neural network.

[0025] In some embodiments of the present invention, the self-attention mechanism in the encoder and the decoder adopts a multi-head self-attention mechanism.

[0026] Compared with the prior art, the advantages and positive effects of the present invention are as follows: in the spatiotemporal multi-site joint prediction method for ozone concentration proposed in the present invention, in order to incorporate empirical information from more perspectives into the modeling analysis in ozone concentration prediction, the distance characteristics between sites, the functional category characteristics of the sites and the correlation of ozone between sites are considered to affect the change of ozone concentration, and a site spatial distance map, a site functional logic map and an ozone correlation map are constructed respectively; the GCN's ability to capture spatial correlation features is utilized to perform convolution operations based on graph structure data to extract spatial dependency features from different perspectives; the extracted feature sequence and multi-source influencing factors are used together as the input sequence of the ozone concentration prediction model Transformer, and the global information causing the change of ozone concentration from multiple perspectives is extracted through the Transformer model to realize the prediction of ozone concentration at different time periods in the future for the target site. The present invention fully considers the spatial correlation, logical similarity, ozone correlation and temporal characteristics of multi-source influencing factors among sites, combines the high efficiency of GCN model in extracting spatial correlation relationships and the accuracy of Transformer model sequence task prediction, and fully considers the multi-perspective feature information that causes changes in ozone concentration. It makes up for the shortcomings of most predictions based only on single-site time series data, and can effectively improve the quality of ozone concentration prediction.

[0027] Other features and advantages of the present invention will become more apparent after reading the detailed description of the embodiments of the present invention in conjunction with the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0029] Figure 1 This is a schematic diagram of the implementation of the spatiotemporal multi-site joint prediction method for ozone concentration proposed in the present invention;

[0030] Figure 2 This is a schematic diagram of the steps of the spatiotemporal multi-site joint prediction method for ozone concentration proposed in the present invention.

[0031] Figure 3 Schematic diagram of the spatial dependency feature extraction method under multiple viewing angles in the present invention;

[0032] Figure 4 FIG. 1 is a schematic diagram of the ozone concentration prediction method of the present invention. DETAILED DESCRIPTION

[0033] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0034] Changes in ozone concentration are affected by many factors, including temperature, humidity, radiation, and boundary layer height. In terms of spatial location, due to factors such as climate change, ozone transmission, and industrial emissions, the measured sequences at different monitoring points in the same city have different degrees of correlation with the historical data of the site and the monitoring data of adjacent or nearby sites: (1) The closer the site is geographically, the more similar the influencing factors are, and the stronger the correlation between the changes in ozone concentration is. (2) The functional category of the area where the monitoring site is located is also an important factor affecting ozone pollution to a large extent. Monitoring sites in tourist areas or industrially intensive areas, although geographically far apart, usually show great similarity in the characteristics of ozone pollution changes. Most existing ozone concentration prediction methods are based on single-site monitoring data for prediction, and some only consider single-variable historical data, resulting in low ozone prediction quality and large deviations in the prediction results of multiple sites between regions.

[0035] In view of the above analysis, when performing predictive modeling on ozone concentration, the present invention conducts modeling and analysis from multiple perspectives of the many influencing factors that cause changes in ozone concentration. While considering multi-source influencing factors such as the air quality index (AQI), air pollutant indicators, and meteorological parameters, it also takes into account the spatial dependence of neighboring or similar sites, providing a more comprehensive source of information for ozone concentration prediction and improving the accuracy and interpretability of the prediction.

[0036] In the spatiotemporal multi-site method for ozone concentration proposed in the present invention, a spatial distance graph is first constructed based on the geographical distance between the target monitoring site and other monitoring sites, a site function logic graph is constructed according to the functional division of the monitoring sites, and an ozone correlation graph is constructed according to the correlation of ozone concentrations at different monitoring sites. Next, the GCN (Graph Convolution Neural Network) model is used to extract the spatial correlation characteristics, logical similarity characteristics, and correlation characteristics of the monitoring sites based on the adjacency matrix of the three feature graphs, respectively, to capture more comprehensive empirical knowledge and provide feature information from more perspectives for ozone concentration. Finally, the extracted feature sequence is spliced ​​with the measured sequences such as the air quality index, air pollutant indicators, and meteorological parameters of the target monitoring site, and used together as the input sequence of the model. The Transformer model is used to extract the global information of the multi-perspective feature sequence to achieve the prediction of ozone concentration in the future period.

[0037] Specific, combined Figures 1 to 4 As shown, the spatiotemporal multi-site joint method of ozone concentration proposed by the present invention includes:

[0038] S1: Construct spatial distance maps, functional logic maps, and correlation maps for multiple ozone concentration monitoring sites.

[0039] (1) Construct a spatial distance map of ozone concentration monitoring stations.

[0040] Since the target monitoring station and its surrounding monitoring stations have similar geographical environment, meteorological conditions, industrial emissions and other factors, the information causing changes in ozone concentration contained in the historical monitoring data of the surrounding monitoring stations has a great correlation with the information needed to extract the ozone concentration of the target monitoring station. The extraction of relevant information from the surrounding monitoring stations can enhance the prediction performance of the ozone concentration at the target monitoring station.

[0041] In the present invention, a spatial distance graph of monitoring sites is constructed based on the distance between the target monitoring site and the surrounding monitoring sites: each monitoring site is taken as a node of the graph, and the node feature data includes measured data sequences such as air quality index, air pollutant index, and meteorological parameters. The spatial relationship between monitoring sites is represented by edges, and the inverse of the Euclidean distance between the target monitoring site and the surrounding monitoring sites is used as the edge weight of the graph.

[0042] The Euclidean distance between two points is expressed as:

[0043] ;

[0044] in Indicates the site and Site The distance between and Represent the latitude and longitude coordinates of the site respectively.

[0045] When the present invention measures the ozone concentration of a target monitoring station, it sorts the Euclidean distances calculated with the surrounding monitoring stations and selects the one with the smallest distance. The site spatial distance map is constructed for each site.

[0046] (2) Construct the functional logic diagram of the ozone concentration monitoring station.

[0047] The functional type of the area to which the target monitoring station belongs has a great impact on the air pollution of the target monitoring station. For example, the air pollution of monitoring stations in the same industrial park is relatively serious, while the air quality of monitoring stations in the same natural scenic area is relatively good. Although they are geographically far away, their air pollution conditions show great similarity.

[0048] The present invention first refers to the functional classification data of each monitoring station in the research city, and marks the functional category of the area to which each monitoring station belongs. When predicting the ozone concentration of the target monitoring station, a functional logic diagram of the target monitoring station and other monitoring stations is constructed: the monitoring station is used as a node of the diagram, and the node feature data includes measured data sequences such as air quality index, air pollutant index, and meteorological parameters. The logical similarity between the stations is represented by edges. If the functional types of the areas to which the two monitoring stations belong are the same, there is an edge connection. If the functional types of the areas to which the two monitoring stations belong are different, there is no edge connection.

[0049] (3) Construct ozone correlation maps for ozone concentration monitoring stations.

[0050] Affected by various factors such as ozone transmission and industrial emissions, the correlation of ozone pollution between monitoring sites changes dynamically. When predicting the ozone concentration at the target monitoring site, real-time extraction of ozone correlation can capture characteristic information from different perspectives.

[0051] In the ozone dynamic correlation graph, each monitoring station is used as a node of the graph. The node attribute data includes measured data sequences such as air quality index, air pollutant index, and meteorological parameters. The dynamic correlation between stations is represented by edges, and the correlation between two stations is used as the edge weight. For the same target monitoring station, the ozone correlation graph constructed at different prediction times is different and has dynamic time-varying properties.

[0052] The construction of the above-mentioned site spatial distance map, logical similarity feature map and ozone dynamic correlation map can provide more knowledge and experience information on the mechanism of air pollution from more perspectives for the ozone concentration prediction task of the target monitoring site, make up for the shortcomings of prediction based only on historical monitoring data of a single site, fully consider the spatial dependencies from multiple perspectives, and provide more information sources for site ozone concentration prediction.

[0053] S2: Taking the constructed spatial distance map, functional logic map and correlation map as input, the GCN model is used to capture the spatial correlation characteristics, logical similarity characteristics and ozone correlation characteristics of ozone concentration changes.

[0054] The present invention adopts the GCN (Graph Convolutional Network) model to extract spatial dependency features. The inputs are the constructed spatial distance graph, functional logic graph and ozone correlation graph respectively. The correlation between network nodes is extracted from the underlying data based on the three feature graphs, and the spatial correlation characteristics, logical similarity characteristics and ozone correlation characteristics of the ozone concentration changes between monitoring stations are effectively captured to obtain more comprehensive knowledge and experience information.

[0055] The change formula of the GCN model is:

[0056] ;

[0057] in, It is The feature matrix of the layer, is the input feature matrix, is the adjacency matrix of the undirected graph G with self-connection added, yes The degree matrix of It is The layer weight matrix, is the activation function.

[0058] The GCN model extracts the relationship and feature information between nodes through graph convolution operations. Its core idea is to update the representation of the central node by using the features of neighboring nodes. The specific operations are: (1) Initialize the feature representation of nodes in the spatial distance graph, functional logic graph, and ozone correlation graph; usually a node feature matrix; (2) Iterate the graph convolution operation, and update the feature representation of the node in each iteration. The information of each node is obtained by the weighted sum of the information of the previous layer node itself and the information of the adjacent nodes, and then undergoes a nonlinear transformation of the linear change W and the activation function σ; (3) Repeat multiple iterations until the feature representation of the node reaches a stable state or reaches a predetermined number of iterations.

[0059] The GCN model gradually integrates the global graph structure information into the feature representation of the node by iteratively aggregating the information of neighboring nodes. In the present invention, each feature graph extracts the spatial dependency features between nodes through a multi-layer GCN network, captures the local and global information between nodes, and thus learns the spatial dependency information of ozone concentration changes at different monitoring stations from three different perspectives: spatial correlation, logical similarity, and ozone correlation. The empirical knowledge of ozone pollution is quantified as the input of the prediction model, providing more comprehensive domain knowledge and enhancing the prediction ability.

[0060] S3: Predict ozone concentration using spatial correlation feature sequence, logical similarity feature sequence, ozone correlation feature sequence and the measured sequence of multi-source influencing factors at the target monitoring station as input to the Transformer model.

[0061] In the present invention, the prediction of ozone concentration adopts the Transformer architecture, which is a deep learning model for processing sequence data. It captures the dependencies between positions in the sequence through the self-attention mechanism and consists of two major parts: an encoder and a decoder. The encoder processes the input sequence and converts the input data into an abstract representation. Each encoding layer contains two main parts: a self-attention mechanism and a feedforward neural network. The output of the encoder is a context-dependent representation that captures the semantics of each word in the input sequence. The task of the decoder is to generate the target sequence. Each decoder layer includes three main parts: a self-attention mechanism, an encoder-decoder attention, and a feedforward neural network.

[0062] In the present invention, the input of Transformer includes feature sequences from four perspectives, namely: spatial correlation feature sequence S, logical similarity feature sequence L, ozone correlation feature sequence R and measured sequence of multi-source influencing factors of target monitoring stations (measured data sequences such as air quality index, air pollutant indicators, meteorological parameters, etc.), and the output is the predicted value of ozone concentration at different time periods in the future for the target monitoring station.

[0063] Transformer consists of an encoder and decoder with a multi-layer structure, and each layer includes a multi-head attention and a feedforward neural network sublayer. The input of the encoder is the concatenated multi-source view feature sequence. Embedding and position encoding are two key components in the Transformer model. Embedding is a vector representation of the input sequence, and then considering the correlation of sequence data in space, the position encoding is added to the embedded input. Use sine and cosine functions of different frequencies to encode the position information:

[0064] ;

[0065] ;

[0066] in, , position encoding .

[0067] The positional encoding is added to the sequence vector as input to the encoder.

[0068] The multi-head self-attention mechanism splits a single self-attention mechanism into multiple subspaces, and performs self-attention on each subspace to learn different attention weights, so as to better capture features at different levels and fuse them in the representation space. The self-attention mechanism is described as:

[0069] ;

[0070] in , and Represents query, key, and value respectively.

[0071] Multi-head self-attention concatenates the vector representations formed by each head combined with the context representation and multiplies them by the weight matrix To aggregate the information on different heads and form the final vector representation combined with the context:

[0072] ;

[0073] ;

[0074] in , Indicates the number of attention heads; , , are the weight matrices for query, key, and value, respectively. A weight matrix representing a linear transformation.

[0075] In the present invention, all inputs in the decoder are historical data of feature sequences without future information, so the Transformer architecture does not use the masked attention mechanism.

[0076] The output vector finally generated by the decoder is converted into the final desired output sequence through the fully connected layer. In the present invention, if the ozone concentration is to be predicted, the loss function is used in the fully connected layer to compare with the ozone concentration value, so that the output gradually reduces the difference with the concentration, and the predicted value of the ozone concentration is obtained when it reaches the preset accuracy. The loss function used is the mean square error (MSE), which can be expressed as:

[0077] ;

[0078] Where n represents the number of samples, is the measured value of ozone concentration, is the predicted value of ozone concentration.

[0079] The spatiotemporal multi-site joint prediction method for ozone concentration given by the present invention first constructs a spatial distance map, a functional logic map, and an ozone correlation map based on the geographical distance, functional attributes, and ozone correlation between the sites, and then uses GCN to extract empirical knowledge information from different perspectives based on the three constructed feature maps. On this basis, the extracted spatial correlation feature sequence, logical similarity feature sequence, ozone correlation sequence, and the measured sequence of multi-source influencing factors of the target monitoring site are spliced ​​together as the input sequence of the ozone concentration prediction model. The Transformer model extracts the global information that causes the change of ozone concentration from multiple perspectives, so as to realize the prediction of ozone concentration at the target monitoring site in different future time periods.

[0080] The following is an example verification of the proposed method:

[0081] (1) Data source: From May 1, 2014 to February 19, 2024, 18 air quality monitoring stations in a certain city, 85,968 monitoring records at each station, each record including air volume index (AQI), air quality evaluation parameters (O3, CO, NO2, SO2, PM 2.5 and PM 10 ) and meteorological parameters (temperature, dew point temperature, wind speed, wind direction, and air pressure). The monitoring records from May 1, 2014 to March 12, 2021 are used as the training data set; the monitoring records from March 13, 2021 to March 5, 2022 are used as the validation data set; and the monitoring records from March 6, 2022 to February 19, 2024 are used as the test data set. At present, there are 35 air quality monitoring stations in the city, including 12 in the urban area, 5 in the northwest, 8 in the northeast, 6 in the southeast, and 4 in the southwest. From the collected historical data, it can be seen that the distribution of stations was updated around January 19, 2021. In order to ensure the time span of the verification data, the stations were screened based on the latitude and longitude and names of the new and old stations, and the stations that existed in both station lists were retained, or the geographically adjacent stations were merged. Therefore, in this example, the experimental data came from 18 stations.

[0082] In this example, the geographical location of the monitoring station is determined by the longitude and latitude of the station. The distance between two points is calculated according to the Euclidean distance formula. The inverse of the Euclidean distance between the target monitoring station and the surrounding monitoring stations is used as the edge weight of the graph to construct the station spatial distance graph.

[0083] In this example, the functional category of each monitoring station is determined according to the logic of air quality monitoring point layout in a certain city. The city's air quality automatic monitoring system consists of 35 monitoring points, including 23 urban environmental evaluation points to evaluate the average status and changing laws of air quality in urban environments, 1 urban clean control point to reflect the background level of air quality in urban areas that are not affected by local urban pollution, 6 regional background transmission points to characterize the regional environmental background level and reflect the transmission of pollution in the region, and 5 traffic pollution monitoring points to monitor the impact of road traffic pollution sources on ambient air quality. When constructing the site function logic diagram, if the functional type of the area to which the two sites belong is the same, there is an edge connection, and if the functional type of the area to which the two sites belong is different, there is no edge connection.

[0084] The ozone correlation between stations is calculated at each prediction time, and an ozone correlation map is constructed. Therefore, for the same target station, the ozone correlation map constructed at different prediction times is different and has dynamic time-varying properties.

[0085] (2) Spatiotemporal multi-site joint prediction of ozone concentration: Based on the method proposed in the present invention, a two-layer GCN network is used to extract spatial correlation features, logical similarity features and ozone correlation features from the three constructed feature maps. The three feature sequences obtained are spliced ​​with the measured sequences of multi-source influencing factors of the target site and used as the input sequence of the ozone concentration prediction model Transformer. The Transformer model extracts global information based on the multi-perspective feature sequence and predicts the ozone concentration of the site. The input of the Transformer model is the three feature sequences extracted by GCN and the air quality index (AQI) of the target site, air pollutant indicators (O3, CO, NO2, SO2, PM 2.5 、PM 10 ) and meteorological parameters (temperature, dew point temperature, wind speed, wind direction, and air pressure). The output is the ozone concentration at the target site at different time periods in the future. The number of layers of the encoder and decoder of the model are both 2, the number of attention heads is 4, the encoder input dimension is 512, and the output dimension is 1024. The decoder input dimension is 1024, and the output dimension is 1, that is, the output of the fully connected layer in the decoder is the ozone concentration value to be predicted. The loss function MSE is used to measure the gap between the predicted value and the measured value of the ozone concentration. During the model training process, the gap between the two is gradually narrowed through parameter optimization to improve the prediction accuracy.

[0086] Taking one of the monitoring stations as the target station, the ozone concentration prediction results for the next 8 hours are compared and analyzed as an example. The 8-hour maximum 8-hour average concentration of ozone (MDA8) is a commonly used indicator for evaluating ozone pollution and management. Therefore, the ozone concentration prediction results for the next 8 hours are used to evaluate the prediction performance of each model.

[0087] The prediction models compared in this example include LSTM, BiLSTM, GRU, CNN, and Transformer, and the comparison index used is R 2 (coefficient of determination), MAE (mean absolute error), RMSE (root mean square error) and MAPE (mean absolute percentage error), and the comparison results are shown in Table 1.

[0088] Table 1 Comparison of prediction results of the method proposed in this invention and other models

[0089] The R of the method proposed by the present invention 2 , MAE, RMSE and MAPE are 0.9704, 5.8935 respectively ,8.6774 and 0.2617, the prediction index R based only on the Transformer model 2 , MAE, RMSE and MAPE are 0.9651, 6.3283 respectively ,9.4269 and 0.284. The four indicators of the method proposed in the present invention are all better than those of the Transformer model, indicating that the multi-perspective feature information constructed in the present invention can provide more comprehensive empirical knowledge information in ozone concentration prediction and effectively improve the ozone concentration prediction performance.

[0090] The proposed method is compared with the other five models: R 2 It was improved by 0.5%-3%, and the other three indicators MAE, RMSE and MAPE were greatly improved, indicating that the prediction performance of the spatiotemporal multi-site joint prediction method of ozone concentration proposed in the present invention is better than that of the other five models.

[0091] This example is a verification example of the proposed method, which can be extended to the prediction of air pollutant concentrations in a region or even nationwide.

[0092] There are many implementation methods of the present invention, and all technical solutions formed by equivalent transformation or equivalent transformation fall within the protection scope of the present invention.

[0093] It should be pointed out that the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by ordinary technicians in this technical field within the essential scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A spatiotemporal multi-site joint prediction method for ozone concentration, characterized in that: include: Construct spatial distance maps, functional logic maps, and correlation maps for multiple ozone concentration monitoring sites; Taking the constructed spatial distance map, functional logic map and correlation map as input, the GCN model is used to capture the spatial correlation characteristics, logical similarity characteristics and correlation characteristics of ozone concentration changes; The ozone concentration is predicted using the spatial correlation feature sequence, logical similarity feature sequence, ozone correlation feature sequence and the measured sequence of multi-source influencing factors at the target monitoring stations as the input of the Transformer model.

2. The spatiotemporal multi-site joint prediction method for ozone concentration according to claim 1 is characterized in that: The GCN model is used to capture the spatial correlation characteristics, logical similarity characteristics and correlation characteristics of ozone concentration changes, including: Initialize the feature representation of nodes in the spatial distance graph, functional logic graph, and correlation graph; The graph convolution operation is performed iteratively, and the feature representation of the node is updated in each iteration. The information of each node is obtained by taking the weighted sum of the information of the previous layer node and the information of the adjacent nodes. Perform linear changes and nonlinear transformations of activation functions; The iteration is repeated until the feature representation of the node reaches a predetermined number of iterations.

3. The spatiotemporal multi-site joint prediction method for ozone concentration according to claim 1 is characterized in that: The Transformer model predicts ozone concentrations, including: Vectorize the input sequence; Use sine and cosine functions of different frequencies to encode position information; The positional encoding is added to the input sequence vector as the input to the encoder of the Transformer model; the encoder is formed by stacking multiple identical layers, each of which contains a self-attention mechanism and a feedforward neural network; The output of the encoder is used as the input of the decoder, and the decoder outputs the predicted value of ozone concentration; the decoder is formed by stacking multiple identical layers, each of which contains a self-attention mechanism, an interaction layer with the encoder, and a feedforward neural network.

4. The spatiotemporal multi-site joint prediction method for ozone concentration according to claim 3 is characterized in that: The self-attention mechanism in the encoder and decoder adopts a multi-head self-attention mechanism.

5. The spatiotemporal multi-site joint prediction method for ozone concentration according to claim 1 is characterized in that: Constructing a spatial distance map includes: Each monitoring station is regarded as a node of the graph, and the spatial relationship between the monitoring stations is used as the edge to construct a spatial distance graph; the inverse of the Euclidean distance between the target monitoring station and the surrounding monitoring stations is used as the edge weight; the node characteristic data includes the measured data of the air quality index, air pollutant indicators, and meteorological parameters.

6. The spatiotemporal multi-site joint prediction method for ozone concentration according to claim 1 is characterized in that: Building a functional logic diagram includes: Each monitoring site is taken as a node of the graph, and a functional logic graph is constructed with the logical similarities between monitoring sites as edges. If the functional types of the areas to which two monitoring sites belong are the same, they are connected by edges; if the functional types of the areas to which the two monitoring sites belong are different, there is no edge connection. The node characteristic data include the measured data of air quality index, air pollutant indicators, and meteorological parameters.

7. The spatiotemporal multi-site joint prediction method for ozone concentration according to claim 1 is characterized in that: Building a dependency graph involves: Each monitoring station is regarded as a node of the graph, and a correlation graph is constructed with the dynamic correlation of ozone between monitoring stations as the edge. The correlation between two monitoring stations is used as the edge weight. The node characteristic data includes the measured data of air quality index, air pollutant indicators and meteorological parameters.

Images