A highly robust comprehensive prediction method for atmospheric pollutant concentrations

Abstract

The present invention discloses a highly robust comprehensive prediction method for atmospheric pollutant concentrations, which belongs to the technical field of pollutant concentration prediction. The present invention uses a variety of environmental data obtained by meteorological environmental monitoring stations as original data, adopts the K-means concept and smoothing filling method to process missing values and outliers, so that the data still has high accuracy under the influence of filling; adopts feature engineering to perform feature screening of training data; trains neural network algorithms and random forest and iterative decision tree algorithms with randomness in parallel and generates initial prediction results; uses real data as a standard, dynamically searches for the optimal combination weights of random forest, iterative decision tree, and neural network prediction results; combines the prediction results with the optimal combination weights to obtain the pollutant prediction results of the final comprehensive prediction model. The present invention can ensure that when there is missing pollution in the quality of the original data, a high-precision prediction result can still be generated through data cleaning and combination methods.

Description

Technical Field

[0001] The present invention relates to the technical field of pollutant concentration prediction, and in particular to a highly robust comprehensive prediction method for atmospheric pollutant concentration. Background Art

[0002] Cities are complex entities, with industrial production being the primary source of urban air pollution. The recent increase in industrial production has made industrial pollutant emissions the primary culprit of air pollution, primarily sulfur dioxide and soot. Another source of urban air pollution is small coal-fired boilers, which primarily rely on coal for heating in winter. These boilers emit low-altitude emissions, directly impacting urban air quality. Furthermore, with the development of urban economies, the number of motor vehicles has increased dramatically, and vehicle exhaust emissions have shown a growing trend year by year. With the concentration of motor vehicles in cities, the main pollutants in vehicle exhaust include suspended solid particles, carbon monoxide, hydrocarbons, nitrogen oxides, lead, and sulfur oxides. Exhaust emissions have become a major source of urban air pollution.

[0003] At present, mainstream intelligent monitoring technology can rely on sensors and various computer technologies to complete air monitoring in relevant areas of industrial parks that produce a large amount of atmospheric pollutants, so as to detect problems in a timely manner and issue early warnings. However, this model of detecting and treating anomalies only after they are detected is often a last-ditch effort to remedy the situation after the atmospheric environment has deteriorated. The best opportunity for pollutant control has been missed, which is not conducive to the long-term healthy development of the environment. Therefore, it is necessary to be able to predict the future concentration of pollutants in the atmosphere so that corresponding measures can be taken in advance to control them. In the existing technology, there are two problems with the solutions for predicting the concentration of atmospheric pollutants: first, the model algorithm for predicting pollutant concentrations is not effective, resulting in insufficient prediction accuracy; second, it is highly dependent on the quality of the original data, and there is no suitable and efficient method to process the original incomplete and polluted data. These problems have led to the inability of atmospheric pollutant prediction models to make accurate predictions and provide inefficient support for atmospheric governance.

[0004] After searching, the Chinese patent application number is 2021110221848, the application date is September 1, 2021, and the name of the invention patent is: A method and system for predicting air pollutant concentrations. The application obtains the spatial relationship parameters of at least two air monitoring stations in the area to be measured, and obtains the monitoring index information of each of the air monitoring stations at the current moment; wherein, the monitoring index information includes air pollutant concentration monitoring values; based on the spatial relationship parameters and the monitoring index information at the current moment, a pre-trained feature extraction model is used to calculate the air quality characteristics at the current moment; the air quality characteristics at the current moment are input into a pre-trained pollutant concentration prediction model to obtain the predicted results of the air pollutant concentration in the area to be measured at the next moment output by the pollutant concentration prediction model. This application improves the prediction accuracy of air pollutant concentrations by extracting the spatial feature information of air monitoring stations and predicting air pollutants based on the time series and spatial feature information. However, this application is highly dependent on the quality of the original data and still needs further improvement. Summary of the Invention

[0005] 1. Technical problem to be solved by the invention

[0006] In view of the problems existing in the above-mentioned prior art, the present invention provides a highly robust comprehensive prediction method for atmospheric pollutant concentrations. The present invention effectively processes the original data and combines the trained random forest, iterative decision tree, and neural network after certain weight adjustments to form a final prediction model, which is applied to pollutant concentration prediction. Compared with traditional prediction methods, it can reduce dependence on data quality and significantly improve the prediction accuracy of atmospheric pollutant concentrations.

[0007] 2. Technical solution

[0008] In order to achieve the above object, the technical solution provided by the present invention is:

[0009] A highly robust comprehensive prediction method for atmospheric pollutant concentrations of the present invention comprises the following steps:

[0010] Step 1: Obtain the original data of environmental attributes within the prediction space. One time point corresponds to one piece of data. Each piece of data is represented by P = [p1, p2, ..., p n ] indicates that n is the number of attributes in the data, p1 to p n For specific attributes;

[0011] Step 2: Clean the original data and use different methods to deal with outliers and missing values in the original data;

[0012] Step 3: After data cleaning, calculate the variance of all attribute features in the data, remove attribute features with variance lower than the specified threshold, and obtain a data set P = [p1, p2, ..., p m ], where m <n;

[0013] Step 4: Use the training set obtained in step 3 to train the random forest, iterative decision tree, and neural network models in parallel to determine the model parameters and generate preliminary prediction results;

[0014] Step 5: Select real data as the data standard and absolute mean error as the evaluation criterion, and conduct a weight grid search from weight 0 to 1 with the goal of minimizing the absolute mean error to determine the combined weights of the random forest, iterative decision tree, and neural network models;

[0015] Step 6: Combine the random forest, iterative decision tree, and neural network according to the optimal weights obtained through grid search to generate prediction results under the comprehensive prediction model.

[0016] Furthermore, in step 2, the small-range outliers in the data attribute values are corrected using a linear smoothing method; for a large area of outlier distribution at the same monitoring site, they are corrected to the normal range through a linear mapping function; for a small range of missing values, the adjacent value filling and regression filling methods are used to complete the data supplement.

[0017] Furthermore, in step 2, the k-means clustering algorithm is used to deal with the situation of continuous large-scale missing values and abnormal values. The specific process is: for a missing attribute of any monitoring point, the attributes of other existing values of the monitoring point are taken as a high-dimensional vector, and the vector represents the monitoring point record at that time. The vector is recorded as Where t represents the recording time and j represents the monitoring point ID. Then a corresponding vector is generated for all other monitoring sites, recorded as Calculate the vector corresponding to the record to be filled Vectors recorded at the same time as other monitoring stations Euclidean distance The formula is as follows:

[0018]

[0019] Where, and is the value of the same attribute of two vectors;

[0020] After calculating the Euclidean distance between the record to be supplemented and all other monitoring points After that, sort them in ascending order, and select other monitoring points with the smallest Euclidean distance among 10% to 20% of the number of monitoring points. Calculate the sum of the values of the to-be-filled attributes recorded at the same time for these monitoring points and take the average. At this time, if there are individual monitoring points with vacancies in the to-be-filled attributes, discard them or replace them with other candidate monitoring points with smaller Euclidean distances. Finally, use this result as the filled value.

[0021] Further, in step 4, the process of training the random forest is as follows:

[0022] Randomly select N data samples with replacement from the training set, and use these N samples to train a decision tree as the sample at the root node of the decision tree; each sample has Q attributes. When each node of the decision tree needs to be split, randomly select q attributes from these Q attributes, satisfying the condition q << Q; then select 1 attribute with a large information gain from these q attributes as the splitting attribute of this node, and the degree of information gain is measured by the degree of change in information entropy; select the optimal feature at this node for splitting, so as to continuously split the training set until a complete decision tree is generated; repeat the above steps until a specified number of decision trees are generated, and combining these trees forms the final random forest model.

[0023] Further, in step 4, the process of training the iterative decision tree model is as follows:

[0024] Initialize the value of f₀(x) in the training set, and this value can make the loss function of the model in the 0th step take the minimum value; calculate the residual R during the process of the number of decision trees from 1 to T t After obtaining the residual R t Fit and learn a decision regression tree to obtain the parameters θ of the tth decision tree t to minimize the loss function; continuously update the model until T decision trees are generated; obtain the final iterative decision model.

[0025] Further, in step 4, the process of training the neural network model is as follows:

[0026] Pass the training data into the BN layer for processing, and then the data is passed from the output layer into the linear layer of the hidden layer, which contains 12 neurons, corresponding to the number of features retained after the data preprocessing process; use the ReLU activation function layer to process the data, and then connect a linear layer containing 60 neurons. The output layer obtains the three-dimensional data passed in by the hidden layer and uses the ReLU activation function for further processing to output the prediction result.

[0027] Further, in step 4, use the grid search method for the random forest and the iterative decision tree model to find the best settings for the number of trees and the depth parameters of a single tree.

[0028] Furthermore, in step 5, the weights of the random forest, decision iterative tree, and neural network models are determined, and the weighted weights of the three models are dynamically searched in the range of 0-1 using real data. After the weights of the three models are determined each time, the mean absolute error (MAE) under this weight distribution is calculated, and the weight with the lowest MAE score is determined as the weight composition of the comprehensive prediction model.

[0029] Furthermore, in step 6, the random forest, decision iteration tree, and neural network models are combined and weighted according to the optimal weight ratio found. The comprehensive prediction model formula is expressed as:

[0030] H=w1h1+w2h2+w3h3

[0031] Among them, h1 is the random forest model, h2 is the iterative decision tree model, h3 is the neural network model, and w1, w2, and w3 are the weights of each model.

[0032] 3. Beneficial effects

[0033] Compared with the existing known technologies, the technical solution provided by the present invention has the following significant effects:

[0034] (1) The present invention provides a highly robust comprehensive prediction method for atmospheric pollutant concentrations. The method uses a comprehensive prediction method on meteorological pollution problems, that is, after certain weight adjustments, the trained random forest, iterative decision tree, and neural network are combined to form a final prediction model, which is applied to pollutant concentration prediction. Compared with traditional prediction models, the method can improve the prediction accuracy of atmospheric pollutants to a certain extent. The method can predict the changes in atmospheric pollutants in a certain area and provide accurate and reliable data support for relevant departments to take relevant meteorological prevention and control measures.

[0035] (2) The highly robust comprehensive prediction method for atmospheric pollutant concentrations of the present invention ensures that the integrated model can still provide a relatively accurate result when each model does not produce a good prediction result or is not properly trained. In addition, the appropriate weight distribution method also reduces the error to a certain extent. The two together ensure that the comprehensive prediction method has higher prediction accuracy than the traditional prediction method.

[0036] (3) The present invention provides a highly robust comprehensive prediction method for atmospheric pollutant concentrations. In view of the fact that the original data may have problems such as data vacancies and anomalies, and data anomalies and data missing are important factors affecting data quality, if missing and abnormal data are not handled properly, it will directly affect the reliability of the results. To solve the above problems, the present invention adopts different methods to effectively process the original data, thereby obtaining more accurate prediction results. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 Schematic diagram of the process of the highly robust comprehensive prediction method for atmospheric pollutant concentrations of the present invention;

[0038] Figure 2 A neural network structure diagram of the present invention;

[0039] Figure 3 This is an implementation effect diagram of the present invention. DETAILED DESCRIPTION

[0040] In order to further understand the content of the present invention, the present invention is described in detail with reference to the accompanying drawings and embodiments.

[0041] Example 1

[0042] As environmental pollution worsens, air quality continues to deteriorate and the concentration of air pollutants continues to rise. Current mainstream air quality monitoring technologies can detect problems promptly and provide early warnings, but the prediction accuracy is insufficient. This embodiment can achieve environmental monitoring within a region, improving the prediction accuracy of air pollutant concentrations such as ozone, nitrogen dioxide, and PM10, providing effective support for subsequent air pollution control.

[0043] Specifically in this embodiment, we first obtain the original data of environmental attributes in the past within the prediction space. We take the meteorological data of 100 meteorological monitoring stations in a certain city from January 1, 2020 to October 6, 2020, totaling 670,000 records. The time interval between adjacent monitoring records at the same monitoring station is 1 hour. Among them, the data between January 1 and October 3 is used as the training set for each model, the data on October 4 is used to determine the weight composition of the comprehensive prediction model, and the data on October 5 and October 6 is used to verify the prediction effect of several prediction models. The pollutants to be predicted are the following three categories: ozone, nitrogen dioxide, and PM10.

[0044] The data attributes mainly include four categories: monitoring site information, pollutant concentration information, meteorological environment information, and indicator information. Monitoring site information includes monitoring time and monitoring station ID, which are used to express the attributes of record collection. Pollutant concentration information includes the concentration of common atmospheric pollutants, such as ozone, carbon monoxide, nitrogen dioxide, sulfur dioxide, PM10, PM2.5 and other pollutants. Meteorological environment information covers a variety of practical environmental information, including wind speed, wind direction, precipitation, dew point temperature, air temperature, air pressure, water vapor pressure, visibility, relative humidity and other attributes. Indicator information is several artificially defined index indicators used to describe the atmospheric pollution situation, such as air quality index, total volatile organic compounds and total suspended particulate matter.

[0045] Each piece of data in the original data consists of multiple attributes. One time point corresponds to one piece of data. Each piece of data is represented by P = [p1, p2, ..., p n ] indicates that n is the number of attributes in the data, p1 to p n The data is specific to the attribute; however, the quality of the original data is not ideal. The data presents the following issues: The range of meteorological environmental information values is abnormal. For example, in this example, the absolute humidity attribute data contains 44,838 negative values, while the humidity attribute has no negative values; small-scale data gaps. The attribute data of all monitoring stations has certain gaps, but these can be repaired using traditional data filling methods; large-scale data gaps. For example, in this example, some monitoring stations have large, continuous gaps in environmental data for certain days in September and October. Traditional data filling methods have difficulty finding suitable data support to fill these gaps.

[0046] Data anomalies and missing data are important factors that affect data quality. If missing and abnormal data are not handled properly, it will directly affect the reliability of the results. This embodiment handles the above problems differently. The abnormal values in a small range of data attribute values are corrected using a linear smoothing method; for the large distribution of abnormal values at the same monitoring site, they are corrected to the normal range through a linear mapping function. For missing values in a small range, two methods, adjacent value filling and regression filling, are used to complete the supplementary correction of the data. Adjacent value filling is to supplement and correct the missing value of each data with the average of the two values above and below the adjacent position. Regression filling is to substitute the known eigenvalues into the regression model to estimate the unknown eigenvalues, and use this estimated value to fill in. The regression model is the regression equation, which uses the least squares function to model the relationship between known eigenvalues.

[0047] The idea of K-means clustering algorithm is used to deal with the situation of continuous large-scale vacant values and outliers. The specific process is: for a vacant attribute of any monitoring point, take the attributes of other existing values of the monitoring point as a high-dimensional vector, and use this vector to represent the monitoring point record at that time. The vector is recorded as Where t represents the recording time of the record, and j represents the monitoring point ID. Then a corresponding vector is generated for all other monitoring sites, which is recorded as According to the idea of k-means clustering algorithm, samples with close Euclidean distance in high-dimensional space can be divided into one category, so the vector corresponding to the record to be supplemented can be calculated. Vectors recorded at the same time as other monitoring stations Euclidean distance The formula is as follows:

[0048]

[0049] Where, and They are the values of the same attribute for two vectors. After calculating the Euclidean distances between the record to be filled and all other monitoring points they are sorted in ascending order, and other monitoring points with the smallest Euclidean distances accounting for 10% to 20% of the number of monitoring points are selected. Calculate the sum of the values of the attribute to be filled recorded at the same time for these monitoring points and take the average result. At this time, if there are individual monitoring points with vacancies in the attribute to be filled, they are discarded or replaced with other candidate monitoring points with smaller Euclidean distances. Finally, this result is used as the filled value.

[0050] After the data cleaning process, calculate the variance values for all attributes in the data. The smaller the variance value of an attribute, the smaller the degree of dispersion of this feature, and the smaller its impact on the change of the pollutant to be predicted. Therefore, these attribute features with smaller variance values can be removed. Specifically in this embodiment, remove the attribute features with variances lower than the specified threshold. The threshold here needs to be set according to the specific scenario, and a data set P = [p1, p2, ……, p m with m attributes for each piece of data is obtained, where m < n. This training set is used as the training set for model training. Through this method, it is calculated that the variance of the air pressure attribute is small, and nearly 99% of the air pressure attribute values fluctuate around the standard atmospheric pressure (±0.5 hPa), so the air pressure attribute is removed.

[0051] Subsequently, for attributes with a large number of missing values and poor data filling effects, including the evaluation indicators for measuring the environment and several pollutant concentration attributes, they are all deleted. To ensure the effectiveness of the time attribute, the original time string is split and converted into three digital attributes of month, date, and hour and added to the original data. At the same time, to examine the impact of whether it is a working day on model training, compare the prediction accuracy of the model before and after adding this feature, and no better effect is obtained, so this newly added feature is abandoned. In addition, considering that there are differences in other conditions such as the geographical location of each monitoring point, a separate prediction model is established for each monitoring point to obtain more accurate prediction results.

[0052] Train the random forest. Randomly select N samples with replacement from the training set from January 1, 2020 to October 3, 2020 (select one sample randomly each time and then return to continue selecting). The N selected samples are used to train a decision tree as the sample at the root node of the decision tree. Each sample has Q attributes. When splitting is required at each node of the decision tree, randomly select q attributes from these Q features, satisfying the condition q << Q. Here, q is taken as the square root of Q, and then select 1 attribute with a large information gain from these q attributes as the splitting attribute of this node. The degree of information gain is measured by the degree of change of information entropy. The formula for information entropy is as follows:

[0053]

[0054] Among them, p(z i ) is the category z i The probability of occurrence, n is the number of categories. It can be seen that the size of entropy is only related to the probability distribution of the variable. The conditional entropy of Y under the condition of X refers to the amount of information (uncertainty) of the variable Y after the information of X. The calculation formula is as follows:

[0055]

[0056] When Entropy is 1, it is the worst state of classification, and when it is 0, it is the state of complete classification. Because entropy equals zero is an ideal state, in general, entropy is between 0 and 1. Select the optimal feature at this node to split, so as to i Continue splitting until a complete decision tree is generated. Repeat the above steps until a specified number of decision trees are generated, and combine these trees to form the final random forest model.

[0057] Train the iterative decision tree model. First, initialize the value of f0(x) in the training set from January 1, 2020 to October 3, 2020. This value can minimize the loss function of the model in step 0. Then, calculate the residual R as t gradually increases from 1 to T (maximum value). t , the residual is defined as:

[0058] R t =y i -f t-1 (x)

[0059] Among them, y i is the label value corresponding to the i-th sample in the training set, f t-1 (x) is the predicted value of the t-1th decision tree. t After that, a decision regression tree is learned by fitting and the parameter θ of the t-th decision tree is obtained. t , so that the loss function is minimized. Continuously update the model until T decision trees are generated. The final boosting tree model is obtained:

[0060]

[0061] Neural network training process: The neural network design used is as follows Figure 2As shown. The training data between October 1 and October 3, 2020, is passed to the BN (Batch Normalization) layer for processing, and then the data is passed from the output layer to the linear layer of the hidden layer, which contains 12 neurons, corresponding to the number of features retained by the data after the data preprocessing process. The data is then processed using the ReLU (Rectified Linear Unit) activation function layer. A linear layer containing 60 neurons is then connected. The output layer obtains the three-dimensional data passed in by the hidden layer and reprocesses it using the ReLU activation function, thereby completing the output of the predicted results for the concentrations of the three pollutants at one time. A neural network model is trained for one monitoring point, and a total of 100 models are trained.

[0062] Then use the grid search method to find the best parameter settings for the random forest and iterative decision tree models. For a single monitoring station, it takes an average of about 15 minutes for the above two models to generate and predict results. For short-term weather forecasts, the training set size of the neural network is mostly controlled at about one week. Therefore, for the training set size of the neural network, in order to prevent overfitting, we control the number of iterations, and finally the comprehensive prediction model selects the best neural network model with a size of nearly 3 days as the base model. After the parallel training of the above three models is completed, the weights of the random forest, iterative decision tree, and neural network models are determined, and the optimal weight combination of the three models is dynamically searched in the range of 0-1. After each weight is taken for the three models, the mean absolute error MAE (Mean Absolute Error) under this weight distribution is calculated using the actual data from October 4 to October 5, 2020. Finally, the weight with the lowest MAE score is selected as the weight composition of the comprehensive prediction model for a single pollutant attribute at a single monitoring station. The MAE formula is:

[0063]

[0064] Among them, u i is the model predicted value of the pollutant, v i is the true value of the pollutant.

[0065] Combine random forest, decision iterative tree, and neural network models and use weighted average method The comprehensive prediction model formula is expressed as:

[0066] H=w1h1+w2h2+w3h3

[0067] Among them, h1 is the random forest algorithm model, h2 is the iterative decision tree algorithm model, h3 is the neural network algorithm model, and w1, w2, and w3 are the weights of the algorithm models respectively.

[0068] In practice, the prediction performance of the integrated prediction model, along with the random forest model, iterative decision tree model, and neural network model for the three pollutants is shown in Tables 1 to 4. As can be seen from the tables, the integrated prediction model achieves improved accuracy compared to the three baseline models in most cases. For the prediction of the three pollutant concentrations, the integrated prediction model achieved a MAE score of 10.79, representing 20.1%, 23.0%, and 68.6% reductions compared to the random forest, iterative decision tree, and neural network models, respectively.

[0069] Table 1 MAE scores of prediction models (three pollutants)

[0070] Model Random Forest Iterative Decision Tree Neural Networks Comprehensive prediction model MAE score 13.51 14.02 34.38 10.79

[0071] Table 2 MAE scores of various prediction models under different training data sizes (ozone)

[0072] Model Random Forest Iterative Decision Tree Neural Networks Comprehensive prediction model MAE score 22.89 23.46 41.67 16.30

[0073] Table 3 MAE scores of prediction models under different training data sizes (nitrogen dioxide)

[0074] Model Random Forest Iterative Decision Tree Neural Networks Comprehensive prediction model MAE score 9.82 9.81 23.86 9.74

[0075] Table 4 MAE scores (PM10) of each prediction model under different training data sizes

[0076] Model Random Forest Iterative Decision Tree Neural Networks Comprehensive prediction model MAE score 10.18 10.70 37.61 7.39

[0077] Figure 3 To illustrate the effect of the invention, in general, the present invention combines the three basic models of random forest, iterative decision tree and neural network with optimal weights, effectively improving the prediction accuracy of atmospheric pollutant concentrations on the basis of data quality optimization, making up for the problem of low prediction accuracy of traditional models, and providing an effective method for predicting and preventing the concentration of urban air pollutants.

[0078] The above is a schematic description of the present invention and its embodiments, which is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. Therefore, if a person skilled in the art is inspired by this and, without departing from the purpose of the present invention, designs a structure and embodiment similar to this technical solution without inventiveness, they shall fall within the scope of protection of the present invention.

Claims

1. A highly robust comprehensive prediction method for atmospheric pollutant concentrations, characterized in that: The steps are: Step 1: Obtain the original data of environmental attributes within the prediction space. One time point corresponds to one piece of data. Each piece of data is represented by P = [p1, p2, ..., p n ] indicates that n is the number of attributes in the data, p1 to p n For specific attributes; Step 2: Clean the original data and use different methods to deal with outliers and missing values in the original data; Step 3: After data cleaning, calculate the variance of all attribute features in the data, remove attribute features with variance lower than the specified threshold, and obtain a data set P = [p1, p2, ..., p m ], where m <n; Step 4: Use the training set obtained in step 3 to train the random forest, iterative decision tree, and neural network models in parallel to determine the model parameters and generate preliminary prediction results; The process of training a random forest is: N data samples are randomly selected with replacement in the training set, and a decision tree is trained using these N data samples as samples at the root node of the decision tree; each sample has Q attributes, and when each node of the decision tree needs to be split, q attributes are randomly selected from these Q attributes, satisfying the condition q < <Q; Then, from these q attributes, an attribute with a large information gain is selected as the splitting attribute of the node. The degree of information gain is measured by the degree of change in information entropy. The optimal feature at this node is selected for splitting, thereby continuously splitting the training set until a complete decision tree is generated. The above steps are repeated until a specified number of decision trees are generated, and these trees are combined to form the final random forest model. The process of training an iterative decision tree model is: Initialize the value of f0(x) in the training set so that the loss function of the model in step 0 is minimized; Calculate the residual R in the process of the number of decision trees from 1 to T t , and get the residual R t After that, a decision regression tree is learned by fitting and the parameter θ of the t-th decision tree is obtained. t , so that the loss function is minimized; the model is continuously updated until T decision trees are generated; the final iterative decision model is obtained; The process of training a neural network model is: The training data is passed to the BN layer for processing. The data is then passed from the output layer to the linear layer of the hidden layer, which contains 12 neurons, corresponding to the number of features retained in the data after the data preprocessing process. The data is processed using the ReLU activation function layer, and then connected to a linear layer containing 60 neurons. The output layer obtains the three-dimensional data passed by the hidden layer and reprocesses it using the ReLU activation function to output the prediction result. Step 5: Select real data as the data standard and absolute mean error as the evaluation criterion, and conduct a weight grid search from weight 0 to 1 with the goal of minimizing the absolute mean error to determine the combined weights of the random forest, iterative decision tree, and neural network models; Step 6: Combine the random forest, iterative decision tree, and neural network according to the optimal weights obtained through grid search to generate prediction results under the comprehensive prediction model.

2. The highly robust comprehensive prediction method for atmospheric pollutant concentrations according to claim 1, characterized in that: In step 2, the outliers in a small range of data attribute values are corrected using a linear smoothing method; for a large area of outlier distribution at the same monitoring site, they are corrected to the normal range through a linear mapping function; for a small range of vacant values, the adjacent value filling and regression filling methods are used to complete the data supplement.

3. A highly robust comprehensive prediction method for atmospheric pollutant concentrations according to claim 1 or 2, characterized in that: In step 2, the k-means clustering algorithm is used to deal with the situation of continuous large-scale missing values and outliers. The specific process is: for a missing attribute of any monitoring point, the attributes of other existing values of the monitoring point are taken as a high-dimensional vector, and the vector represents the monitoring point record at that time. The vector is recorded as Where t represents the recording time and j represents the monitoring point ID. Then a corresponding vector is generated for all other monitoring sites, recorded as Calculate the vector corresponding to the record to be filled Vectors recorded at the same time as other monitoring stations Euclidean distance The formula is as follows: Where, and is the value of the same attribute of two vectors; After calculating the Euclidean distance between the record to be supplemented and all other monitoring points Then, sort them from small to large, and select 10% to 20% of the monitoring points with the smallest Euclidean distance. Calculate the sum of the attribute values to be filled recorded by these monitoring points at the same time and take the average result. At this time, if there are vacancies in the attributes to be filled for individual monitoring points, they will be discarded or replaced with other candidate monitoring points with a smaller Euclidean distance. Finally, the result is used as the filling value.

4. The highly robust comprehensive prediction method for atmospheric pollutant concentrations according to claim 3, characterized in that: In step 4, a grid search method is used to find the number of trees with the best effect and the depth parameter setting of a single tree for the random forest and iterative decision tree models.

5. The highly robust comprehensive prediction method for atmospheric pollutant concentrations according to claim 4, characterized in that: In step 5, the weights of the random forest, decision iterative tree, and neural network models are determined, and the weighted weights of the three models are dynamically searched in the range of 0-1 using real data. After the weights of the three models are determined each time, the mean absolute error (MAE) under the weight distribution is calculated, and the weight with the lowest MAE score is determined as the weight composition of the comprehensive prediction model.

6. The highly robust comprehensive prediction method for atmospheric pollutant concentrations according to claim 5, characterized in that: In step 6, the random forest, decision iterative tree, and neural network models are combined and weighted according to the optimal weight ratio found. The comprehensive prediction model formula is expressed as: H=w1h1+w2h2+w3h3 Among them, h1 is the random forest model, h2 is the iterative decision tree model, h3 is the neural network model, and w1, w2, and w3 are the weights of each model.

Images