PM2.5 and O3 combined estimation method, medium and equipment

Abstract

The invention discloses a PM2.5 and O3 combined estimation method, medium and equipment, and relates to the field of environmental pollution prediction.The PM2.5 and O3 combined estimation method mainly comprises the steps that PM2.5 and O3 influence factor data are preprocessed, and a training set, a verification set and a test set are obtained; constructing a PM2.5 and O3 prediction model, and training and verifying the PM2.5 and O3 prediction model by using the training set and the verification set to obtain a final PM2.5 and O3 prediction model; and according to the test set, utilizing the final PM2.5 and O3 prediction model to obtain a PM2.5 and O3 combined estimation result. By implementing the PM2.5 and O3 joint estimation method, the medium and the equipment provided by the invention, the precision and robustness of PM2.5 and O3 estimation can be improved.

Description

Technical Field

[0001] The present invention relates to the field of environmental pollution prediction, and more specifically, to a PM 2.5 and O3 combined estimation methods, media and equipment. Background Art

[0002] It is widely recognized that air pollution poses a serious threat to human health and ecosystems, and there is a large body of evidence linking air pollution to a variety of adverse effects. The increasing levels of pollutants emitted into the atmosphere by human activities are the main cause of the decline in air quality. Among them, PM 2.5 O2 and O3 are two typical and important air pollutants, there are complex chemical reactions between them and they influence each other.

[0003] Scientific research shows that the satellite-derived aerosol optical depth (AOD) and PM 2.5 There is a generally high correlation between the two, which makes it possible to monitor and analyze air pollution through satellite technology. A variety of machine learning models have been developed to infer air pollutants such as PM from satellite images. 2.5 Although these models have achieved good results, most of them still adopt the modeling method for single pollutants. This single method estimates PM 2.5 The methods of modeling PM2.5 and PM3 are not only repetitive and time-consuming, but also often ignore the interactions and common influencing factors between different pollutants. Compared with traditional single modeling methods, multi-task learning can effectively improve the quality of PM2.5 and PM3. 2.5 There is a significant advantage in jointly estimating PM and O3 concentrations. 2.5 Since these two pollutants share common precursors, this method can also provide a better understanding of PM 2.5 There is a complex interaction between PM and O3. However, at this stage, PM 2.5 The joint estimation of PM and O3 concentrations is limited by many factors and is difficult. Therefore, a new PM 2.5 and O3 concentration collaborative estimation method. Summary of the invention

[0004] The object of the present invention is to provide a PM 2.5 Combined estimation methods, media and equipment with O3 can improve PM 2.5 and the accuracy and robustness of O3 estimation.

[0005] The present invention provides a PM 2.5 The combined estimation method of PM and O3 includes the following steps: S1: obtaining PM 2.5 and O3 influencing factors data, PM2.5 Preprocess the data of O3 influencing factors to obtain training set, validation set and test set; S2: Construct PM 2.5 and O3 prediction models, using training and validation sets to predict PM 2.5 The final PM 2.5 and O3 prediction model; S3: Based on the test set, using the final PM 2.5 and O3 prediction model to obtain PM 2.5 Combined estimation results with O3.

[0006] Furthermore, the above PM 2.5 PM and O3 combined estimation method 2.5 and O3 influencing factors including PM 2.5 and O3 data, daily average aerosol optical depth data, daily average NO2 data and global climate and weather reanalysis data sets; the global climate and weather reanalysis data sets include relative humidity, ozone mass mixing ratio, 2-meter height temperature, east component of wind speed at 10 meters height, north component of wind speed at 10 meters height, surface pressure, surface net solar radiation, surface net thermal radiation, surface downward ultraviolet radiation, total ozone column, planetary boundary layer height, evapotranspiration and total precipitation.

[0007] Furthermore, the above PM 2.5 Step S1 of the combined estimation method of PM and O3 specifically includes: 2.5 The O3 influencing factor data are normalized to obtain normalized data; the normalized data are divided according to a preset ratio to obtain a training set, a validation set, and a test set.

[0008] Furthermore, the above PM 2.5 PM and O3 combined estimation method 2.5 and O3 prediction models including PM 2.5 task model and O3 task model; the O3 task model includes a first convolutional neural network module, a first long short-term memory module and a first deep convolutional neural network module connected in sequence; PM 2.5 The task model includes a second convolutional neural network module, a second long short-term memory module, and a second deep convolutional neural network module; Furthermore, the above PM 2.5 The first convolutional neural network module and the second convolutional neural network module of the O3 joint estimation method share the same parameters.

[0009] Furthermore, the above PM 2.5 The first convolutional neural network module and the second convolutional neural network module of the O3 joint estimation method share the same parameters; the first long short-term memory module and the second long short-term memory module share the same parameters.

[0010] Furthermore, the above PM 2.5 The O3 joint estimation method combines the output of the second convolutional neural network module with the output of the first convolutional neural network module as the input of the second long short-term memory module, and combines the output of the second long short-term memory module with the output of the first long short-term memory module as the input of the second deep convolutional neural network module; superimposes the output of the first convolutional neural network module with the output of the second convolutional neural network module as the input of the first long short-term memory module, and superimposes the output of the first long short-term memory module with the output of the second long short-term memory module as the input of the first deep convolutional neural network module.

[0011] Furthermore, the above PM 2.5The first convolutional neural network module and the second convolutional neural network module of the O3 joint estimation method are used to explain spatial dependence; the first convolutional neural network module and the second convolutional neural network module each include 1 input layer, 1 batch normalization layer and 5 convolutional layers; the input layer size of the first convolutional neural network module is 7×7×4×19, and the input layer size of the second convolutional neural network module is 7×7×4×14; the batch normalization layer is used to normalize the original data so that the data maintains a stable distribution before entering the next layer of the network, thereby accelerating model learning; 5 The convolution kernel size of the convolution layer is 3×3, and the number of channels is 64, 128, 256, 128 and 64 respectively. The five convolution layers are used to extract data spatial features. The first convolution neural network module is connected to the first long short-term memory module through a 3D pooling layer, and the second convolution neural network module is connected to the second long short-term memory module through a 3D pooling layer. The size of the 3D pooling layer is 3×3×1. The 3D pooling layer is used to reduce the spatial size of the data and the computational complexity of the model while retaining important feature information to avoid the algorithm from falling into local features and causing overfitting. The short-term memory module and the second long short-term memory module are used to extract time information. The first long short-term memory module and the second long short-term memory module both adopt a double-layer LSTM structure. The number of parameters of each layer of LSTM is 64, of which the output size of the first layer is 1×4, and the output size of the second layer is 1×1; the first long short-term memory module is connected to the first deep convolutional neural network module through a batch normalization layer and a response function ReLU, and the second long short-term memory module is connected to the second deep convolutional neural network module through a batch normalization layer and a response function ReLU, wherein the batch normalization layer is used to control data distribution, and the response function ReLU is used as an optimizer and activation function for the model; the first deep convolutional neural network module and the second deep convolutional neural network module are used to integrate spatial and temporal characteristics and increase the nonlinear fitting effect of pollutants. The first deep convolutional neural network module and the second deep convolutional neural network module both adopt a deep convolutional neural network with 64 hidden layers, and the first deep convolutional neural network module and the second deep convolutional neural network module both use the ReLU function as the optimizer and activation function of the model.

[0012] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which implements the PM as described above when the computer program is executed by a processor. 2.5 and O3 joint estimation method steps.

[0013] The present invention also provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the PM as described above when executing the computer program. 2.5 and O3 joint estimation method steps.

[0014] Implementation of the PM provided by the present invention 2.5 The combined estimation method, medium and equipment with O3 have the following beneficial effects: The present invention uses the CNN model to replace the site PM with the features of the surrounding pixels. 2.5 The weighted features of the proposed method are used to evaluate the contribution of each neighborhood to the estimated ozone level. A batch normalization layer (BN) is added before the first convolution layer to replace the normalization of the original data. A BN layer is also added between the convolution function and the activation function (ReLU) to keep the data in a stable distribution before entering the next layer of the network, thereby accelerating model learning. In order to further utilize the spatiotemporal information of AOD and precursor NO2, the present invention adds an LSTM module after the CNN module, takes the spatiotemporal three-dimensional data set as a sample, and extracts their spatiotemporal information at the same time. The present invention adds a DNN module with 64 DNN units after the CNN and LSTM modules, and also adds a BN layer between the DNN layer and the activation function. The DNN model combines the feature-rich information passed down by the CNN and LSTM layers, which can effectively weigh the importance of different features and provide a final output. The present invention adopts the CLDNN model for PM 2.5 and O3 estimation, by effectively utilizing multi-source data and embedding prior knowledge, improving PM 2.5 The combined estimation accuracy of PM and O3 is 2.5 It can jointly estimate the concentration of PM2.5 and O3, extract features from multidimensional environmental variables, and more accurately consider the interaction between factors; avoid the uncertainty and complexity caused by actual measurements of surrounding locations, and greatly improve PM2.5 2.5 and O3 estimation accuracy and robustness. 2.5 The two pollutants, PM and O3, share common precursors, so the present invention can also help us better understand PM 2.5 The complex interaction between PM and O3 is considered, and the physical and chemical relationship between them is considered to obtain a higher precision 2.5 and O3 pollutants are estimated and mapped. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] The present invention will be further described below with reference to the accompanying drawings and embodiments, in which: Figure 1 The PM provided by the present invention 2.5 Flowchart of the joint estimation method with O3; Figure 2 The PM based on the CLDNN model provided by the present invention 2.5 and a flow chart of the O3 estimation method and system; Figure 3It is the PM of the CLDNN model provided by the present invention 2.5 and O3 estimation effect diagram; Figure 4 It is a structural diagram of the CNN model provided by the present invention; Figure 5 It is the LSTM structure diagram provided by the present invention; Figure 6 It is a DNN structure diagram provided by the present invention; Figure 7 Schematic diagram of the CLDNN model shared parameter comparison test scheme (a) provided by the present invention; Figure 8 Schematic diagram of the CLDNN model shared parameter comparison test scheme (b) provided by the present invention; Fig. 9 Schematic diagram of the CLDNN model shared parameter comparison test scheme (c) provided by the present invention; Fig.10 The PM based on the CLDNN model provided by the present invention 2.5 and the schematic diagram of the structure of the O3 estimation system; Fig.11 It is a structural block diagram of the computer device provided by the present invention. DETAILED DESCRIPTION

[0016] In order to have a clearer understanding of the technical features, purposes and effects of the present invention, specific embodiments of the present invention are now described in detail with reference to the accompanying drawings.

[0017] Figure 1 The PM of this embodiment is shown 2.5 Schematic diagram of the combined estimation method of PM and O3. 2.5 The combined estimation method with O3 includes the following steps: S1: Get PM 2.5 and O3 influencing factors data, PM 2.5 Preprocess the O3 influencing factor data to obtain the training set, validation set and test set; Specifically, PM 2.5 and O3 influencing factors including PM 2.5 and O3 data, daily mean aerosol optical depth data, daily mean NO2 data and global climate and weather reanalysis data sets; global climate and weather reanalysis data sets include relative humidity, ozone mass mixing ratio, 2-meter height temperature, east component of wind speed at 10 meters, north component of wind speed at 10 meters, surface pressure, surface net solar radiation, surface net thermal radiation, surface downward ultraviolet radiation, total ozone column, planetary boundary layer height, evapotranspiration and total precipitation; Specifically, PM 2.5The data on factors affecting O3 include: PM2.5 at 1,641 stations per hour in 2019 2.5 and O3 data; used to estimate PM 2.5 The daily average aerosol optical depth (AOD) data comes from the Long-Term Gap-Free High-Resolution Air Pollutant Concentration Dataset (LGHAP); used for collaborative estimation of PM 2.5 The daily average NO2 data of O3 and NO2 come from the ChinaHighNO2 dataset; the meteorological data come from the fifth generation European Centre for Medium-Range Weather Forecasts (ECMWF) Global Climate and Weather Reanalysis Dataset (ERA5); Specifically, the main meteorological factors in the ERA5 data include: relative humidity (RH), ozone mass mixing ratio (OR), temperature at 2 meters (T2M), east component of wind speed at 10 meters (U10), north component of wind speed at 10 meters (V10), surface pressure (SP), surface net solar radiation (SSR), surface net thermal radiation (STR), surface downward ultraviolet radiation (UVB), total ozone column (TCO3), planetary boundary layer height (BLH), evapotranspiration (E) and total precipitation (TP); among them, relative humidity is calculated on the pressure level data at the pixel scale through the near-surface altitude; Specifically, step S1 includes: S11: To PM 2.5 The O3 influencing factor data is normalized to obtain the normalized data, as shown in the formula:

[0018] in, represents the i-th normalized data, Represents the original PM 2.5 and the i-th data of O3 influencing factor data, Represents the original PM 2.5 and O3 influencing factor data, Indicates the maximum value, Indicates the minimum value; S12: Divide the normalized data according to a preset ratio to obtain a training set, a validation set, and a test set; Specifically, the preset ratio is 7:1:2; S2: Build PM 2.5and O3 prediction models, using training and validation sets to predict PM 2.5 The final PM 2.5 and O3 prediction model; Specifically, PM 2.5 and O3 prediction models including PM 2.5 task model and O3 task model; the O3 task model includes a first convolutional neural network module, a first long short-term memory module and a first deep convolutional neural network module connected in sequence; PM 2.5 The task model includes a second convolutional neural network module, a second long short-term memory module, and a second deep convolutional neural network module; Optionally, the first convolutional neural network module and the second convolutional neural network module share the same parameters; Optionally, the first convolutional neural network module and the second convolutional neural network module share the same parameters; the first long short-term memory module and the second long short-term memory module share the same parameters; Preferably, the output of the second convolutional neural network module is combined with the output of the first convolutional neural network module as the input of the second long short-term memory module, and the output of the second long short-term memory module is combined with the output of the first long short-term memory module as the input of the second deep convolutional neural network module; the output of the first convolutional neural network module is superimposed with the output of the second convolutional neural network module as the input of the first long short-term memory module, and the output of the first long short-term memory module is superimposed with the output of the second long short-term memory module as the input of the first deep convolutional neural network module; Specifically, the first convolutional neural network module and the second convolutional neural network module are used to explain spatial dependence; the first convolutional neural network module and the second convolutional neural network module each include 1 input layer, 1 batch normalization layer and 5 convolutional layers; the input layer size of the first convolutional neural network module is 7×7×4×19, and the input layer size of the second convolutional neural network module is 7×7×4×14; the batch normalization layer is used to normalize the original data so that the data maintains a stable distribution before entering the next layer of the network, thereby accelerating model learning; the 5 convolutional layers The convolution kernel size is 3×3, and the number of channels is 64, 128, 256, 128 and 64 respectively. Five convolutional layers are used to extract data spatial features. The first convolutional neural network module is connected to the first long short-term memory module through a 3D pooling layer, and the second convolutional neural network module is connected to the second long short-term memory module through a 3D pooling layer. The size of the 3D pooling layer is 3×3×1. The 3D pooling layer is used to reduce the spatial size of the data and the computational complexity of the model while retaining important feature information to avoid the algorithm from falling into local features and causing overfitting. The first long short-term memory The first long short-term memory module and the second long short-term memory module are used to extract time information. The first long short-term memory module and the second long short-term memory module both adopt a double-layer LSTM structure. The number of parameters of each layer of LSTM is 64, wherein the output size of the first layer is 1×4, and the output size of the second layer is 1×1; the first long short-term memory module is connected to the first deep convolutional neural network module through a batch normalization layer and a response function ReLU, and the second long short-term memory module is connected to the second deep convolutional neural network module through a batch normalization layer and a response function ReLU, wherein the batch normalization layer is used to control data distribution, and the response function ReLU is used as an optimizer and activation function for the model; the first deep convolutional neural network module and the second deep convolutional neural network module are used to integrate spatial and temporal features and increase the nonlinear fitting effect of pollutants. The first deep convolutional neural network module and the second deep convolutional neural network module both adopt a deep convolutional neural network with 64 hidden layers, and the first deep convolutional neural network module and the second deep convolutional neural network module both use the ReLU function as the optimizer and activation function of the model; Specifically, the mean absolute error (MAE), root mean square error (RMSE) and determination coefficient (R-Square, R²) are used to evaluate the model performance. The Adam optimizer is used to set the preset learning rate and the preset number of iterations to determine the final model parameters and obtain the final PM. 2.5 And the O3 prediction model, such as the formula:

[0019]

[0020]

[0021] in, is the true value, is the predicted value, is the average value corresponding to the true value; Specifically, the model parameters are continuously corrected and the model parameters are determined when the result prediction MAE value is the smallest; Specifically, the optimizer uses Adam optimizer, the activation function uses ReLU, Batchsize=512, the learning rate Lr=0.00001, and the number of iterations Epoch=5 to determine the final model parameters; Preferably, step S2 further comprises: evaluating the model using a ten-fold cross validation method; S3: Based on the test set, use the final PM 2.5 and O3 prediction model to obtain PM 2.5 Combined estimation results with O3.

[0022] In some embodiments, the PM 2.5 The combined estimation method with O3 can also be implemented in the following way.

[0023] like Figure 2 As shown, this embodiment collects the influencing factor data of a specified year period, PM 2.5 and O3 data, process the influencing factor data, divide the data set, build the CLDNN model and set up the shared parameter comparison test, train the model, train and tune the model parameters, and obtain PM 2.5 The model is estimated jointly with O3, and the test set data is finally input into the model to obtain PM 2.5 and O3 results; the specific steps are as follows: Step 100: Collect PM data from each monitoring site in 2019 2.5 and O3 observation data, satellite data, reanalysis data, and meteorological assistance data; Step 200: Process the data set and divide the data set into a training set, a validation set, and a test set; Step 300: Build a CLDNN model and set up a comparative experiment; Step 400: Use the training set and the validation set to train and validate the CLDNN model, adjust the model parameters according to the training results, and obtain PM 2.5 and O3 estimation model; Step 500: Input the test set data into the CLDNN model to obtain PM 2.5 Combined estimation results with O3; Specifically, in step S100, it is determined that the influencing factor data includes the hourly PM of 1641 stations in 2019. 2.5 and O3 data; used to estimate PM 2.5 The daily average aerosol optical depth (AOD) data comes from the High Resolution Air Pollutant Concentration Dataset (LGHAP); used for collaborative estimation of PM 2.5 The daily average NO2 data of O3 and NO2 are from ChinaHighNO2 dataset; the meteorological data are from the fifth generation European Centre for Medium-Range Weather Forecasts (ECMWF) Global Reanalysis of Climate and Weather (ERA5) dataset; Specifically, the main influencing factors of the ERA5 data determined in this embodiment include: relative humidity (RH), ozone mass mixing ratio (OR), temperature at 2 meters (T2M), east component of wind speed at 10 meters (U10), north component of wind speed at 10 meters (V10), surface pressure (SP), net surface solar radiation (SSR), net surface thermal radiation (STR), downward ultraviolet radiation (UVB), total ozone column (TCO3), planetary boundary layer height (BLH), evapotranspiration (E) and total precipitation (TP); wherein, relative humidity is calculated on the pressure level data at the pixel scale by considering the near-surface altitude; Specifically, regarding the data processing part: a 0.01° grid is created for data integration and model building; in order to ensure consistency in spatial and temporal scales, satellite data and meteorological reanalysis data are resampled to 0.01° in spatial scale by nearest-neighbor interpolation; in temporal scale, meteorological reanalysis data are resampled by daily average sliding window; the dataset is constructed starting from January 4, 2019, and the first three days are used to process time series information. For all pixels of each day, a two-dimensional spatial grid (7×7) is constructed with each pixel as the center, and then the station data and corresponding satellite data (AOD and NO2) that are not missing on the day and the previous three days are selected, so as to construct a three-dimensional space-time matrix (7×7×4) for each input feature; based on the 3-D space-time feature set, a collaborative estimation dataset is constructed using multi-source satellite data; specifically, PM 2.5 and O3 are two single-task datasets. Then, based on the similarities and differences between the two tasks, shared inputs and specific inputs of the two tasks are set to perform multi-task collaborative estimation; used for PM 2.5The data information of the estimation task includes 14 variables (7×7×4×14), including satellite data AOD, meteorological data U10, V10, T2M, PBLH, E, SP, TP, RH and 5 spatiotemporal data. The O3 estimation task adds 5 radiation-related meteorological variables that affect the photochemical reaction of ozone, including UVB, SSR, STR, TCO3 and OR. The satellite data used in the O3 estimation task is the NO2 product, one of the ozone precursors, with a total of 19 characteristic variables (7×7×4×19). Specifically, step S200 includes: performing normalization processing based on the influencing factor data set and dividing the data; wherein, for PM 2.5 The O3 dataset is divided into training set, validation set and test set, including: 2.5 The O3 data set is normalized and divided into training set, validation set and test set according to the preset division ratio; since different influencing factors have different ranges and time scales, it is necessary to normalize the data to achieve a reasonable prediction effect. The obtained data is normalized as follows, such as the formula:

[0024] in, represents the i-th normalized data, Represents the original PM 2.5 and the i-th data of O3 influencing factor data, Represents the original PM 2.5 and O3 influencing factor data, Indicates the maximum value, Indicates the minimum value; The normalized data is divided into training set, validation set and test set in a ratio of 7:1:2; Specifically, step S300 includes: using a CNN model to interpret spatial dependencies; using an LSTM model to extract temporal information; using a DNN model to integrate spatial and temporal features; Figure 3 The figure shows the verification result of the estimation accuracy, and the prediction results of the embodiment of the present invention are compared with the PM measured at the ground-based site. 2.5 Schematic diagram of the comparison results with O3; Specifically, the weighted average feature derived from the pollutant concentration collected around the target pixel has been proven to be effective in previous studies; however, due to the uneven distribution of stations, especially in areas with sparse stations, this additional feature will increase the uncertainty and complexity of the model accordingly, limiting its further application; in order to avoid this problem, this embodiment uses the features of the surrounding pixels instead of the station PM 2.5 The weighted features of are used to evaluate the contribution of each neighborhood to the estimated ozone level, as Figure 4 As shown in the figure; Convolutional neural network (CNN) is very suitable for interpreting these spatial dependencies because its neural layers can focus on specific areas (called receptive fields) in the input data; the final ozone concentration at a specific point is determined by the surrounding area rather than by nearby ground measurements, so the contribution of each neighborhood to the estimated ozone level can be evaluated; specifically, a batch normalization (Batch Normalization, BN) layer is added before the first convolution layer to replace the normalization of the original data; a BN layer is also added between the convolution function and the activation function (ReLU) to keep the data stably distributed before entering the next layer of the network, thereby speeding up model learning; Specifically, the CNN model used in this embodiment includes: 1 input layer → 1 batch normalization layer BN → 5 convolutional layers, where the input layer size is 7×7×4×14 or 7×7×4×19 (depending on whether it is PM 2.5 Or O3); Batch normalization layer BN is used to normalize the original data so that the data maintains a stable distribution before entering the next layer of the network, thereby accelerating model learning; the convolution kernel size of the 5 convolutional layers is 3×3, and the number of channels is 64-128-256-128-64, which is used to extract data spatial features; like Figure 5 As shown, in order to further utilize the spatiotemporal information of AOD and precursor NO2, this embodiment adds an LSTM module after the CNN module, takes the spatiotemporal three-dimensional data set as a sample, and extracts their spatiotemporal information at the same time; Specifically, a 3D pooling layer (size 3×3×1) is used to connect the CNN model to the next LSTM model. The purpose is to reduce the spatial size of the data (convert the data into 1×4 dimensions to facilitate the learning of the next LSTM model), reduce the computational complexity of the model, and retain important feature information to avoid the algorithm from falling into local features and causing overfitting. Specifically, the LSTM model adopts a two-layer LSTM structure; the number of parameters in each layer is 64, the output size of the first layer is 1×4, and the output size of the second layer is 1×1; the two-layer design is used to better obtain the time series characteristics of the data; Specifically, the LSTM model is linked to the next DNN model using a BN layer and a response function (ReLU); the BN layer is used to control the data distribution to make it more stable; ReLU is used as the optimizer and activation function of the model; like Figure 6As shown in the figure, the DNN model combines the feature-rich information passed down by the CNN and LSTM layers, which can effectively weigh the importance of different features and provide the final output; a DNN module with 64 DNN units is added after the CNN-LSTM module, and a BN layer is also added between the DNN layer and the activation function; Specifically, the DNN model uses a DNN with 64 hidden layers to increase the PM 2.5 The nonlinear fitting effect of pollutants such as O3 and others; the ReLU function is also used as the optimizer and activation function of the model; like Figure 7 , Figure 8 and Fig. 9 As shown, this embodiment sets up a shared input and shared weight experiment, that is, designs three collaborative estimation model constraint schemes, including: (1) Hard Parameters Sharing A:PM 2.5 The prediction task of PM shares a CNN module with O3, but has different LSTM and DNN modules; specifically, PM 2.5 The inputs of the two prediction tasks O1 and O2 share the same parameters in the same CNN module, and then are output to two different LSTM modules, and finally sent to their respective DNN modules for estimation; Figure 7 It is a schematic diagram of the CLDNN model shared parameter comparison test scheme (a); (2) Hard Parameters Sharing B:PM 2.5 The prediction tasks of and O3 share a CNN module and a LSTM module, but have different DNN modules; specifically, the inputs of the two tasks share the same parameters in a CNN module, then are output to a shared LSTM module, and finally sent to two different DNN modules for final estimation; Figure 8 is a schematic diagram of the CLDNN model shared parameter comparison test scheme (b); (3) Soft Parameters Sharing:PM 2.5 The prediction tasks of PM and O3 use different CNN, LSTM and DNN modules, but allow data interaction between modules; specifically, the inputs of the two tasks are sent to different CLDNN models, each task has its own model and parameters, and these model parameters are mutually constrained; that is, in PM 2.5 In the two subtasks of O3, PM 2.5 The output of the selected module in the task is merged with the output of the corresponding module in the O3 task and transmitted to PM 2.5 The subsequent network module of the task; the output of the same module in the O3 task is combined with the PM 2.5The output of the task module is superimposed and placed in the next module of the O3 task, considering PM 2.5 Impact on O3, such as Fig. 9 (c) is a schematic diagram of the CLDNN model shared parameter comparison test plan.

[0025] As shown in Table 1, the comparison results of this embodiment and other studies are provided to further evaluate the CLDNN model of this embodiment in PM 2.5 By making full use of the correlation between pollutants, the joint estimation model can improve the performance of the two subtasks, especially PM 2.5 Estimation task, R 2 From 0.904 to 0.915. In addition, the proposed joint estimation method outperforms the current best knowledge information neural network joint estimation method in both subtasks, confirming the efficiency of the designed joint estimation method.

[0026] Table 1: Comparison results of CLDNN model with other models

[0027] Based on the above embodiment, the CLDNN model is trained and verified using the training set and the verification set, and the model parameters are adjusted according to the training results to obtain PM. 2.5 Combined with O3 estimation model, including: using mean absolute error (MAE), root mean square error (RMSE) and R-Square (R²) to evaluate model performance, using Adam optimizer, setting preset learning rate and preset number of iterations, determining final model parameters, and outputting PM 2.5 and O3 estimation model.

[0028] Specifically, this embodiment trains the model through CLDNN model parameters, and adjusts the parameters according to the training results to determine the final model parameters.

[0029] Calculated by MAE, RMSE and R² as follows:

[0030]

[0031]

[0032] in, is the true value, is the predicted value, is the average value corresponding to the true value.

[0033] The model parameters are continuously corrected. When the predicted MAE value is the smallest, the model parameters are determined: Adam optimizer, ReLU activation function, Batchsize=512, learning rate Lr=0.00001, Epoch number=5, and the final model parameters are determined.

[0034] Based on the above embodiment, as a preferred embodiment of the CLDNN model estimation method provided in this embodiment, it also includes: using a ten-fold cross validation method to evaluate the model, based on samples, based on sites and time. The expected error indicates that the data whose error between the estimated value and the observed value falls within the expected error range accounts for 68% of the total data volume.

[0035] This embodiment provides a PM based on the CLDNN model. 2.5 The O3 joint estimation system includes: an acquisition module for collecting data information for a preset year time period; a construction module for building a CLDNN model and performing shared input and shared weight experiments based on the model; a partitioning module for partitioning the new typhoon data set to obtain a training set, a validation set, and a test set; a training module for training and validating the CLDNN model using the training set and the validation set, adjusting the model parameters according to the training results, and obtaining PM 2.5 and O3 estimation model; estimation module, used to input the test set data into PM 2.5 and O3 estimation model to obtain PM 2.5 and O3 estimation results; specifically, the system structure is as follows Fig.10 As shown, the system includes: an acquisition module 1101, a construction module 1102, a division module 1103, a training module 1104 and an estimation module 1105, wherein: the acquisition module 1101 is used to collect data information of a preset year time period; the construction module 1102 is used to build a CLDNN model, and conduct shared input and shared weight experiments based on the model; the division module 1103 is used to divide the new typhoon data set to obtain a training set, a validation set and a test set; the training module 1104 uses the training set and the validation set to train and verify the CLDNN model, adjusts the model parameters according to the training results, and obtains PM 2.5 and O3 estimation model; the estimation module 1105 is used to input the data of the test set into PM 2.5 and O3 estimation model to obtain PM 2.5 and O3 estimation results.

[0036] This embodiment provides a computer-readable storage medium having a computer program stored thereon, which implements the PM described above when the computer program is executed by a processor. 2.5The steps of the combined estimation method of O3. The storage medium may be a magnetic disk, an optical disk, a read-only memory (ROM), a random access memory (RAM), a flash memory, a hard disk drive (HDD) or a solid-state drive (SSD), etc. The storage medium may also include a combination of the above-mentioned types of memory.

[0037] This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the above-mentioned PM is implemented. 2.5 and O3 joint estimation method steps.

[0038] like Fig.11 As shown, the computer device may include: at least one processor 121, such as a CPU (Central Processing Unit), at least one communication interface 123, a memory 124, and at least one communication bus 122. The communication bus 122 is used to realize the connection and communication between these components. The communication interface 123 may include a display screen (Display), a keyboard (Keyboard), and the optional communication interface 123 may also include a standard wired interface and a wireless interface. The memory 124 may be a high-speed RAM memory (Random Access Memory, volatile random access memory), or a non-volatile memory (non-volatile memory), such as at least one disk memory. The memory 124 may also be at least one storage device located away from the aforementioned processor 121. The memory 124 stores an application program, and the processor 121 calls the program code stored in the memory 124 to perform any of the above method steps. The communication bus 122 may be a peripheral component interconnect standard (PCI) bus or an extended industry standard architecture (EISA) bus, etc. The communication bus 122 can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Fig.11Only one line is used to represent it, but it does not mean that there is only one bus or one type of bus. Among them, the memory 124 may include a volatile memory (English: volatile memory), such as a random access memory (English: random-access memory, abbreviated: RAM); the memory may also include a non-volatile memory (English: non-volatile memory), such as a flash memory (English: flash memory), a hard disk (English: hard disk drive, abbreviated: HDD) or a solid-state drive (English: solid-state drive, abbreviated: SSD); the memory 124 may also include a combination of the above-mentioned types of memory. Among them, the processor 121 may be a central processing unit (English: central processing unit, abbreviated: CPU), a network processor (English: network processor, abbreviated: NP) or a combination of CPU and NP. Among them, the processor 121 may further include a hardware chip. The above-mentioned hardware chip may be an application-specific integrated circuit (English: application-specific integrated circuit, abbreviated: ASIC), a programmable logic device (English: programmable logic device, abbreviated: PLD) or a combination thereof. The PLD can be a complex programmable logic device (CPLD), a field-programmable gate array (FPGA), a generic array logic (GAL), or any combination thereof. Optionally, the memory 124 is also used to store program instructions. The processor 121 can call program instructions to implement the PLD of the present embodiment. 2.5 and O3 joint estimation method.

[0039] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation modes, which are merely illustrative rather than restrictive. Under the guidance of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the present invention and the claims, all of which are within the protection of the present invention.

Claims

1. A PM 2.5 and O3 joint estimation method, characterized in that, The following steps are involved: S1: Get PM 2.5 and O3 influencing factors data, the PM 2.5 Preprocess the O3 influencing factor data to obtain the training set, validation set and test set; S2: Build PM 2.5 and O3 prediction model, using the training set and the validation set to 2.5 The final PM 2.5 and O3 prediction model; S3: Based on the test set, using the final PM 2.5 and O3 prediction model to obtain PM 2.5 Combined estimation results with O3.

2. The PM according to claim 1 2.5 and O3 joint estimation method, characterized in that, The PM 2.5 and O3 influencing factors including PM 2.5 and O3 data, daily average aerosol optical depth data, daily average NO2 data and global climate and weather reanalysis data sets; the global climate and weather reanalysis data sets include relative humidity, ozone mass mixing ratio, 2-meter height temperature, east component of wind speed at 10 meters height, north component of wind speed at 10 meters height, surface air pressure, surface net solar radiation, surface net thermal radiation, surface downward ultraviolet radiation, total ozone column, planetary boundary layer height, evapotranspiration and total precipitation.

3. The PM according to claim 1 2.5 and O3 joint estimation method, characterized in that, Step S1 specifically includes: 2.5 The O3 influencing factor data is normalized to obtain normalized data; the normalized data is divided according to a preset ratio to obtain a training set, a validation set and a test set.

4. The PM according to claim 1 2.5 and O3 joint estimation method, characterized in that, The PM 2.5 and O3 prediction models including PM 2.5 task model and O3 task model; the O3 task model includes a first convolutional neural network module, a first long short-term memory module and a first deep convolutional neural network module connected in sequence; the PM 2.5 The task model includes a second convolutional neural network module, a second long short-term memory module and a second deep convolutional neural network module.

5. The PM according to claim 4 2.5 and O3 joint estimation method, characterized in that, The first convolutional neural network module and the second convolutional neural network module share the same parameters.

6. The PM according to claim 4 2.5 and O3 joint estimation method, characterized in that, The first convolutional neural network module and the second convolutional neural network module share the same parameters; the first long short-term memory module and the second long short-term memory module share the same parameters.

7. The PM according to claim 4 2.5 and O3 joint estimation method, characterized in that, The output of the second convolutional neural network module is combined with the output of the first convolutional neural network module as the input of the second long short-term memory module, and the output of the second long short-term memory module is combined with the output of the first long short-term memory module as the input of the second deep convolutional neural network module; the output of the first convolutional neural network module is superimposed with the output of the second convolutional neural network module as the input of the first long short-term memory module, and the output of the first long short-term memory module is superimposed with the output of the second long short-term memory module as the input of the first deep convolutional neural network module.

8. The PM according to claim 4 2.5 and O3 joint estimation method, characterized in that, The first convolutional neural network module and the second convolutional neural network module are used to explain spatial dependence; the first convolutional neural network module and the second convolutional neural network module each include 1 input layer, 1 batch normalization layer and 5 convolutional layers; wherein the input layer size of the first convolutional neural network module is 7×7×4×19, and the input layer size of the second convolutional neural network module is 7×7×4×14; the batch normalization layer is used to normalize the original data so that the data maintains a stable distribution before entering the next layer of the network, thereby accelerating model learning; the convolution kernel size of the 5 convolutional layers is 3×3, and the number of channels is 64, 128, 256, 128 and 64 respectively, and the 5 convolutional layers are used to extract data spatial features; The first convolutional neural network module is connected to the first long short-term memory module through a 3D pooling layer, and the second convolutional neural network module is connected to the second long short-term memory module through a 3D pooling layer. The size of the 3D pooling layer is 3×3×1. The 3D pooling layer is used to reduce the spatial size of the data and the computational complexity of the model while retaining important feature information to avoid the algorithm from falling into local features and causing overfitting; The first long short-term memory module and the second long short-term memory module are used to extract time information, and the first long short-term memory module and the second long short-term memory module both adopt a double-layer LSTM structure, and the number of parameters of each layer of LSTM is 64, wherein the output size of the first layer is 1×4, and the output size of the second layer is 1×1; The first long short-term memory module is connected to the first deep convolutional neural network module through a batch normalization layer and a response function ReLU, and the second long short-term memory module is connected to the second deep convolutional neural network module through a batch normalization layer and a response function ReLU, wherein the batch normalization layer is used to control data distribution, and the response function ReLU is used as an optimizer and activation function for modeling; The first deep convolutional neural network module and the second deep convolutional neural network module are used to integrate spatial and temporal features and increase the nonlinear fitting effect of pollutants. The first deep convolutional neural network module and the second deep convolutional neural network module both adopt a deep convolutional neural network with 64 hidden layers. The first deep convolutional neural network module and the second deep convolutional neural network module both use the ReLU function as the optimizer and activation function of the model.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the PM described in any one of claims 1 to 8 is implemented. 2.5 and O3 joint estimation method steps.

10. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the PM according to any one of claims 1 to 8 is implemented 2.5 and O3 joint estimation method steps.

Images