An atmospheric pollutant concentration prediction method based on ICLU-CWGAN model

Abstract

The present application belongs to the field of artificial intelligence and environmental monitoring technology, and particularly relates to an atmospheric pollutant concentration prediction method based on an ICLU-CWGAN model. By obtaining and preprocessing atmospheric pollutant concentration, meteorological and other data, the Inception multi-scale convolution idea is fused with the ConvLSTM network to construct a multi-scale ConvLSTM network; under the condition of the gradient-penalized Wassertein generative adversarial network framework, the multi-scale ConvLSTM is taken as a generator, and the improved U-Net is taken as a discriminator to construct an ICLU-CWGAN model; the model training is completed by using the pretreated training grid data to predict the atmospheric pollutant concentration. The present application can solve the problems of weak modeling capability for long-time sequence complex space-time correlation, spatial distortion and unstable training in the prior art, thereby improving the accuracy and stability of the atmospheric pollutant concentration space-time distribution prediction.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and environmental monitoring technology, specifically relating to an atmospheric pollutant concentration prediction method based on the ICLU-CWGAN model, which aims to improve the accuracy of intelligent prediction. Background Technology

[0002] Air pollution has become a major global environmental and public health problem. Accurately predicting the spatiotemporal changes in pollutant concentrations is crucial for implementing effective pollution control measures and issuing public health warnings. Existing methods for predicting atmospheric pollutant concentrations mainly include physical models, statistical models, traditional machine learning models, and deep learning models. Physical diffusion models (such as WRF-Chem) are based on atmospheric dynamics principles, but their computational complexity is extremely high, heavily relying on pollutant emission inventories that are difficult to obtain accurately, and dynamic calibration is challenging. Mathematical statistical models (such as ARIMA and geographic weighted regression), while computationally simple, cannot effectively capture the nonlinear spatiotemporal interaction characteristics under the coupling of meteorological and topographical factors, and have poor generalization ability. Traditional machine learning models (such as random forests and XGBoost), while capable of handling nonlinear relationships, still fall short in characterizing the spatiotemporal dynamics of pollutant diffusion and transport. In recent years, deep learning models have demonstrated significant advantages in spatiotemporal prediction tasks due to their powerful nonlinear modeling capabilities. Models such as Convolutional Neural Networks (CNN), Long Short-Term Memory Networks (LSTM), and their combinations (such as ConvLSTM and CNN-LSTM) have been applied to predicting atmospheric pollutant concentrations, exhibiting superior predictive performance compared to traditional methods. However, these advanced deep learning models still face three main challenges when applied to high-precision, gridded spatiotemporal concentration field prediction:

[0003] (1) Weak ability to model complex spatiotemporal correlations over long time: Existing models lack adaptive mechanisms to simultaneously capture the multi-scale coupling effect between local pollution events and regional diffusion processes, and are difficult to effectively capture and maintain spatiotemporal dependence features over long time spans, with obvious memory degradation.

[0004] (2) Spatial detail blurring and topological distortion in the generated prediction results: Existing models are prone to boundary blurring in pixel-level prediction, making it difficult to accurately reconstruct complex spatial topological structures (such as concentration gradients, extreme regions, and administrative boundaries). This results in significant differences between the predicted results and the actual distribution in terms of spatial details.

[0005] (3) Dynamic instability of model training and gradient signal failure: Under the adversarial training framework, the discriminator is prone to converge to the ideal state too early, resulting in gradient vanishing or saturation, which makes the generator unable to obtain effective learning signals, and the training process becomes unstable or even stagnant.

[0006] Therefore, there is an urgent need in this field for a new method that can overcome the above-mentioned defects, achieve high accuracy, high stability, and effectively preserve spatial details for long-term spatiotemporal prediction of atmospheric pollutant concentrations. Summary of the Invention

[0007] To address the problems existing in the prior art, this invention provides a method for predicting atmospheric pollutant concentrations based on the ICLU-CWGAN model, comprising the following steps:

[0008] S1. Acquire pollutant concentration data and meteorological data from monitoring stations in the target area, perform preprocessing, and create grid data for model training and prediction of pollutant concentration.

[0009] S2. Integrate the multi-scale convolution algorithm in Inception with ConvLSTM to construct a multi-scale ConvLSTM network;

[0010] S3. Based on the Conditional Wasserstein Generative Adversarial Network (CWGAN-GP) framework with gradient penalty, the generator uses a multi-scale ConvLSTM network and the discriminator uses an improved U-Net network to construct the ICLU-CWGAN model.

[0011] S4. Use the training concentration grid data obtained from step S1 to train the ICLU-CWGAN model;

[0012] S5. Apply the trained ICLU-CWGAN prediction model to predict the concentration of air pollutants in the target area over a future period of time.

[0013] Furthermore, the multi-scale ConvLSTM network hierarchical processing design scheme in step S2:

[0014] (1) To address the differential diffusion characteristics of pollutant concentration grid data at different time distances, a hierarchical processing scheme is designed. Small-scale convolution, batch normalization and residual connection are used to capture local details in the adjacent prediction period, medium-scale convolution, group normalization and residual connection are used to extract regional features in the middle period, and large-scale convolution, instance normalization and residual connection are used to capture macroscopic distribution patterns in the far period.

[0015] (2) The gating mechanism is used to regulate the information flow. The input gate determines the degree of retention of new features through the activation coefficient, the forget gate filters the historical information to be forgotten with the weight coefficient, and the output gate dynamically controls the contribution of the current cell state to the final output.

[0016] Furthermore, the following improvements are made to the discriminator U-Net network:

[0017] (1) Apply spectral normalization to the weights of each convolutional kernel and remove all in-layer convolutional kernels. The activation function ensures that each layer of the discriminator satisfies Lipschitz continuity.

[0018] (2) Introduce binary geographic awareness masks into the feature maps of each layer in the network structure, add feature extraction guidance in the skip connections, and dynamically adjust the fusion ratio of encoder and decoder features through channel attention weights;

[0019] (3) Add instance normalization to the shallow layer of the encoder;

[0020] (4) Perform the following steps on the feature map of the last layer: Convolution transformation.

[0021] Furthermore, the ICLU-CWGAN model described in step 4 is trained through the following steps:

[0022] (1) Historical PM 2.5 The concentration grid sequence is input into a multi-scale ConvLSTM network, which extracts spatiotemporal features step by step and outputs a predicted concentration grid sequence.

[0023] (2) Input the concentration distribution map output by the multi-scale ConvLSTM and the corresponding real observation data at the same time step into the improved U-Net network to evaluate the spatial authenticity of the generated data;

[0024] (3) The discriminator error and the generator error propagate back and update the parameters, forming a continuously iterative adversarial competitive game state. In the end, the generator can generate a pollutant concentration grid sequence that is difficult for the discriminator to judge as true or false.

[0025] The present invention has the following beneficial effects:

[0026] (1) This invention, based on the Conditional Wasserstein Generative Adversarial Network (CWGAN-GP) framework with gradient penalty, introduces geographic-aware mask constraints to both the generator and discriminator to optimize model performance. The generator employs a multi-scale ConvLSTM structure, enhancing its ability to extract multi-scale spatiotemporal features while effectively suppressing boundary blurring in the predicted image. The discriminator is designed as a U-Net network based on a dual attention mechanism, where the spatial attention module improves the accuracy of pixel-level feature extraction, and the channel attention module dynamically adjusts the feature fusion ratio between the encoder and decoder, reducing redundant information transmission. This structure effectively alleviates the problem of premature convergence in pixel-level training of CWGAN-GP.

[0027] (2) The loss function of this invention integrates adversarial loss, feature matching loss and spatiotemporal smoothing constraint to optimize gradient propagation efficiency. At the same time, it uses the historical pollutant grid sequence as the sliding window condition and uses the multi-level features of the discriminator to iteratively correct the generated sequence, which can suppress the recursive propagation of the initial prediction error and effectively alleviate the long-term memory degradation problem.

[0028] (3) The present invention can accurately predict the concentration of air pollutants in the target city in the future, providing a basis for decision-making for the prevention and control of air pollution and government governance. Attached Figure Description

[0029] Figure 1 This is a flowchart of the atmospheric pollutant concentration prediction method based on the ICLU-CWGAN model described in this invention;

[0030] Figure 2 This invention relates to a multi-scale convolutional structure design scheme for a multi-scale ConvLSTM network generator.

[0031] Figure 3 This is a diagram of the multi-scale ConvLSTM network structure of the generator in this invention;

[0032] Figure 4 This is an improved diagram of the U-Net network structure of the discriminator of this invention;

[0033] Figure 5 These are the PMs in the embodiments of the present invention. 2.5 A graph showing the fit between the average observed values ​​and the predicted values ​​at the monitoring stations. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments in this specification without creative effort are within the scope of protection of this application.

[0035] A method for predicting atmospheric pollutant concentrations based on the ICLU-CWGAN model includes the following steps, as shown in the appendix. Figure 1 As shown:

[0036] S1: Acquire pollutant concentration data, meteorological data, and satellite remote sensing aerosol optical thickness data from monitoring stations in the target area, and perform preprocessing to create grid data for model training and model prediction;

[0037] The pollutant concentration data includes PM2.5 concentrations at N stations in the target area. 2.5The concentrations of NO2, SO2, and O3, and street-level pollutant concentration data monitored by ground-based mobile monitoring vehicles; the meteorological data include temperature, air pressure, rainfall, wind direction, and wind speed.

[0038] The preprocessing steps include:

[0039] (1) The initial concentration field is generated by integrating satellite remote sensing aerosol optical thickness data with ground monitoring data using a Bayesian model;

[0040] (2) Spatial interpolation was performed using the inverse distance weighting method to construct grid data of pollutant concentrations;

[0041] (3) Use Kalman filtering to collaboratively update the grid data and output a spatiotemporally continuous concentration grid map;

[0042] (4) Geographic constraints are applied to the initial concentration grid using geographic information system spatial analysis technology, and grid data of the disturbed areas (such as the estuary and the main stream and tributary waters of the river) are spatially cropped.

[0043] (5) Standardize the grid data after spatial clipping, i.e. perform Box-Cox transformation to reconstruct the concentration grid data into a normal state, and then use bilinear interpolation guided by the anisotropic diffusion equation to repair the missing data and the abnormal data detected by the flexible detection.

[0044] The concentration grid data is reconstructed using a Box-Cox transformation according to the following formula:

[0045]

[0046] In the formula, These are the original pollutant concentration observations, in μg / m³. To transform the parameters, the maximum likelihood estimation is used to iteratively optimize within the interval [-2, 2]. The target variable is the one that, after transformation, follows a quasi-normal distribution.

[0047] Bilinear interpolation guided by anisotropic diffusion equations is used to repair missing data and anomalous data detected by flexible detection. The calculation formula is as follows. In the outlier control phase, a dynamic threshold strategy is adopted, setting PM2.5 limits based on the classification limits in the "Technical Regulations for Ambient Air Quality Index (AQI)". 2.5 >500μg / m 3 O3 > 1200 μg / m 3 SO2 > 800 µg / m 3 NO2 > 3840 µg / m 3 CO > 150 mg / m³ 3 It employs rigid filtering conditions while simultaneously performing flexible detection based on the Local Outlier Factor (LOF).

[0048]

[0049] In the formula, The weighting coefficients are calculated by quantifying the spatial autocorrelation using a semi-variogram function, and satisfy the following conditions: Furthermore, it is negatively correlated with spatial distance; This represents the spatial coordinates of the four valid grid cells surrounding the missing point. To standardize the concentration data using a sliding window Z-score based on the rolling mean and rolling variance of historical data, the concentration data values ​​of the four nearest grid cells are used.

[0050] S2 integrates the multi-scale convolution algorithm from Inception with ConvLSTM, and constructs a multi-scale ConvLSTM network through the following hierarchical processing design:

[0051] (1) Perform multi-scale convolution on the input data to generate a feature matrix.

[0052] To address the differential diffusion characteristics of pollutant concentration grid data across different time intervals, a hierarchical processing scheme is designed: adjacent prediction time periods are processed using... Convolution, batch normalization, and residual connections capture local details, while intermediate time intervals utilize... Convolution, group normalization, and residual connections are used to extract regional features, while features in distant time periods are extracted through... Convolution, instance normalization, and residual connections capture macroscopic distribution patterns, as shown in the appendix. Figure 2 As shown.

[0053] (2) The gating mechanism is used to regulate the information flow. The input gate determines the degree of retention of new features through the activation coefficient, the forgetting gate filters the historical information to be forgotten through the weight coefficient, and the output gate dynamically controls the contribution of the current cell state to the final output, such as... Figure 3 As shown in the figure Represents initialization state and time step Input; Represents initialization state and time step The hidden state; Represents initialization state and time step cellular state; Represents initialization state and time step The convolution operation; σ is the Sigmoid activation function; These are activation functions, representing the modified linear unit and the parameterized linear unit, respectively. ; Yes Do Activate; ⊙ represents the Hadamard product, which is element-wise multiplication; + represents element-wise addition.

[0054] Multi-scale convolution kernels are calculated using the following formula:

[0055]

[0056] In the formula, Calculate dimensions from model input Convolution operation, It is the size of the convolution kernel; Different convolution kernel sizes Corresponding weights; It is a convolution operation; It is a bias term; It is the input feature map.

[0057] Secondly, the output of the forget gate is determined based on different convolution kernels. The output of the input gate Output of the output gate Current time step Candidate cell state Current cell state and the hidden state of the current time step Calculation and Update:

[0058]

[0059] In the formula, It is a time step Model input; It is the current time step; It was a step from a previous time; , and These represent small, medium, and large convolutional kernels, respectively. It is the hidden state of the previous time step; This refers to the Sigmoid activation function; It is the previous time step cellular state; This represents the Hadamard product, which is the product of elements. These are activation functions, representing the modified linear unit and the parameterized linear unit, respectively. ; Yes Do activation.

[0060] S3 is based on the conditional Wasserstein generative adversarial network framework with gradient penalty. The generator uses a multi-scale ConvLSTM network, and the discriminator uses an improved U-Net network to construct the ICLU-CWGAN model.

[0061] The discriminator U-Net network is improved as follows:

[0062] (1) Apply spectral normalization to the weights of each convolutional kernel and remove all in-layer convolutional kernels. The activation function ensures that each layer of the discriminator satisfies Lipschitz continuity. The spectral normalized weights can be calculated using the following formula:

[0063]

[0064] In the formula, These are the original convolutional kernel weights. ( ) is the spectral norm of the weight matrix, which is approximated by the power iteration method; The weights are obtained after spectral normalization. For spectrum normalization;

[0065] (2) Given that the mesh graph dataset only contains valid information in the target area, a spatial attention mechanism with binary geographic-aware mask constraints is introduced into the feature maps of each network layer to enhance the pixel-level feature extraction accuracy of skip connections and convolutional layers. Simultaneously, a feature guidance mechanism is deployed in the skip connection path, dynamically adjusting the fusion ratio of encoder and decoder features through channel attention weights to effectively suppress redundant information transmission. Channel attention weights The output feature map after adaptive weighted fusion. The calculation formula is as follows:

[0066]

[0067] In the formula, These are channel attention weights, ranging from [0,1]. This indicates activation via the Sigmoid function. Generate attention weights between 0 and 1; It is a corrected linear unit; , It is a 1×1 convolution weight matrix; It is a binary geographic sensing mask matrix, with the target area set to 1 and the rest to 0. and It refers to the height, number of rows, width, or number of columns of the feature map; These are encoder features. It is a decoder feature. This indicates the concatenation of channels for encoder and decoder features;

[0068] (3) Add instance normalization to the shallow layer of the encoder, the mathematical definition of which is:

[0069]

[0070]

[0071]

[0072] In the formula, It is the first instance after normalization. The sample, the first Each channel, location eigenvalues; For the learnable first Each channel scaling parameter; For the first instance before normalization The sample, the first Each channel, location eigenvalues; , These represent the channel mean and channel variance. For the learnable first Each channel translation parameter This is a numerical stability constant to prevent the denominator from being zero; and It refers to the height, number of rows, width, or number of columns of the feature map.

[0073] (4) Perform the following steps on the feature map of the last layer: Convolutional transformation allows the U-Net network to directly output the original score map, preventing the discriminator from failing to accurately represent the difference between real and generated data, and providing the generator with a continuous and stable gradient signal. With this improvement, the generator can obtain a smoother gradient through backpropagation, and the discriminator can better represent the difference between real and generated data. The last linear layer output It can be calculated using the following formula:

[0074]

[0075] In the formula, The weights are those of a 1×1 convolution kernel. These represent the kernel height, kernel width, number of input feature map channels, and number of output feature map channels, respectively. This is a convolution operation; This is the feature map output by the last convolutional layer of U-Net; This is a bias term.

[0076] The discriminant and generator loss functions of the generative adversarial network are as follows:

[0077] Discriminator loss function

[0078] Generator loss function

[0079] In the formula, It is the discriminator that generates the data. The output value, It is the discriminator's judgment of real samples The output value; It generates data distribution Expected rating It is the actual data distribution The expected score, the difference between the two is the Wasserstein distance estimate between the generated data distribution and the real data distribution; It is the gradient penalty coefficient; It is a gradient penalty term; Weight coefficients of feature matching loss; It is feature matching loss. It is a spacetime smoothing term. The weight coefficients of the feature matching loss, It is the weighting coefficient of the spatiotemporal smoothing term.

[0080] The feature matching loss, spatiotemporal smoothing term, and random interpolation between real and generated data can be expressed as:

[0081]

[0082]

[0083]

[0084] In the formula, For real data After inputting the improved discriminator U-Net network, the first Feature map of the layer; It is a generator The corresponding number Feature map of the layer; It is a hierarchical weighting system; yes Norm; Indicates the generator at the 1st... Output at each time step; It is real data. and generating data Random interpolation between; It is a random variable that follows a uniform distribution in the range [0,1].

[0085] S4. Using the training concentration grid data obtained from step S1, train the ICLU-CWGAN model. The model training is performed through the following steps, as shown in the appendix. Figure 1 As shown:

[0086] (1) Input the historical pollutant concentration grid sequence into a multi-scale ConvLSTM, extract the spatiotemporal features step by step, and output the predicted concentration grid sequence;

[0087] (2) Input the concentration distribution map output by the multi-scale ConvLSTM and the corresponding real observation data at the same time step into the U-Net network to evaluate the spatial authenticity of the generated data;

[0088] (3) The discriminator error and the generator error propagate back and update the parameters, forming a continuously iterative adversarial competitive game state, making it difficult for the generator to generate pollutant concentration grid sequences that the discriminator can judge as true or false.

[0089] S5, apply the trained ICLU-CWGAN prediction model to predict the concentration of air pollutants in the target area over a future period. Taking City A as an example, the prediction is performed as follows:

[0090] City A, a renowned manufacturing city, relies heavily on industrial emissions (electronics manufacturing, plastics processing, etc.), vehicle exhaust, and dust as its main sources of pollution. The primary air pollutants involved include fine particulate matter (PM2.5). 2.5 The data includes ozone (O3), nitrogen dioxide (NO2), sulfur dioxide (SO2), and carbon monoxide (CO). The data primarily consists of four sources: pollutant concentrations from 36 monitoring stations within City A and meteorological observation data (including wind direction, wind speed, rainfall, and solar radiation density). PM2.5 was selected as the primary air pollutant. 2.5 As an example, the ICLU-CWGAN model of the present invention is trained and predicted. Through... Figure 1 Methods and processes for acquiring grid map data, resulting in PM 2.5 Each image in the grid map is 50×75 pixels with 3750 pixels. The value of each pixel is the near-ground atmospheric pollutant concentration within that grid area. The actual area of ​​each grid is approximately 1 km². 2 PM 2.5 The concentration data spans from 00:00 on October 1, 2021 to 22:00 on May 7, 2022, with each time interval being 1 hour.

[0091] The ICLU-CWGAN model training configuration uses 70% of the data as the training set, 20% as the validation set, and 10% as the test set to ensure the independence of model training, tuning, and evaluation. During model training, an early stopping strategy dynamically controls the training cycle; training terminates when the validation set loss does not decrease for 10 consecutive epochs to avoid overfitting. The network structure weights are initialized using the Xavier method, which matches the characteristics of the activation function, ensuring gradient stability during forward and backward propagation. The optimizer uses the Adam algorithm (initial learning rate set to 0.0005), dynamically adjusting the parameter update step size through adaptive momentum estimation, and introducing Dropout regularization (dropout rate = 0.2) to randomly mask neurons to improve the model's generalization ability.

[0092] Selection of time sliding window length: Comparing the training effects of the ICLU-CWGAN model with time sliding windows of 5h, 10h, 15h, 20h, 25h, 30h, 35h, and 40h, it can be seen that the 12h prediction space accuracy with a sliding window of 20h is the highest, the temporal information is relatively completely preserved, and the overfitting phenomenon is the least.

[0093] The training parameters of the generator multi-scale ConvLSTM network are shown in Table 1. f is the kernel size, s is the kernel stride, p is the padding size, d is the number of kernels in the layer, seq_len is the number of historical time steps of the input, and pre_len is the number of predicted time steps of the output.

[0094] Table 1. Training parameters of the generator multi-scale ConvLSTM network

[0095]

[0096] Based on this, to comprehensively evaluate the model's capabilities and compare it with other predictive models on the same dataset, the root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²) were used. 2 RMSE and MAE are used as metrics to measure the effectiveness of a method. The smaller the RMSE and MAE values, the higher the model accuracy and R². 2 The closer the value is to 1, the stronger the model's explanatory power. The formula for calculating the index is as follows:

[0097]

[0098]

[0099]

[0100] In the formula, It is the number of test set samples; , This refers to the grid dimensions (number of rows and columns). It is the total number of time steps in the test set; yes Time step position , The true value; yes Time step position , The predicted value; yes Time step observation grid mean; yes Time step prediction grid mean; It is the average of the average values ​​of all observed grids.

[0101] Tables 2 and 3 respectively show the multi-timescale PM performance of the ICLU-CWGAN model, baseline model, and advanced model across the entire test set. 2.5 Concentration prediction performance. Table 2 compares the performance of this model with all baseline models on short-term prediction tasks, using historical 7-hour PM2.5 concentration prediction data. 2.5 Using a concentration grid map as input, predict PM2.5 levels in the target city for the next 3 hours. 2.5 Concentration grid plot. As shown in Table 2, the ICLU-CWGAN model has an RMSE of 2.77, which is 40.93% lower than the ConvLSTM model. 2 The score was 5.48, a 14.10% improvement over the CNN-LSTM model, indicating that the model's short-term (3-hour) prediction performance significantly outperformed mainstream baseline models. Table 3 compares the performance of this model with three advanced models on long-term prediction tasks, using historical 8-hour PM data. 2.5 Using a concentration grid map as input, predict PM2.5 levels in the target city for the next 12 hours. 2.5 Concentration grid diagram. Table 3 shows the performance indicators RMSE, MAE, and R... 2 The scores were 4.61, 6.42, and 0.80, respectively, all outperforming the three advanced models ConvLSTM, CNN-LSTM, and ConvGRU. Compared with the 3-hour prediction performance metrics, the RMSE decreased by only 1.84, while the other three advanced models decreased by at least 4.23 (ConvLSTM), indicating that the ICLU-CWGAN model can alleviate the problem of long-term memory degradation.

[0102] Table 2 Comparison of 3-hour prediction performance between ICLU-CWGAN and mainstream models

[0103]

[0104] Table 3. Performance Comparison of ICLU-CWGAN and Advanced Models in 12-Hour Prediction

[0105]

[0106] Figure 5 The results demonstrate the generalization ability of the ICLU-CWGAN model for heterogeneous geographic units, evaluated using a 72-hour validation set based on 12 monitoring stations in City A. The ICLU-CWGAN regional concentration prediction bias remained stable within ±3 μg / m³, with an overall bias rate of less than 15%, representing a 57% improvement in spatial prediction accuracy compared to the ±7 μg / m³ fluctuation range observed in the ConvLSTM test during the same period. Particularly in high concentration gradient regions, such as A7 with an observed value of 35.7 μg / m³ and a predicted value of 36.9 μg / m³, the prediction error rate remained stable within 3.4%, significantly better than the typical error rate of 12%-18% for the ConvLSTM model in such regions. The prediction-observation scatter plot showed R²=0.74 and RMSE=4.61 μg / m³, further confirming strong correlation and the absence of systematic shifts, validating the enhancing effect of the generative adversarial mechanism on the spatiotemporal decoupling capability of ConvLSTM.

[0107] Table 4 PM2.5 levels at various stations in City A generated by the ICLU-CWGAN model 2.5 Concentration comparison

[0108]

[0109] Those skilled in the art should understand that the above embodiments are merely illustrative and are not intended to imply that the scope of the invention is limited to these examples. Within the framework of this invention, technical features of the above embodiments or different embodiments can be combined, steps can be implemented in any order, and many other variations of the different aspects of the invention as described above exist, which are not provided in detail for the sake of brevity. Any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for predicting atmospheric pollutant concentration based on ICLU-CWGAN model, characterized in that, The method comprises the following steps: S1, obtaining target area monitoring station pollutant concentration data, meteorological data and satellite remote sensing aerosol optical thickness data, and preprocessing to create model training and model prediction grid data; S2, fusing the multi-scale convolution algorithm in Inception and ConvLSTM to construct a multi-scale ConvLSTM network; S3, based on the conditional Wasserstein generative adversarial network framework with gradient penalty, the generator adopts a multi-scale ConvLSTM network, and the discriminator utilizes an improved U-Net network to construct an ICLU-CWGAN model; the U-Net network is improved as follows: (1) Apply spectral normalization to each layer's convolution kernel weights, remove all layer rectified linear units activation functions, ensuring that the discriminator satisfies Lipschitz continuity at each layer, the spectral normalized weights are computed by the following formula: In the formula, is the original convolution kernel weight, is the spectral norm of the weight matrix, which is approximated by the power iteration method; is the weight obtained after spectral normalization, is the spectral normalization;​ (2) A binary geographical perception mask is introduced to each layer feature map in the network structure, feature extraction guidance is added in the skip connection, and the fusion ratio of the encoder and decoder features is dynamically adjusted through the channel attention weight and the output feature map after adaptive weighted fusion The calculation formula is as follows: wherein, is the channel attention weight, ranging [0, 1]; denotes passing through the Sigmoid activation function , generating attention weights between 0-1; is the rectified linear unit; , is the 1x1 convolution weight matrix; is the binary geoinference mask matrix, the target area is 1, and the rest is 0, and is the height or row number, width or column number of the feature map; is the encoder feature, is the decoder feature, denotes the channel splicing of the encoder and decoder features; (3) adding instance normalization at the shallow layer of the encoder, and calculating the output feature value from the following formula; wherein is the normalized feature value of the i-th sample, the j-th channel, and the k-th position of the example; is the normalized feature value of the i-th sample, the j-th channel, and the k-th position of the example; is the normalized feature value of the i-th sample, the j-th channel, and the k-th position of the example; is the normalized feature value of the i-th sample, the j-th channel, and the k-th position of the example; is the i-th learnable channel scaling parameter; is the i-th learnable channel scaling parameter; is the feature value of the i-th sample, the j-th channel, and the k-th position of the example before normalization; is the feature value of the i-th sample, the j-th channel, and the k-th position of the example before normalization; is the feature value of the i-th sample, the j-th channel, and the k-th position of the example before normalization; , , is the channel mean and channel variance; is the i-th learnable channel shift parameter, is the i-th learnable channel shift parameter, is a numerical stability constant to prevent the denominator from being 0; and is the height or number of rows, and the width or number of columns of the feature map; (4) The feature map of the last layer is subjected to a convolution transformation, so that the U-Net network directly outputs the original score map, and the discriminator outputs the last linear layer output of the real sample may be calculated by the following formula:​​ In the formula, is a weight of a 1x1 convolution kernel, respectively represent a convolution kernel height, a convolution kernel width, an input feature map channel number, and an output feature map channel number; is a convolution operation; is a feature map output by the last layer of the U-Net; is a bias term; S4, training the ICLU-CWGAN model by using the training concentration grid data obtained by preprocessing in step S1; S5, applying the trained ICLU-CWGAN prediction model to the prediction of atmospheric pollutant concentration in the target area in the future period of time.

2. The method of claim 1, wherein the ICLU-CWGAN model is based on, The preprocessing method in step S1 is as follows: integrating satellite remote sensing aerosol optical thickness data and ground monitoring data by using a Bayesian model to generate an initial concentration field; constructing pollutant concentration grid data by using the inverse distance weighted method for spatial interpolation; outputting a spatiotemporally continuous concentration grid map by using Kalman filtering for collaborative updating of the grid data; implementing geographic constraints on the initial concentration grid by using geographic information system spatial analysis technology, and spatially cropping grid data in the interference area; and normalizing the spatially cropped grid data by using Box-Cox transformation, normalizing and reconstructing the concentration grid data, and repairing data missing and abnormal data detected by flexibility by using an anisotropic diffusion equation guided bilinear interpolation. 3.The method of claim 1, wherein the ICLU-CWGAN model is based on a convolutional neural network (CNN) and a generative adversarial network (GAN) model. The specific steps for constructing the multi-scale ConvLSTM network in step S2 are as follows: (1) designing a hierarchical processing scheme for the difference diffusion characteristics of pollutant concentration grid data at different time periods, using small-scale convolution, batch normalization and residual connection to capture local details for adjacent prediction periods, using medium-scale convolution, group normalization and residual connection to extract regional features for intermediate periods, and using large-scale convolution, instance normalization and residual connection to capture macroscopic distribution rules for remote periods; (2) using a gating mechanism to regulate information flow, using an input gate to determine the retention degree of new features through an activation coefficient, using a forgetting gate to filter historical information to be forgotten through a weight coefficient, and using an output gate to dynamically control the contribution of the current cell state to the final output; The multi-scale size convolution kernel is calculated by the following formula: wherein is a convolution operation that sizes the model input , is the size of the convolution kernel; are different convolution kernel sizes corresponding weights; is a convolution operation; is a bias term; is an input feature map; the output of the forget gate according to different convolution kernels the output of the input gate the output of the output gate the current time step the candidate cell state the current cell state and the hidden state of the current time step compute and update: where, is the model input at time step ; is the current time step; is the previous time step; , and represent small, medium and large size convolutional kernels, respectively; is the hidden state of the previous time step; is the Sigmoid activation function; is the cell state of the previous time step ; denotes the Hadamard product, i.e., element-wise multiplication; is the activation function, representing the rectified linear unit and the parametric ReLU, respectively; is the activation of ; and is the activation of 4.The method of predicting atmospheric pollutant concentration based on ICLU-CWGAN model according to claim 1, characterized in that, Step S4 is specifically implemented as follows: the historical PM 2.5 The concentration grid sequence is input into the multi-scale ConvLSTM network, the spatio-temporal features are extracted step by step, and the predicted concentration grid sequence is output; the concentration distribution map output by the multi-scale ConvLSTM network and the real observation data corresponding to the same time step are input into the U-Net network to evaluate the spatial authenticity of the generated data; the discriminator error and the generator error are back propagated and the parameters are updated, forming a kind of constantly iterative state of competitive game, so that the generator generates the pollutant concentration grid sequence which is difficult for the discriminator to judge.

Images