A deep neural network-based PM2.5 full-coverage prediction method

Abstract

The application discloses a PM2.5 full coverage prediction method based on a deep neural network, preprocessing of previously acquired air pollution concentration, meteorological data and land use data, and division into training data and test data; a deep neural network prediction model STA-ConvLSTM is constructed and trained; STA-ConvLSTM takes a CNN network as a bottom layer, extracts the spatial correlation of grid data through convolution; STA-LSTM with stacked multiple layers of space-time memory units and space memory units is taken as an intermediate layer of the prediction model, used for extracting the time correlation and the spatial correlation features; the last layer uses a CNN layer to decode the features extracted by the STA-LSTM unit. The application integrates multi-source heterogeneous data, considers more influencing factors, and considers the problems of space-time correlation, reduction of overfitting, avoidance of gradient disappearance and gradient explosion, full coverage prediction of regions, use of time attention mechanism and space attention mechanism, and improvement of prediction accuracy and other advantages.

Description

Technical Field

[0001] This invention belongs to the field of urban air pollutant concentration monitoring and early warning, specifically involving a comprehensive prediction method for PM2.5 based on deep neural networks. Background Technology

[0002] In recent years, against the backdrop of China's rapid industrialization and technological development, air pollution has become a prominent problem, posing significant threats to public health. Air pollution has become a major concern for society and researchers alike. Fine particulate matter (PM2.5) is the most dangerous component, as it can directly enter the lungs and cause serious harm. Therefore, developing an efficient and accurate PM2.5 prediction method is essential.

[0003] Currently, numerous methods for predicting air pollution have been proposed, including traditional statistical methods, machine learning, and artificial neural networks. Machine learning, in particular, is widely used and includes algorithms such as Support Vector Machines (SVM), Decision Trees, Random Forests, and Neural Networks. However, these methods have limitations in terms of accuracy and real-time performance. In recent years, with the development of deep learning technology, significant breakthroughs have been achieved in various research fields, and air pollution prediction methods based on deep learning models have gradually become a research hotspot. Effective training of large amounts of data using deep learning techniques can effectively extract the spatiotemporal correlations between data points.

[0004] While existing deep learning methods have shown promising performance in air pollution prediction, most models rely on single air pollution data points, neglecting the influence of meteorological and topographical factors on air pollution propagation. Current models primarily predict air pollutants at individual sites, ignoring comprehensive air pollution coverage and exhibiting the problem of underestimating high-value air pollutants. Most current deep learning models struggle to capture long-term dependent patterns, as their spatiotemporal correlations decay over time, leading to significant loss of prediction accuracy. Furthermore, most current models cannot simultaneously integrate spatiotemporal features to effectively predict pollutant conditions over a future period. Summary of the Invention

[0005] Purpose of the invention: This invention proposes a method for full-coverage prediction of PM2.5 based on deep neural networks. Its purpose is to overcome the shortcomings and defects of the existing technology, achieve full-coverage prediction of PM2.5, and obtain future spatiotemporal evolution data.

[0006] Technical Solution: This invention provides a method for full-coverage prediction of PM2.5 based on deep neural networks, specifically including the following steps:

[0007] (1) The pre-acquired air pollution concentration, meteorological data and land use data are preprocessed and divided into training data and test data;

[0008] (2) Construct a deep neural network prediction model STA-ConvLSTM to predict the pollutant concentration in the target area; the STA-ConvLSTM uses a CNN network as the bottom layer to process the input data to extract spatial features and extracts the spatial correlation of grid data through convolution; it uses stacked STA-LSTM with multiple layers of spatiotemporal memory units and spatial memory units as the middle layer of the prediction model to extract features of temporal correlation and spatial correlation; the last layer uses a CNN layer combined with the features extracted by the STA-LSTM units for decoding;

[0009] (3) Train the deep neural network prediction model STA-ConvLSTM using training data;

[0010] (4) Use the trained model to predict the PM2.5 concentration in the target area for the next N hours.

[0011] Furthermore, the implementation process of step (1) is as follows:

[0012] The collected multi-source heterogeneous data were interpolated using Kriging space interpolation, with the data stations' latitude and longitude used for interpolation filling. The data was interpolated into a 100*100 grid, generating the original feature matrix. An exponential model was then used to fit the exponential model of the semivariogram. The interpolation result can be expressed as:

[0013]

[0014] Where Z(x) represents the spatial interpolation of the unknown point, Z(xi) represents the data value of the known point, and h ij Represents a known point x i The Euclidean distance between the unknown point x and the unknown point x, α is the correlation length parameter in the semivariogram function, which is used to control the rate at which spatial autocorrelation gradually decreases with increasing distance, n is the number of known points, and ∈ represents the error term;

[0015] The original feature matrix is ​​upgraded by using 1×1 convolution kernels and filters to increase the number of output channels, enabling information from different channels to interact and merge. Then, 1×1 convolution kernels are used to aggregate the input data into a two-dimensional feature vector, which is compatible with the input of subsequent models and normalizes the data.

[0016] Furthermore, the process for extracting features with temporal and spatial correlations is as follows:

[0017] S1: Input the spatial features extracted by the CNN network into the prediction unit based on multi-layer STA_LSTM stacking. The input of the first prediction unit is the feature X extracted by the CNN network. t ∈R B×C×H×W B and C represent the batch size and channels of the feature map; temporal memory. Spatial memory and hidden state l represents the current prediction unit layer number, and t represents the current time. After the first prediction unit extracts the temporal characteristics of pollutant concentration, meteorological, and land use data for the prediction area, the data is encoded to generate new data. Spatial memory and hidden state Will and Add them to the hidden state and time memory lists respectively. The current hidden state and spatial memory, as well as the historical hidden state and time memory, serve as the input for the next layer.

[0018] S2: Add SAM at the top layer of the prediction unit, merge the hidden states of each layer output of STA_LSTM into attention keys and values, use the current output as a query to form the corresponding SAM output, so that more information from the lower layers can be integrated into the top layer to improve prediction.

[0019] S3: When time t reaches prediction stage k, TAM is embedded into the prediction unit of each layer. The historical feature grid data is connected in the channel size as the key and value of attention. The channel attention function softmax is used to use the current output as the query. The current output of TAM and the historical hidden state are used as the query and input of SAM.

[0020] Furthermore, in step (2), the features extracted using the CNN layer combined with the STA_LSTM unit are used for decoding, resulting in:

[0021]

[0022]

[0023] Among them, f SL For STA_LSTM functions, Let be the output of the STA-LSTM at time t and the hidden state of each encoder layer, respectively. As a unit of historical memory, Y t+1The predicted output of pollutant concentration at time t is given by W, where W represents the network parameters and b represents the bias. Finally, the features extracted by the STA-LSTM unit are decoded using a CNN layer to complete the prediction of pollutant concentration at time t+1.

[0024] Furthermore, the implementation process of step S1 is as follows:

[0025] The spatial features X extracted from the CNN network layers t ∈R B×C×H×W The initialized temporal memory, spatial memory, and hidden state serve as the raw inputs to the STA-LSTM, allowing... To be a layer of time memory above the current time. To be a spatial memory layer above the current time. The current time step represents the hidden state of the previous layer, is the Hadamard product, σ represents the sigmoid activation function, and * represents the convolution operator. The feature extraction process in the prediction unit STA-LSTM is as follows:

[0026]

[0027]

[0028]

[0029]

[0030]

[0031]

[0032]

[0033] Among them, W hf W hi W h3 These represent the forget gate, input gate, output gate, and h during the feature extraction process, respectively. t-1 The weighting coefficient, W xf W xi W x3 These represent the forget gate, input gate, output gate, and x during the feature extraction process, respectively. t The weighting coefficient, b i b f b3 and b4 represent the bias values ​​in the forget gate, input gate, output gate, and feature extraction process, respectively. These represent the temporal memory, spatial memory, and hidden state of the l-th layer at time t, respectively.

[0034] In standard LSTM, for Using the original door, another set of door structures was constructed in the same way to accommodate... At this point, the STA-LSTM cell simultaneously contains both temporal and spatial memory units; the final hidden state. Based on spatiotemporal memory fusion; to connect memories from different directions, the STA-LSTM unit uses a shared output gate to process the two types of memories, thus achieving seamless memory fusion; in addition, dimensionality reduction is performed using 1×1 convolutional layers to hide the hidden states. It has the same dimensions as a memory cell.

[0035] Furthermore, step S2 is implemented as follows:

[0036] The original hidden state of the top layer The reshaping operation can be used to convolve the query Q. S ∈R B ×C×(H*W) B and C represent the batch size and channels of the feature map, respectively; hidden states at the same time step in different layers. Also generates key K S ∈R B×((L-1)*C)×(H*W) Sum V S ∈R B×((L-1)*C)×(H*W) New hidden state Obtained through the following methods:

[0037]

[0038] Reshaping the size back back, With the original hidden state The sums are then normalized by the LayerNormalization layer and used as the output of the SAM module.

[0039] Furthermore, step S3 is implemented as follows:

[0040] Pass the current hidden state on the convolutional layer. Query Q is generated through plasticity operations. t ∈R B ×C×(H*W) Similarly, from historical input The key K is obtained through two independent convolutions. T ∈R B×(t*C)×(H*W) Sum V T ∈R B×(t*C)×(H*W) Then, the new hidden state. Calculations can be made using the time-based formula:

[0041]

[0042] Finally, Reshape to the same size as the original hidden state and add the original hidden state normalized by the LayerNormalization layer. As the output of the TAM module.

[0043] Furthermore, the implementation process of step (3) is as follows:

[0044] The mean squared error function is used as the loss function for the model:

[0045]

[0046] in, and X A+1:T These are the predicted data and the actual data, respectively. H and W are the height and width of the grid data, and (i,j) represents the position in the grid area.

[0047] Furthermore, the implementation process of step (4) is as follows:

[0048] The predicted output yields a grid of data. To obtain the pollutant concentration values ​​for the target area grid, the row and column coordinates of the target grid are precisely located based on the interpolated station's latitude and longitude coordinates. The data in the target grid represents the predicted pollutant concentration value for a specific target station area. The formula for obtaining the latitude and longitude coordinates is:

[0049]

[0050]

[0051] y = Y i，j

[0052] Where l3n_PuerT and lat_querT are the longitude and latitude of the query station, respectively; l3n_max and lat_max are the longitude and latitude with the largest interpolation, respectively; l3n_min and lat_min are the longitude and latitude with the smallest interpolation, respectively; and height and width represent the height and width of the grid.

[0053] Beneficial effects: Compared with the prior art, the beneficial effects of the present invention are as follows:

[0054] 1. This invention combines pollutant concentration, meteorological, and land use data from multiple sites to effectively fuse the spatiotemporal features of multi-source heterogeneous data. By using a multidimensional convolutional CNN network, it combines the features of multi-source data from multiple regions to extract deep spatial correlations and achieve full-coverage prediction. The model adopts a fully convolutional method, which can effectively eliminate the large amount of feature loss caused by pooling layers, thereby better extracting the spatial features of pollutant, meteorological, and land use data.

[0055] 2. The model is built on a stacked STA-LSTM architecture of deep learning. It uses convolutional neural networks to capture spatial interaction relationships and long short-term memory networks to capture temporal correlations. It incorporates an attention mechanism to capture global information and improve the modeling ability of spatiotemporal information. Regularization technology is used to reduce the overfitting problem of the model and avoid problems such as gradient vanishing and gradient exploding. The time attention module is embedded in the model to solve the problem of historical information being lost over time. The spatial attention module is embedded to solve the problem of information loss from the bottom to the top of the stacked features of multi-layer prediction units, which enhances the local prediction ability and further improves the accuracy of the prediction model from the perspective of time series data.

[0056] 3. This invention uses multi-source heterogeneous data, fully considers spatiotemporal issues, and overcomes the problems of insufficient feature extraction and weak data correlation in prediction. The prediction results can be linked with the predicted PM2.5 concentration data, and the accuracy of continuous prediction of PM2.5 concentration over a future period is greatly improved.

[0057] 4. Existing deep learning-based PM2.5 prediction methods rarely consider full data coverage. This invention addresses this issue by constructing a deep neural network model that exhibits good prediction performance for cities with densely packed stations. Attached Figure Description

[0058] Figure 1 This is a schematic diagram of the STA-ConvLSTM model structure constructed in this invention;

[0059] Figure 2 This is a schematic diagram of the STA-LSTM prediction unit structure;

[0060] Figure 3 This is a schematic diagram of the SAM module structure;

[0061] Figure 4 This is a schematic diagram of the TAM module structure;

[0062] Figure 5 This is a diagram of spatiotemporal evolution. Detailed Implementation

[0063] The present invention will now be described in further detail with reference to the accompanying drawings.

[0064] This invention proposes a comprehensive PM2.5 prediction method based on deep neural networks, specifically including the following steps:

[0065] Step 1: Collect air pollution concentration, meteorological data and land use data from environmental monitoring data, and preprocess the collected multi-source heterogeneous data and divide it into training data and test data.

[0066] The collected air quality, meteorological, and land use data were filled with missing values ​​using Kriging spatial interpolation based on the latitude and longitude of the data stations. First, the data was preprocessed. Then, Kriging spatial interpolation was used, selecting an appropriate interpolation method based on the data type. This invention uses an exponential model to fit the exponential model of the semivariogram; the interpolation result can be expressed as:

[0067]

[0068] Where Z(x) represents the spatial interpolation of the unknown point, Z(xi) represents the data value of the known point, and h ij Represents a known point x i The Euclidean distance between the unknown point x and the unknown point x, where α is the correlation length parameter in the semivariogram function, used to control the rate at which spatial autocorrelation gradually weakens with increasing distance, n is the number of known points, and ∈ represents the error term.

[0069] The data is interpolated into a 100*100 grid to generate the original feature matrix. The original feature matrix is ​​then increased in dimensionality using 1×1 convolutional kernels and filters to enhance the number of output channels, allowing information from different channels to interact and merge, thus improving the model's ability to extract nonlinear features. Then, a 1×1 convolutional kernel is used to aggregate the input data into a two-dimensional feature vector, which is compatible with the subsequent model input. The data is then normalized. The dataset uses the first two years' data as the training set and the data from the following year as the test set, completing the initialization of the deep neural network prediction model.

[0070] Step 2: As Figure 1 As shown, a deep neural network prediction model STA-ConvLSTM is constructed to predict the pollutant concentration in the target area.

[0071] (2.1) STA-ConvLSTM uses a CNN network as its underlying layer to process input data and extract spatial features. It extracts spatial correlations of grid data through convolution. Multi-source heterogeneous data is transformed into a two-dimensional matrix with time series that can be accepted by the CNN network and input into the CNN network to extract spatial features, which are then used as input to the stacked STA-LSTM.

[0072] (2.2) As Figure 2 As shown, a multi-layered STA-LSTM with spatiotemporal memory units and spatial memory units is used as the intermediate layer of the prediction model to extract features with temporal and spatial correlation.

[0073] The trained CNN weight parameters are input into a prediction unit based on a multi-layer STA-LSTM stack. The input to the first prediction unit is the feature X extracted by the CNN network. t ∈R B×C×H×W (B and C represent the batch size and channels of the feature map), temporal memory Spatial memory and hidden state (Except for the initial zero of input X, where t is the current time), after the first-layer prediction unit extracts the temporal characteristics of pollutant concentration, meteorological, and land use data in the prediction area, the data is encoded to generate new data. Spatial memory and hidden state Will and Add them to the hidden state and time memory lists respectively. The current hidden state and spatial memory, as well as the historical hidden state and time memory, serve as the input for the next layer, and so on.

[0074] CNN network training generates X t ∈R B×C×H×W The generated X t The initialized temporal memory, spatial memory, and hidden state serve as the raw inputs to the STA-LSTM, allowing... For the next level of time memory, For the memory of the next higher level of space, The hidden state of the previous layer is represented by °, the Hadamard product is represented by σ, the activation function sigmoid is represented by *, and the convolution operator is represented by *. The feature extraction process in the prediction unit STA-LSTM can be represented by the following formula:

[0075]

[0076]

[0077]

[0078]

[0079]

[0080]

[0081]

[0082] Among them, W hf W hi W h3 These represent the forget gate, input gate, output gate, and h during the feature extraction process, respectively. t-1The weighting coefficient, W xf W xi W x3 These represent the forget gate, input gate, output gate, and x during the feature extraction process, respectively. t The weighting coefficient, b i b f b3 and b4 represent the bias values ​​in the forget gate, input gate, output gate, and feature extraction process, respectively. These represent the temporal memory, spatial memory, and hidden state of the l-th layer at time t, respectively.

[0083] In standard LSTM, for The original door was used, while another set of door structures was constructed using the same method to accommodate it. At this point, the STA_LSTM unit contains both temporal and spatial memory units. The final hidden state. Fusion based on spatiotemporal memory. To connect memories from different directions, the STA_LSTM unit uses a shared output gate to process both types of memories, thus achieving seamless memory fusion. Furthermore, dimensionality reduction is achieved through the use of 1×1 convolutional layers, hiding the state. It has the same dimensions as the memory unit. Unlike simple memory concatenation, this method can effectively simulate spatial changes and trajectories in a spatiotemporal sequence.

[0084] Adding SAM at the top layer of the prediction unit merges the hidden states of each layer of TAM into attention keys and values, forming the corresponding SAM output, thus integrating more information from lower layers into the top layer to improve prediction.

[0085] Specifically, the SAM process is as follows: Figure 3 As shown. The original hidden state of the top layer. The reshaping operation can be used to convolve the query Q. S ∈R B×C×(H*W) Here, B and C represent the batch size and channels of the feature map, respectively. Then, the hidden states at the same time step in different layers... Also generates key K S ∈R B×((L-1)*C)×(H*W) Sum V S ∈R B×((L-1)*C)×(H*W) Finally, the new hidden state It can be obtained in the following ways:

[0086]

[0087] Reshaping the size back back, With the original hidden state The sums are then normalized by the LayerNormalization layer and used as the output of the SAM module.

[0088] When time t reaches prediction stage k, the TAM is embedded into the prediction unit of each layer. Historical feature grid data is concatenated in the channel dimension as attention keys and values. Using the channel attention function softmax, the current output is used as the query. The TAM process is as follows: Figure 4 As shown, the output of TAM is used as the input of SAM (Spatial Attention Module).

[0089] Pass the current hidden state on the convolutional layer. Query Q is generated through plasticity operations. t ∈R B ×C×(H*W) Similarly, from historical input The key K can be obtained through two independent convolutions. T ∈R B ×(t*C)×(H*W) Sum V T ∈R B×(t*C)×(H*W) Then, the new hidden state. Calculations can be made using the time-based formula:

[0090]

[0091] Finally, Reshape to the same size as the original hidden state and add the original hidden state normalized by the LayerNormalization layer. As the output of the TAM module.

[0092] (2.3) The last layer uses CNN layers combined with features extracted by STA-LSTM units for decoding.

[0093]

[0094]

[0095] Where f SL For STA_LSTM functions, Let be the output of STA_LSTM at time t and the hidden state of each encoder layer, respectively. As a unit of historical memory, Y t+1 The predicted output of pollutant concentration at time t is given by W, where W represents the network parameters and b represents the bias. Finally, the features extracted by the STA_LSTM unit are decoded using a CNN layer to complete the prediction of pollutant concentration at time t+1.

[0096] Step 3: Train the deep neural network prediction model using the training data.

[0097] During the training of the STA-ConvLSTM prediction model, the mean squared error function is used as the model's loss function, and the calculation formula is as follows:

[0098]

[0099] Where Y A+1:T It can also be expressed as and X A+1:T These are the predicted data and the actual data, respectively. H and W are the height and width of the grid data, and (i,j) represents the position in the grid area.

[0100] Step 4: Use the trained model to predict the PM2.5 concentration in the target area for the next N hours.

[0101] like Figure 5 As shown, the final prediction output is a grid of data. To obtain the pollutant concentration values ​​for the target area grid or monitoring station, it is necessary to accurately locate the row and column coordinates of the target grid based on the interpolated station's latitude and longitude coordinates. Obtaining the data of the current grid is the predicted pollutant concentration value for a specific target station area. The formula for obtaining latitude and longitude coordinates is:

[0102]

[0103]

[0104] T = Y j , j

[0105] Where l3n_PuerT and lat_PuerT are the longitude and latitude of the query station, respectively; l3n_max and lat_max are the longitude and latitude with the largest interpolation, respectively; l3n_min and lat_min are the longitude and latitude with the smallest interpolation, respectively; and height and width represent the height and width of the grid.

[0106] It should be understood that the specific embodiments described above are merely illustrative or explanatory of the principles of the invention and do not constitute a limitation thereof. Therefore, any modifications, equivalent substitutions, improvements, etc., made without departing from the spirit and scope of the invention should be included within the protection scope of the invention. Furthermore, the appended claims are intended to cover all variations and modifications falling within the scope and boundaries of the appended claims, or equivalent forms of such scope and boundaries.

Claims

1. A method for full-coverage prediction of PM2.5 based on deep neural networks, characterized in that, Includes the following steps: (1) The air pollution concentration, meteorological data and land use data obtained in advance are preprocessed and divided into training data and test data; (2) Construct a deep neural network prediction model STA-ConvLSTM to predict the pollutant concentration in the target area; the STA-ConvLSTM uses a CNN network as the bottom layer to process the input data to extract spatial features and extract the spatial correlation of grid data through convolution. The prediction model uses a stacked STA-LSTM with multiple layers of spatiotemporal memory units and spatial memory units as the intermediate layer to extract features with temporal and spatial correlation; the last layer uses a CNN layer combined with the features extracted by the STA-LSTM units for decoding. (3) Train the deep neural network prediction model STA-ConvLSTM using training data; (4) Use the trained model to predict the PM2.5 concentration in the target area for the next N hours; The process of extracting temporal and spatial correlation features in step (2) is as follows: S1: The spatial features extracted by the CNN network are input into the prediction unit based on a multi-layer STA_LSTM stack. The inputs of the first prediction unit are the features extracted by the CNN network. , and Represents the batch size and channels of the feature map; temporal memory. Spatial memory and hidden state , Let t be the current prediction unit layer number and t be the current time. After the first prediction unit extracts the temporal characteristics of pollutant concentration, meteorological, and land use data for the prediction area, the data is encoded to generate new data. Spatial memory and hidden state ;Will and Add them to the hidden state and time memory lists respectively. The current hidden state and spatial memory, as well as the historical hidden state and time memory, serve as the input for the next layer. S2: Add SAM at the top layer of the prediction unit, merge the hidden states of each layer output of STA_LSTM into attention keys and values, use the current output as a query to form the corresponding SAM output, so that more information from the lower layers can be integrated into the top layer to improve prediction. S3: When time t reaches prediction stage k, TAM is embedded into the prediction unit of each layer. The historical feature grid data is connected in the channel size as the key and value of attention. The channel attention function softmax is used to use the current output as the query. The current output of TAM and the historical hidden state are used as the query and input of SAM.

2. The PM2.5 full-coverage prediction method based on deep neural networks according to claim 1, characterized in that, The implementation process of step (1) is as follows: The collected multi-source heterogeneous data were interpolated using Kriging space interpolation, with the data stations' latitude and longitude used for interpolation filling. The data was interpolated into a 100*100 grid, generating the original feature matrix. An exponential model was then used to fit the exponential model of the semivariogram. The interpolation result can be expressed as: in, Spatial interpolation for unknown points, This represents the data value of a known point. Represents a known point and unknown points The Euclidean distance between them It is the correlation length parameter in the semivariance function, used to control the rate at which spatial autocorrelation gradually decreases with increasing distance. It is the number of known points. Indicates the error term; The original feature matrix is ​​upgraded by using 1×1 convolution kernels and filters to increase the number of output channels, enabling information from different channels to interact and merge. Then, 1×1 convolution kernels are used to aggregate the input data into a two-dimensional feature vector, which is compatible with the input of subsequent models and normalizes the data.

3. The PM2.5 full-coverage prediction method based on deep neural networks according to claim 1, characterized in that, In step (2), the features extracted using the CNN layer combined with the STA_LSTM unit are used for decoding, resulting in: in, For STA_LSTM functions, Let be the output of the STA-LSTM at time t and the hidden state of each encoder layer, respectively. As a unit for historical memory, This is the predicted output of the pollutant concentration at time t. Represents network parameters, This represents the bias value. Finally, a CNN layer is used to decode the features extracted by the STA-LSTM unit to complete the process. Predicting pollutant concentrations at any given time.

4. The PM2.5 full-coverage prediction method based on deep neural networks according to claim 1, characterized in that, The implementation process of step S1 is as follows: Spatial features extracted from CNN network layers The initialized temporal memory, spatial memory, and hidden state serve as the raw inputs to the STA-LSTM, allowing... To be a layer of time memory above the current time. To be a spatial memory layer above the current time. This is the hidden state above the current time. For Hadamard product, This represents the activation function sigmoid. The convolution operator is represented, and the feature extraction process in the prediction unit STA-LSTM is as follows: in, , , , These represent the forget gate, input gate, output gate, and feature extraction process, respectively. The weighting coefficients, , , These represent the forget gate, input gate, output gate, and feature extraction process, respectively. The weighting coefficients, , and , These represent the bias values ​​for the forget gate, input gate, output gate, and feature extraction process, respectively. , , They represent Time of the first The temporal memory, spatial memory, and hidden state of layers; In standard LSTM, for Using the original door, another set of door structures was constructed in the same way to accommodate... At this point, the STA-LSTM cell simultaneously contains both temporal and spatial memory units; the final hidden state. Based on spatiotemporal memory fusion; to connect memories from different directions, the STA-LSTM unit uses a shared output gate to process the two types of memories, thus achieving seamless memory fusion; in addition, dimensionality reduction is performed using 1×1 convolutional layers to hide the hidden states. It has the same dimensions as a memory cell.

5. The PM2.5 full-coverage prediction method based on deep neural networks according to claim 1, characterized in that, The implementation process of step S2 is as follows: The original hidden state of the top layer Through the reshaping operation, convolution can be used to convert queries. ; and These represent the batch size and channels of the feature map, respectively; and the hidden states at the same time step in different layers. Also generate keys Sum New hidden state Obtained through the following methods: Reshaping the size back back, With the original hidden state The sums are then normalized by a Layer Normalization layer and used as the output of the SAM module.

6. The PM2.5 full-coverage prediction method based on deep neural networks according to claim 1, characterized in that, The implementation process of step S3 is as follows: Pass the current hidden state on the convolutional layer. Generate queries through integer operations ; Similarly, from historical input The key is obtained through two independent convolutions. Sum Then, the new hidden state. Calculations can be made using the time-based formula: Finally, Reshape to the same size as the original hidden state and add the original hidden state normalized by Layer Normalization. As the output of the TAM module.

7. The PM2.5 full-coverage prediction method based on deep neural networks according to claim 1, characterized in that, The implementation process of step (3) is as follows: The mean squared error function is used as the loss function for the model: in, ( )and These are the predicted data and the actual data, respectively. and It refers to the height and width of the grid data. Indicates the location within the grid area.

8. The PM2.5 full-coverage prediction method based on deep neural networks according to claim 1, characterized in that, The implementation process of step (4) is as follows: The predicted output yields a grid of data. To obtain the pollutant concentration values ​​for the target area grid, the row and column coordinates of the target grid are precisely located based on the interpolated station's latitude and longitude coordinates. The data in the target grid represents the predicted pollutant concentration value for a specific target station area. The formula for obtaining the latitude and longitude coordinates is: in, and These are the longitude and latitude of the query site, respectively. and These are the longitude and latitude with the largest interpolation values, respectively. and These are the longitude and latitude with the smallest interpolation values, respectively. and This indicates the height and width of the grid.

Images