Air pollutant concentration prediction method, device, equipment and medium

Abstract

The invention discloses an air pollutant concentration prediction method and device, equipment and a medium, and relates to the technical field of deep learning, and the method comprises the steps: obtaining site monitoring data and site longitude and latitude, carrying out the missing value filling and normalization processing of the site monitoring data, obtaining a time sequence, and constructing a geographic relation graph; performing trend feature extraction and feature processing on the time sequence by using an exponential moving average method and based on a gating circulation unit to obtain trend component features; performing periodic feature extraction on the time sequence; performing batch normalization processing, function activation and vector splicing on the noise sequence to obtain noise component features; splicing and fusing the trend component feature, the periodic component feature and the noise component feature to obtain a first prediction feature; solving and coding the time sequence to obtain a second prediction feature; and performing gating fusion and graph attention fusion on the first prediction feature and the second prediction feature based on a geographic relation graph to obtain an air pollutant concentration prediction result.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, and in particular to a method, apparatus, equipment and medium for predicting air pollutant concentrations. Background Technology

[0002] Currently, deep learning models that embed physical knowledge lack explicit decomposition of historical time series data. Using a single model to simultaneously characterize the trend component, periodic component, and noise component in historical time series data has the following shortcomings: (1) Severe information coupling: The trend component, periodic component, and noise component overlap and interfere with each other, making it difficult for the model to distinguish the change signals from different sources; (2) Limited prediction accuracy: The model must take into account the features of different categories within the same framework, which easily leads to overfitting of the periodic component or underfitting of the trend component, thereby reducing the overall prediction accuracy; (3) Insufficient robustness: When faced with sudden emissions, extreme weather, or missing observation data, the model lacks an independent decomposition mechanism to isolate abnormal disturbances, affecting stability and generalization ability.

[0003] As can be seen from the above, how to solve the problem of insufficient extraction of discontinuous state information by traditional gating loop mechanisms, improve the accuracy, robustness and interpretability of air pollutant concentration prediction, and improve the accuracy of long-term time series modeling are problems to be solved in this field. Summary of the Invention

[0004] In view of this, the purpose of this invention is to provide a method, apparatus, device, and medium for predicting air pollutant concentrations, which can solve the problem of insufficient extraction of discontinuous state information by traditional gating loop mechanisms, improve the accuracy, robustness, and interpretability of air pollutant concentration prediction, and improve the accuracy of long-term time series modeling. The specific solution is as follows: In a first aspect, this application discloses a method for predicting air pollutant concentrations, including: Acquire site monitoring data and site latitude and longitude, perform missing value imputation and normalization on the site monitoring data to obtain a time series, and construct a geographic relationship map of the sites based on the site latitude and longitude; The time series is subjected to trend feature extraction and feature processing using the exponential moving average method and based on a gated cyclic unit to obtain trend component features; The time series is subjected to periodic feature extraction using an attention mechanism to obtain periodic component features; Based on the time series, the trend component features, and the periodic component features, a noise series is determined. The noise series is then subjected to batch normalization, function activation, and vector concatenation to obtain the noise component features. The trend component feature, the periodic component feature, and the noise component feature are spliced ​​and fused to obtain the first prediction feature; The time series is solved using the advection diffusion equation and based on the neuron ordinary differential equation, and the solution is encoded to obtain the second predictive feature; Based on the geographic relationship map, gating fusion and graph attention fusion are performed on the first prediction feature and the second prediction feature to obtain the air pollutant concentration prediction result.

[0005] Optionally, the steps of acquiring site monitoring data and site latitude and longitude, performing missing value imputation and normalization on the site monitoring data to obtain a time series, and constructing a geographic relationship map of the sites based on the site latitude and longitude include: Acquire site monitoring data and site latitude and longitude; the site monitoring data includes particulate matter concentration, temperature, humidity, pressure, wind direction, and wind speed; Missing values ​​in the site monitoring data are filled by interpolating adjacent means, and the filled site monitoring data is then normalized to obtain a time series. The half-sine distance between stations is calculated based on their latitude and longitude to construct a geographic relationship map.

[0006] Optionally, the step of using the exponential moving average method and based on a gated cyclic unit to extract and process trend features from the time series to obtain trend component features includes: The time series was modeled using the exponential moving average method, and trend features were extracted to obtain the trend components. The trend component is input into a gated loop unit for feature processing to obtain the trend component feature.

[0007] Optionally, the step of using an attention mechanism to extract periodic features from the time series to obtain periodic component features includes: An attention mechanism is used, and an attention score is calculated based on the time series. The attention score is concatenated with the time series, and the concatenated data is normalized using a multilayer perceptron to calculate the periodic component features.

[0008] Optionally, the step of determining the noise sequence based on the time series, the trend component features, and the periodic component features, and then performing batch normalization, function activation, and vector concatenation on the noise sequence to obtain the noise component features includes: Calculate the residual sequence between the time series, the trend component feature, and the periodic component feature, and use the residual sequence as a noise sequence; The noise sequence is batch normalized using multi-scale convolution, activated using the GeLU function, and then vector concatenation and residual concatenation are performed based on the noise sequence to obtain noise component features.

[0009] Optionally, the step of using the advection-diffusion equation and solving the time series based on the neural ordinary differential equation, and encoding the solution results to obtain the second predicted feature, includes: Define the advection-diffusion equation, diffusion diagram, and advection diagram; Using the advection-diffusion equation, the diffusion map, and the advection map, and solving the time series based on the neural ordinary differential equation, the initial prediction results are obtained; The initial prediction result is encoded using a gated loop unit to generate a second prediction feature that conforms to physical laws.

[0010] Optionally, the step of performing gating fusion and graph attention fusion on the first prediction feature and the second prediction feature based on the geographic relationship graph to obtain the air pollutant concentration prediction result includes: The first and second predicted features are concatenated and fused using a linear layer to obtain the initial fused features; The gate coefficients are determined using a softmax layer, and the initial fused features are then subjected to gated fusion based on the gate coefficients to obtain the fused features. Based on the geographic relationship graph, graph attention fusion is performed on the fusion features to obtain the target graph attention fusion features; The air pollutant concentration prediction result is obtained by decoding the attention fusion features of the target map using a decoder.

[0011] Secondly, this application discloses an air pollutant concentration prediction device, comprising: The data processing module is used to acquire site monitoring data and site latitude and longitude, perform missing value imputation and normalization processing on the site monitoring data to obtain a time series, and construct a geographic relationship map of the site based on the site latitude and longitude. The geographic relationship graph processing module is used to extract and process trend features of the time series using the exponential moving average method and based on a gated loop unit to obtain trend component features. The periodic feature extraction module is used to extract periodic features from the time series using an attention mechanism to obtain periodic component features. The noise sequence processing module is used to determine the noise sequence based on the time series, the trend component features, and the periodic component features, and to perform batch normalization, function activation, and vector concatenation on the noise sequence to obtain the noise component features. The first prediction feature determination module is used to splice and fuse the trend component feature, the periodic component feature, and the noise component feature to obtain the first prediction feature; The second predictive feature determination module is used to solve the time series using the advection diffusion equation and based on the constant differential equation, and to encode the solution results to obtain the second predictive feature. The fusion module is used to perform gating fusion and graph attention fusion on the first prediction feature and the second prediction feature based on the geographic relationship graph to obtain the air pollutant concentration prediction result.

[0012] Thirdly, this application discloses an electronic device, including: Memory, used to store computer programs; A processor is used to execute the computer program to implement the aforementioned method for predicting air pollutant concentrations.

[0013] Fourthly, this application discloses a computer storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the steps of the aforementioned disclosed air pollutant concentration prediction method.

[0014] As can be seen, this application provides a method for predicting air pollutant concentrations, including acquiring station monitoring data and station latitude and longitude, performing missing value imputation and normalization on the station monitoring data to obtain a time series, and constructing a geographic relationship map of the stations based on the station latitude and longitude; using the exponential moving average method and based on a gated cyclic unit to extract and process trend features of the time series to obtain trend component features; using an attention mechanism to extract periodic features of the time series to obtain periodic component features; determining a noise sequence based on the time series, the trend component features, and the periodic component features, performing batch normalization, function activation, and vector concatenation on the noise sequence to obtain noise component features; concatenating and fusing the trend component features, periodic component features, and noise component features to obtain a first prediction feature; using the advection diffusion equation and based on the neural network ordinary differential equation to solve the time series, and encoding the solution to obtain a second prediction feature; and based on the geographic relationship map, performing gated fusion and graph attention fusion on the first prediction feature and the second prediction feature to obtain the air pollutant concentration prediction result. This application performs missing value imputation and normalization on site monitoring data, avoiding problems such as gradient vanishing or numerical instability. It constructs a geographic relationship map of the sites based on their latitude and longitude, and uses the exponential moving average method with gated recurrent units to extract and process trend features from time series data. This addresses the problem of insufficient extraction of discontinuous state information by traditional gated recurrent mechanisms, improving the accuracy of long-term time series modeling. It also utilizes an attention mechanism to extract periodic features from the time series, identifying noise sequences based on time series, trend component features, and periodic component features. The noise sequences undergo batch normalization, function activation, and vector concatenation. Finally, the trend component features, periodic component features, and noise component features are concatenated and fused to obtain the first predictive feature, thus solving the problems of existing technologies. The lack of explicit decomposition of data patterns during the procedure makes the learning of data features more accurate, improving the robustness and interpretability of the prediction results. By using the advection diffusion equation and solving the time series based on the neural ordinary differential equation, and encoding the solution results, a second prediction feature is obtained. This combines the ability of deep learning models to learn data distribution with the ability of physical perception models to constrain objective physical laws. Based on the geographic relationship map, the first and second prediction features are gated and fused with graph attention to obtain the air pollutant concentration prediction results, improving the accuracy of air pollutant concentration prediction. The prediction results not only conform to the physical laws of the objective world, but also can quickly predict results within seconds, which has reference significance for emergency response to some public emergencies, such as industrial leaks. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0016] Figure 1 This is a flowchart of an air pollutant concentration prediction method disclosed in this application; Figure 2 This is an example diagram of a periodic alignment disclosed in this application; Figure 3 This is a schematic diagram of noise component processing disclosed in this application; Figure 4 This is a structural diagram of a decoder disclosed in this application; Figure 5 This is a structural diagram of a physical perception temporal decomposition deep learning model disclosed in this application; Figure 6 This is a schematic diagram of the structure of an air pollutant concentration prediction device disclosed in this application; Figure 7 This application provides a structural diagram of an electronic device. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Currently, deep learning models that embed physical knowledge lack explicit decomposition of historical time-series data. Using a single model to simultaneously characterize the trend, periodic, and noise components in historical time-series data has the following shortcomings: (1) Severe information coupling: Trend, periodic, and noise components overlap and interfere with each other, making it difficult for the model to distinguish changes from different sources; (2) Limited prediction accuracy: The model must take into account the features of different categories within the same framework, which easily leads to overfitting of periodic components or underfitting of trend components, thereby reducing the overall prediction accuracy; (3) Insufficient robustness: When facing sudden emissions, extreme weather, or missing observation data, the model lacks an independent decomposition mechanism to isolate abnormal disturbances, affecting stability and generalization ability. As can be seen from the above, how to solve the problem of insufficient extraction of discontinuous state information by the traditional gated loop mechanism, improve the accuracy, robustness, and interpretability of air pollutant concentration prediction, and improve the accuracy of long-term time-series modeling are problems to be solved in this field.

[0019] See Figure 1 As shown in the figure, an embodiment of the present invention discloses a method for predicting air pollutant concentrations, which may specifically include: Step S11: Obtain site monitoring data and site latitude and longitude, perform missing value imputation and normalization on the site monitoring data to obtain a time series, and construct a geographic relationship map of the site based on the site latitude and longitude.

[0020] In this embodiment, station monitoring data and station latitude and longitude are acquired; the station monitoring data includes particulate matter concentration, temperature, humidity, pressure, wind direction, and wind speed; missing values ​​are filled by interpolating adjacent mean values ​​of the station monitoring data, and the filled station monitoring data is normalized to obtain a time series; the half-sine distance between stations is calculated based on the station latitude and longitude to construct a geographic relationship map.

[0021] The air pollutant concentration prediction method proposed in this application can be implemented using the PyTorch framework and deployed in a GPU (Graphics Processing Unit) / CPU (Central Processing Unit) server environment. First, it acquires monitoring data from various monitoring stations; the number of monitoring stations is... N ,include N Monitoring data from each site ,in, T The length of the historical time series data used for prediction (with a time step of 3 hours, using historical 72-hour data to predict the concentration of the next 72 hours, therefore...) T =24), DThe characteristic dimensions (which include PM2.5 concentration, temperature, humidity, pressure, wind direction, and wind speed) D =6).

[0022] Furthermore, the latitude and longitude of the stations are obtained, and the half-sine distance between the stations is calculated using the latitude and longitude to construct a geographic relationship map. G This is an npz file (a file storage format) that stores the adjacency matrix A and the edge attribute matrix Edge. Matrix A uses 300km as a threshold; if the station... i , j If the half-sine distance between stations is less than the threshold of 300 km, then the stations are considered to be... i , j There are potential transmission paths (i.e.) The edge attribute matrix Edge is a matrix of size (E, 2), where E represents the total number of potential transmission paths. Each row stores two values: the half-sine distance of the current edge and the azimuth of the edge.

[0023] In this embodiment, the station monitoring data (PM2.5 concentration, temperature, humidity, pressure, wind direction, and wind speed) are interpolated by adjacent mean to fill in missing values, and then normalized to avoid gradient vanishing or numerical instability during model training.

[0024] Step S12: Using the exponential moving average method and based on a gated cyclic unit, the time series is subjected to trend feature extraction and feature processing to obtain trend component features.

[0025] In this embodiment, the time series is modeled using the exponential moving average method and trend features are extracted to obtain trend components; the trend components are then input into a gated loop unit for feature processing to obtain trend component features.

[0026] First, the time series... X To perform exponential moving average modeling, we will... X By time step T Written The form of trend components It can be recursively defined by the following formula: ; in, This is the smoothing coefficient.

[0027] The extracted trend components are then input into TAGRU (Top-k Attention Gated Recurrent Unit) for feature processing. For ease of explanation, we will take a time step... tIts gate control unit is defined as: ; ; ; ; ; in, It is a hidden state table that stores... t All hidden states before time (i.e., from) arrive Top-k Attention is a discrete attention mechanism, which... As a query vector, with Attention scores are calculated using key-value pairs, and the top-scoring pairs are ranked... k The hidden attention state is obtained by normalizing and weighting the high scores. ; To update the gate, use the current input. With hidden attention state Multiply by different transformation matrices and After adding, add the bias. It is obtained by activation through a sigmoid function (S-shaped growth curve); Forget gate, based on the input at the current moment. Compared to the previous hidden state Multiply by different transformation matrices and After adding, add the bias. It is obtained by activating a sigmoid function; The candidate hidden state is determined by the input at the current time step. Hiding attention state With the Gate of Oblivion And the previous hidden state The element-wise products are multiplied by different transformation matrices, summed, and then the bias is added. It is obtained by activating a tanh function (hyperbolic tangent function); The hidden state output at the current moment is determined by... and According to the updated door The result is obtained by element-wise multiplication. The entire process described above only yields the time step. t The hidden state output requires looping through this process. T Only then can the trend component characteristics be obtained. .

[0028] TAGRU is a variant of GRU (Gated Recurrent Unit), with the main changes being: (1) A discrete attention mechanism is introduced to capture long-term sequence features; (2) The long-term sequence features captured in (1) were added to the update gate; (3) Add the long-term sequence features captured in (1) to the candidate hidden states; The specific formula is as follows: ; This formula describes the process by which the discrete attention mechanism captures features from long-term sequences. It is a hidden state table that stores... t All hidden states prior to that time. As the current input state, As a query vector, with Attention scores are calculated using key-value pairs, and the top-scoring pairs are ranked... k The hidden attention state is obtained by normalizing and weighting the high scores. ; ; This formula describes the update gate. The solution process starts from the input at the current moment. With hidden attention state Multiply by different transformation matrices and After adding, add the bias. It is obtained by activating a sigmoid function; ; This formula describes the forgetting gate. The solution process starts from the input at the current moment. Compared to the previous hidden state Multiply by different transformation matrices and After adding, add the bias. It is obtained by activating a sigmoid function; ; This formula describes the candidate hidden state. The solution process starts from the input at the current moment. Hiding attention state With the Gate of Oblivion And the previous hidden state The element-wise products are multiplied by different transformation matrices, summed, and then the bias is added. It is obtained by activation through a tanh function; ; This formula describes the final output. The solution process is as follows: and According to the updated door It is obtained by multiplying each element.

[0029] Step S13: Use an attention mechanism to extract periodic features from the time series to obtain periodic component features.

[0030] In this embodiment, an attention mechanism is used, and an attention score is calculated based on the time series. The attention score is then concatenated with the time series, and the concatenated data is normalized using a multilayer perceptron to calculate the periodic component features.

[0031] To capture the periodic relationships present in the input sequence, this application designs MPAA (Multi-Granularity Periodic Alignment Attention). For ease of explanation, let's take a time step... t Let time granularity set P It stores several different time granularities. For example, to consider the cyclical changes at the same time every day, we use 8 as the time granularity (because one time step is 3 hours, and a time granularity of 8 is exactly 24 hours). For each time granularity... Both can yield a set of aligned historical fragments. Example of periodic alignment: Figure 2 As shown, the arrow indicates the current time in the example.

[0032] Input at the current moment As a query vector, different granularities Calculate attention scores using key-value pairs respectively. Then and The obtained query vectors are concatenated and fed into an MLP (Multi-layer Perceptron), and then normalized using softmax (normalized exponential function) to obtain each attention score. weight Finally, element-wise multiplication of the two yields the periodic characteristics at the current moment. This process is repeated. T The periodic component characteristics can be obtained in one step. .

[0033] The calculation formula and symbols for MPAA in this application are explained as follows: ; in, t For the current time step, p For a specific periodic granularity, such as 24 hours or 168 hours (one week), To step at the current time t For periodic granularity p The calculated periodic attention representation, qt In time step t The query vector, For the corresponding periodic granularity p The key vector matrix of historical moments, For the corresponding periodic granularity p The value vector matrix of historical moments, for transpose, The dimension of the key vector; ; in, In time step t The model is of periodic granularity. p The dynamically assigned importance weights, with canat[] representing the concatenation operation; ; in, In time step t Ultimately, it integrates a comprehensive representation that incorporates all periodic information. P For the set of all preset periodic granularities, For sets P All periodic granularities p The calculation results are summed.

[0034] Step S14: Based on the time series, the trend component features, and the periodic component features, a noise sequence is determined. The noise sequence is then subjected to batch normalization, function activation, and vector concatenation to obtain the noise component features.

[0035] In this embodiment, the residual sequence between the time series, the trend component features, and the periodic component features is calculated, and the residual sequence is used as a noise sequence. The noise sequence is batch normalized using multi-scale convolution, and the GeLU function is used for function activation. Based on the noise sequence, vector concatenation and residual concatenation are performed to obtain the noise component features.

[0036] Specifically, using time series XSubtract trend component features and periodic component characteristics The resulting residual sequence is the noise sequence. This application introduces a multi-scale convolutional module to process the noise sequence. The network structure is similar to the Inception structure, and the noise component processing diagram is shown below. Figure 3 As shown.

[0037] Here, conv1, conv3, and conv5 represent one-dimensional convolutions with kernel sizes of 1, 3, and 5, respectively. The noise sequence is processed through three parallel paths, followed by Batch Normalization, then activation via the GeLU function. Finally, the three paths are concatenated into vectors and residually connected to the original noise sequence to obtain the noise component features. .

[0038] Step S15: The trend component feature, the periodic component feature, and the noise component feature are spliced ​​and fused to obtain the first prediction feature.

[0039] In this embodiment, after obtaining the trend component characteristics Periodic component characteristics and noise component characteristics Then, these three parts are spliced ​​and fused to obtain the first predictive feature that conforms to the distribution of historical time series data, denoted as... .

[0040] Step S16: Using the advection diffusion equation and based on the constant differential equation, the time series is solved, and the solution is encoded to obtain the second prediction feature.

[0041] In this embodiment, an advection-diffusion equation, a diffusion map, and an advection map are defined; the time series is solved using the advection-diffusion equation, the diffusion map, and the advection map, based on the normal differential equation of the nervous system, to obtain an initial prediction result; the initial prediction result is encoded using a gated loop unit to generate a second prediction feature of prediction data that conforms to physical laws.

[0042] In this embodiment, a physical constraint equation is defined—a source-sink term is established. pollutant concentration X The time-evolving differential equation BA-DAE (Discrete Boundary-Aware Diffusion-Advection Equation) is a partial differential equation derived from the naive diffusion-advection equation through modification and simplification. It simulates the two core physical processes of atmospheric pollutants in an open system: diffusion and advection. The overall equation is expressed as follows: ; in, It is a gating coefficient, estimated by a linear layer. It is the diffusion coefficient. It is a correction coefficient, and both are obtained by a recurrent network estimator. and These are the Laplace operators for the diffusion plot and the flow plot, respectively; Define a diffusion graph—a diffusion graph is a static undirected graph with the same size as the adjacency matrix A, but it stores edge weights equal to the node weights. i , j Half-sine distance between The reciprocal of, that is: ; Define the advection graph—the advection graph is a dynamic directed graph of the same size as A, but the weights of the stored edges are determined by the wind vector and the directed edges at each time step (i.e., there is an advection graph at each time step, not a static one). Its edge weights are defined as follows: ; in, For wind vectors, For nodes point to For directed edges, the value of the adjacency matrix A first determines the direction of the edge. Does it exist? If it doesn't exist, then... If so, there is no need to construct this edge weight. Otherwise, we can further use a ReLU function to determine whether the dot product between the two vectors is a positive number (that is, the angle between the wind direction and the direction of the directed edge is an acute angle, and the wind can promote the advection of pollutants along the direction of the directed edge) or a non-positive number (the angle between the wind direction and the directed edge is a right angle or an obtuse angle, and the wind cannot push the pollutants to flow along the direction of the directed edge).

[0043] In this embodiment, the BA-DAE equation is solved using a constant differential equation—based on the above differential equation, the pollutant concentration at several future time points is calculated. Specifically, using... t Input of time Using the initial values, solve using the constant differential equation. Output the future The concentration of the step is used to analyze the future output using TAGRU. The concentration of each step is encoded to generate a second predictive feature that conforms to physical laws. Defined as: ; Where ODESolve represents the solution of the BA-DAE of the divine ordinary differential equation, and its input variables include initial conditions. ,differential Solve for the length interval Learnable parameter set .

[0044] Step S17: Based on the geographic relationship map, perform gating fusion and graph attention fusion on the first prediction feature and the second prediction feature to obtain the air pollutant concentration prediction result.

[0045] In this embodiment, the first and second predicted features are concatenated and fused using a linear layer to obtain an initial fused feature; a softmax layer is used to determine the gating coefficient, and the initial fused feature is gated and fused based on the gating coefficient to obtain a fused feature; graph attention fusion is performed on the fused feature based on the geographic relationship graph to obtain a target graph attention fused feature; the target graph attention fused feature is decoded using a decoder to obtain the air pollutant concentration prediction result.

[0046] In this embodiment, the first prediction feature conforming to the distribution of historical time-series data was obtained through the above steps. ( ) and the second predictive feature that conforms to physical laws ( ), N Indicates the number of sites. Indicates the time step of the prediction. With input sequence X Feature Dimensions D different, It is a higher-dimensional hidden space dimension.

[0047] To combine the advantages of these two different methods, feature fusion is further performed on the two data representations to obtain more accurate prediction results. The steps are as follows: Gating fusion—performed within each site and The integration, specifically, firstly, will and The data is then concatenated, fused through a linear layer, and then passed through a softmax layer to obtain the gating coefficients. g Next, a gating fusion is performed, defined as: ; Graph Attention Fusion—Based on the previously defined adjacency matrix A, it fuses possible spatial influence relationships using a graph attention mechanism. N Each site As a node attribute, the edge attribute is the reciprocal of the half-sine distance between stations (i.e., the edge attribute defined in the diffusion graph), and the stacked 3-layer graph attention network learns the attention weights between nodes.

[0048] After feature fusion, the attention fusion features of the target map were obtained. Then, a decoder was defined to decode the output, and finally the predicted air pollutant concentration results were obtained. 1 represents the final pollutant concentration dimension, requiring only a single scalar value. The decoder structure is as follows: Figure 4 As shown, Conv2d represents two-dimensional convolution, and the following is the transformation of its internal feature dimensions.

[0049] This application proposes a physical-sensory temporal decomposition deep learning model for pollutant concentration prediction, specifically for predicting air quality concentrations such as PM2.5. The model mainly includes a dual-branch processing module and a feature fusion module. The structure of the physical-sensory temporal decomposition deep learning model is as follows: Figure 5 As shown: The dual-branch processing module consists of two parts: trend-period-noise decomposition and physical perception modeling. (1) The trend-cycle-noise decomposition part extracts the trend component using the exponential moving average method and extracts the trend features using TAGRU; then, MPAA is used to capture the cycle component features; finally, the noise component is obtained by subtracting the trend component and cycle component from the original time series data, and the noise component is processed by extracting features in parallel using a multi-scale convolution module. The results of the three components are concatenated and encoded to obtain the first predictive feature that conforms to the distribution of historical time series data; (2) The physical perception modeling part introduces the discrete boundary perception diffusion-advection equation and solves it using the constant differential equation. Then, the solution is encoded using the discrete attention gating loop mechanism to obtain the second prediction feature that conforms to the physical law.

[0050] The feature fusion module consists of two parts: gated fusion and graph attention fusion. First, the outputs of the two parts of the dual-branch processing module are subjected to station-by-station gated fusion. Then, graph attention fusion is performed based on the spatial relationship between stations (i.e., the geographic relationship map). Finally, the air pollutant concentration prediction result is obtained through a decoder.

[0051] This application addresses the problem of existing solutions lacking explicit decomposition of data patterns, making the learning of data features more accurate and improving the robustness and interpretability of prediction results. The proposed TAGRU solves the problem of insufficient extraction of discontinuous state information by traditional gated loop mechanisms, improving the accuracy of long-term time series modeling. With the embedding of physical knowledge constraints, it combines the ability of deep learning models to learn data distribution with the ability of physical perception models to constrain objective physical laws. The model's prediction results not only conform to the physical laws of the objective world, but also can quickly predict results within seconds, which has reference value for emergency response to some public events (such as industrial leaks).

[0052] In this embodiment, station monitoring data and station latitude and longitude are acquired. Missing values ​​are imputed and normalized in the station monitoring data to obtain a time series. A geographic relationship map of the stations is constructed based on the station latitude and longitude. Trend features are extracted and processed from the geographic relationship map using the exponential moving average method and a gated cyclic unit to obtain trend component features. Periodic features are extracted from the geographic relationship map using an attention mechanism to obtain periodic component features. A noise sequence is determined based on the time series, the trend component features, and the periodic component features. The noise sequence is batch normalized, activated by a function, and concatenated with vectors to obtain noise component features. The trend component features, periodic component features, and noise component features are concatenated and fused to obtain a first prediction feature. The time series is solved using the advection diffusion equation and based on the neural network ordinary differential equation, and the solution is encoded to obtain a second prediction feature. Gated fusion and graph attention fusion are performed on the first and second prediction features to obtain the predicted air pollutant concentration. This application performs missing value imputation and normalization on site monitoring data, avoiding problems such as gradient vanishing or numerical instability. A geographic relationship map of the sites is constructed based on their latitude and longitude. Using the exponential moving average method and gated recurrent units, trend features are extracted and processed from the geographic relationship map. This addresses the problem of insufficient extraction of discontinuous state information by traditional gated recurrent mechanisms, improving the accuracy of long-term time series modeling. An attention mechanism is used to extract periodic features from the geographic relationship map. Based on time series, trend component features, and periodic component features, a noise sequence is identified. This noise sequence undergoes batch normalization, function activation, and vector concatenation. The trend component features, periodic component features, and noise component features are then concatenated and fused to obtain the first predictive feature. To address the lack of explicit decomposition of data patterns in existing technologies, this method enables more accurate learning of data features and improves the robustness and interpretability of prediction results. It utilizes the advection-diffusion equation and solves the time series based on the neural network ordinary differential equation, encoding the solution to obtain a second predictive feature. This approach combines the data distribution learning capabilities of deep learning models with the constraint capabilities of physical perception models on objective physical laws. Gating fusion and graph attention fusion are then applied to the first and second predictive features to obtain air pollutant concentration prediction results, improving the accuracy of air pollutant concentration prediction. The prediction results not only conform to the physical laws of the objective world but also provide rapid predictions within seconds, offering valuable reference for emergency responses to public events such as industrial leaks.

[0053] See Figure 6 As shown, an embodiment of the present invention discloses an air pollutant concentration prediction device, which may specifically include: The data processing module 11 is used to acquire site monitoring data and site latitude and longitude, perform missing value filling and normalization processing on the site monitoring data to obtain a time series, and construct a geographical relationship map of the site based on the site latitude and longitude. The geographic relationship graph processing module 12 is used to extract and process trend features of the geographic relationship graph using the exponential moving average method and based on a gated loop unit to obtain trend component features. Periodic feature extraction module 13 is used to extract periodic features from the geographic relationship map using an attention mechanism to obtain periodic component features; The noise sequence processing module 14 is used to determine a noise sequence based on the time series, the trend component features, and the periodic component features, and to perform batch normalization, function activation, and vector concatenation on the noise sequence to obtain noise component features. The first prediction feature determination module 15 is used to splice and fuse the trend component feature, the periodic component feature, and the noise component feature to obtain the first prediction feature; The second prediction feature determination module 16 is used to solve the time series using the advection diffusion equation and based on the constant differential equation, and to encode the solution result to obtain the second prediction feature. The fusion module 17 is used to perform gating fusion and graph attention fusion on the first prediction feature and the second prediction feature to obtain the air pollutant concentration prediction result.

[0054] In this embodiment, station monitoring data and station latitude and longitude are acquired. Missing values ​​are imputed and normalized in the station monitoring data to obtain a time series. A geographic relationship map of the stations is constructed based on the station latitude and longitude. Trend features are extracted and processed from the geographic relationship map using the exponential moving average method and a gated cyclic unit to obtain trend component features. Periodic features are extracted from the geographic relationship map using an attention mechanism to obtain periodic component features. A noise sequence is determined based on the time series, the trend component features, and the periodic component features. The noise sequence is batch normalized, activated by a function, and concatenated with vectors to obtain noise component features. The trend component features, periodic component features, and noise component features are concatenated and fused to obtain a first prediction feature. The time series is solved using the advection diffusion equation and based on the neural network ordinary differential equation, and the solution is encoded to obtain a second prediction feature. Gated fusion and graph attention fusion are performed on the first and second prediction features to obtain the predicted air pollutant concentration. This application performs missing value imputation and normalization on site monitoring data, avoiding problems such as gradient vanishing or numerical instability. A geographic relationship map of the sites is constructed based on their latitude and longitude. Using the exponential moving average method and gated recurrent units, trend features are extracted and processed from the geographic relationship map. This addresses the problem of insufficient extraction of discontinuous state information by traditional gated recurrent mechanisms, improving the accuracy of long-term time series modeling. An attention mechanism is used to extract periodic features from the geographic relationship map. Based on time series, trend component features, and periodic component features, a noise sequence is identified. This noise sequence undergoes batch normalization, function activation, and vector concatenation. The trend component features, periodic component features, and noise component features are then concatenated and fused to obtain the first predictive feature. To address the lack of explicit decomposition of data patterns in existing technologies, this method enables more accurate learning of data features and improves the robustness and interpretability of prediction results. It utilizes the advection-diffusion equation and solves the time series based on the neural network ordinary differential equation, encoding the solution to obtain a second predictive feature. This approach combines the data distribution learning capabilities of deep learning models with the constraint capabilities of physical perception models on objective physical laws. Gating fusion and graph attention fusion are then applied to the first and second predictive features to obtain air pollutant concentration prediction results, improving the accuracy of air pollutant concentration prediction. The prediction results not only conform to the physical laws of the objective world but also provide rapid predictions within seconds, offering valuable reference for emergency responses to public events such as industrial leaks.

[0055] In some specific embodiments, the data processing module 11 may specifically include: The acquisition module is used to acquire site monitoring data and site latitude and longitude; the site monitoring data includes particulate matter concentration, temperature, humidity, pressure, wind direction, and wind speed; The normalization processing module is used to fill in missing values ​​in the site monitoring data by interpolating adjacent means, and to normalize the filled site monitoring data to obtain a time series. The geographic relationship map construction module is used to calculate the half-sine distance between stations based on their latitude and longitude to construct a geographic relationship map.

[0056] In some specific embodiments, the geographic relationship map processing module 12 may specifically include: The exponential moving average modeling and trend feature extraction module is used to perform exponential moving average modeling and trend feature extraction on the geographic relationship map using the exponential moving average method to obtain trend components. The feature processing module is used to input the trend component into the gated loop unit for feature processing to obtain the trend component features.

[0057] In some specific embodiments, the periodic feature extraction module 13 may specifically include: An attention score calculation module is used to calculate an attention score based on the geographic relationship map, utilizing an attention mechanism. The geographic relationship map stitching module is used to stitch the attention score with the geographic relationship map, and to use a multilayer perceptron to normalize the stitched data in order to calculate the periodic component features.

[0058] In some specific embodiments, the noise sequence processing module 14 may specifically include: The noise sequence determination module is used to calculate the residual sequence between the time series and the trend component features and the periodic component features, and to use the residual sequence as the noise sequence; The vector concatenation and residual concatenation module is used to perform batch normalization on the noise sequence using multi-scale convolution, activate the function using the GeLU function, and perform vector concatenation and residual concatenation based on the noise sequence to obtain noise component features.

[0059] In some specific embodiments, the second predictive feature determination module 16 may specifically include: The definition module is used to define the advection-diffusion equation, diffusion diagram, and advection diagram. The solution module is used to solve the time series using the advection-diffusion equation, the diffusion map, and the advection map, and based on the normal differential equation, to obtain the initial prediction results; The encoding module is used to encode the initial prediction result using a gated loop unit to generate prediction data and a second prediction feature that conforms to physical laws.

[0060] In some specific embodiments, the fusion module 17 may specifically include: The initial fusion feature determination module is used to concatenate the first and second predicted features and fuse them using a linear layer to obtain the initial fusion features. The gated fusion module is used to determine the gate coefficients using the softmax layer, and to perform gated fusion on the initial fusion features based on the gate coefficients to obtain fusion features; The graph attention fusion module is used to perform graph attention fusion on the fusion features to obtain the target graph attention fusion features; The decoding module is used to decode the attention fusion features of the target image using a decoder to obtain the air pollutant concentration prediction result.

[0061] Figure 7 This is a schematic diagram of an electronic device provided in an embodiment of this application. The electronic device 20 may specifically include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 stores a computer program, which is loaded and executed by the processor 21 to implement the relevant steps in the air pollutant concentration prediction method performed by the electronic device disclosed in any of the foregoing embodiments.

[0062] In this embodiment, the power supply 23 is used to provide operating voltage for each hardware device on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and external devices, and the communication protocol it follows can be any communication protocol applicable to the technical solution of this application, and is not specifically limited here; the input / output interface 25 is used to acquire external input data or output data to the outside world, and its specific interface type can be selected according to specific application needs, and is not specifically limited here.

[0063] In addition, the memory 22, as a carrier for resource storage, can be a read-only memory, random access memory, disk or optical disk, etc. The resources stored on it include operating system 221, computer program 222 and data 223, etc., and the storage method can be temporary storage or permanent storage.

[0064] The operating system 221 manages and controls the various hardware devices on the electronic device 20 and the computer program 222 to enable the processor 21 to perform calculations and processing on the data 223 in the memory 22. The operating system 221 can be Windows, Unix, Linux, etc. The computer program 222, in addition to including a computer program capable of performing the air pollutant concentration prediction method executed by the electronic device 20 as disclosed in any of the foregoing embodiments, may further include computer programs capable of performing other specific tasks. The data 223 may include data received by the air pollutant concentration prediction device from external devices, as well as data collected by its own input / output interface 25.

[0065] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0066] Furthermore, embodiments of this application also disclose a computer-readable storage medium storing a computer program. When the computer program is loaded and executed by a processor, it implements the steps of the air pollutant concentration prediction method disclosed in any of the foregoing embodiments.

[0067] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0068] The present invention provides a detailed description of an air pollutant concentration prediction method, apparatus, device, and storage medium. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for predicting air pollutant concentrations, characterized in that, include: Acquire site monitoring data and site latitude and longitude, perform missing value imputation and normalization on the site monitoring data to obtain a time series, and construct a geographic relationship map of the sites based on the site latitude and longitude; The time series is subjected to trend feature extraction and feature processing using the exponential moving average method and based on a gated cyclic unit to obtain trend component features; The time series is subjected to periodic feature extraction using an attention mechanism to obtain periodic component features; Based on the time series, the trend component features, and the periodic component features, a noise series is determined. The noise series is then subjected to batch normalization, function activation, and vector concatenation to obtain the noise component features. The trend component feature, the periodic component feature, and the noise component feature are spliced ​​and fused to obtain the first prediction feature; The time series is solved using the advection diffusion equation and based on the neuron ordinary differential equation, and the solution is encoded to obtain the second predictive feature; Based on the geographic relationship map, gating fusion and graph attention fusion are performed on the first prediction feature and the second prediction feature to obtain the air pollutant concentration prediction result.

2. The air pollutant concentration prediction method according to claim 1, characterized in that, The process of acquiring site monitoring data and site latitude and longitude, performing missing value imputation and normalization on the site monitoring data to obtain a time series, and constructing a geographic relationship map of the sites based on the site latitude and longitude includes: Acquire site monitoring data and site latitude and longitude; the site monitoring data includes particulate matter concentration, temperature, humidity, pressure, wind direction, and wind speed; Missing values ​​in the site monitoring data are filled by interpolating adjacent means, and the filled site monitoring data is then normalized to obtain a time series. The half-sine distance between stations is calculated based on their latitude and longitude to construct a geographic relationship map.

3. The method for predicting air pollutant concentration according to claim 1, characterized in that, The process of extracting and processing trend features from the time series using the exponential moving average method and based on a gated cyclic unit to obtain trend component features includes: The time series was modeled using the exponential moving average method, and trend features were extracted to obtain the trend components. The trend component is input into a gated loop unit for feature processing to obtain the trend component feature.

4. The method for predicting air pollutant concentration according to claim 1, characterized in that, The process of extracting periodic features from the time series using an attention mechanism to obtain periodic component features includes: An attention mechanism is used, and an attention score is calculated based on the time series. The attention score is concatenated with the time series, and the concatenated data is normalized using a multilayer perceptron to calculate the periodic component features.

5. The method for predicting air pollutant concentrations according to claim 1, characterized in that, The noise sequence is determined based on the time series, the trend component features, and the periodic component features. Batch normalization, function activation, and vector concatenation are then performed on the noise sequence to obtain the noise component features, including: Calculate the residual sequence between the time series, the trend component feature, and the periodic component feature, and use the residual sequence as a noise sequence; The noise sequence is batch normalized using multi-scale convolution, activated using the GeLU function, and then vector concatenation and residual concatenation are performed based on the noise sequence to obtain noise component features.

6. The method for predicting air pollutant concentration according to claim 1, characterized in that, The process involves using the advection-diffusion equation and solving the time series based on the ordinary differential equation, then encoding the solution to obtain the second predicted feature, including: Define the advection-diffusion equation, diffusion diagram, and advection diagram; Using the advection-diffusion equation, the diffusion map, and the advection map, and solving the time series based on the neural ordinary differential equation, the initial prediction results are obtained; The initial prediction result is encoded using a gated loop unit to generate a second prediction feature that conforms to physical laws.

7. The method for predicting air pollutant concentrations according to any one of claims 1 to 6, characterized in that, The step of performing gating fusion and graph attention fusion on the first and second prediction features based on the geographic relationship graph to obtain the air pollutant concentration prediction result includes: The first and second predicted features are concatenated and fused using a linear layer to obtain the initial fused features; The gate coefficients are determined using a softmax layer, and the initial fused features are then subjected to gated fusion based on the gate coefficients to obtain the fused features. Based on the geographic relationship graph, graph attention fusion is performed on the fusion features to obtain the target graph attention fusion features; The air pollutant concentration prediction result is obtained by decoding the attention fusion features of the target map using a decoder.

8. An air pollutant concentration prediction device, characterized in that, include: The data processing module is used to acquire site monitoring data and site latitude and longitude, perform missing value imputation and normalization processing on the site monitoring data to obtain a time series, and construct a geographic relationship map of the site based on the site latitude and longitude. The geographic relationship graph processing module is used to extract and process trend features of the time series using the exponential moving average method and based on a gated loop unit to obtain trend component features. The periodic feature extraction module is used to extract periodic features from the time series using an attention mechanism to obtain periodic component features. The noise sequence processing module is used to determine the noise sequence based on the time series, the trend component features, and the periodic component features, and to perform batch normalization, function activation, and vector concatenation on the noise sequence to obtain the noise component features. The first prediction feature determination module is used to splice and fuse the trend component feature, the periodic component feature, and the noise component feature to obtain the first prediction feature; The second predictive feature determination module is used to solve the time series using the advection diffusion equation and based on the constant differential equation, and to encode the solution results to obtain the second predictive feature. The fusion module is used to perform gating fusion and graph attention fusion on the first prediction feature and the second prediction feature based on the geographic relationship graph to obtain the air pollutant concentration prediction result.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the air pollutant concentration prediction method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, Used to store a computer program; wherein, when the computer program is executed by a processor, it implements the air pollutant concentration prediction method as described in any one of claims 1 to 7.

Images