A high-precision atmospheric pollutant concentration prediction method based on machine learning

Abstract

This invention relates to the field of atmospheric environment monitoring technology and discloses a high-precision method for predicting atmospheric pollutant concentrations based on machine learning. The key technical points are: multi-source data acquisition and correlation preprocessing, correlation feature engineering, adaptive data transformation, hybrid machine learning model construction, model optimization, and concentration prediction. A closed-loop technical solution is adopted, employing multi-source data correlation preprocessing, dynamic feature engineering, improved RBF kernel SVR-ELM hybrid model construction, and adaptive PSO optimization, to achieve high-precision prediction of hourly concentrations of different types of atmospheric pollutants, providing reliable data support for atmospheric pollution prevention and control and public health protection.

Description

Technical Field

[0001] This invention relates to the field of atmospheric environment monitoring technology, and more specifically, to a high-precision method for predicting atmospheric pollutant concentrations based on machine learning. Background Technology

[0002] Existing methods for predicting atmospheric pollutant concentrations often lack correlation analysis of multi-source data in the data preprocessing stage, resulting in insufficient utilization of key information; feature engineering often ignores spatiotemporal dynamic correlations, affecting the model's ability to characterize complex pollution processes; and most models use a single machine learning algorithm, which has limited generalization performance, especially with large errors in predicting pollutant concentration peaks.

[0003] Therefore, the present invention provides a high-precision method for predicting atmospheric pollutant concentrations based on machine learning, thereby improving the above-mentioned technical problems. Summary of the Invention

[0004] This invention aims to address the shortcomings of existing technologies by providing a high-precision atmospheric pollutant concentration prediction method based on machine learning. Through multi-source data correlation preprocessing, dynamic feature extraction, improved hybrid model construction, and adaptive optimization, this invention significantly improves prediction accuracy and peak capture capability, providing reliable technical support for precise prevention and control of air pollution.

[0005] The above-mentioned technical objective of this invention is achieved through the following technical solution: a high-precision method for predicting atmospheric pollutant concentrations based on machine learning, comprising the following steps: S1. Multi-source data acquisition and correlation preprocessing: Collect multi-source data containing pollutant concentration, meteorological factors and time information in the target area. Use an improved polynomial-extreme learning machine hybrid imputation method to fill missing values. Combine the improved IQR method and hysteresis feature verification to perform adaptive outlier detection and replacement. S2. Relationship feature engineering: Perform dynamic feature extraction, including converting time information into trigonometric function features, constructing lag features and rolling features, and using an improved correlation coefficient method that introduces time weights and outlier correction for feature selection, and constructing a feature matrix; S3. Adaptive Data Transformation: Adaptive data transformation is performed on the filtered feature data by combining Yeo-Johnson transformation and min-max normalization. The transformation parameters are dynamically adjusted by the improved correlation coefficient between the feature and the target pollutant. S4. Hybrid Machine Learning Model Construction: Construct a hybrid prediction model based on support vector regression and extreme learning machine with improved radial basis function kernel. The radial basis function kernel incorporates the ensemble weights of random forest. The final output of the hybrid model is the weighted sum of the predictions of the two models. The weights are dynamically determined by their respective errors on the training set. S5. Model Optimization and Concentration Prediction: An improved particle swarm optimization algorithm is adopted, with the weighted average absolute error and root mean square error as objective functions, to optimize the key parameters of the hybrid model. The optimized model is used for prediction, and the peak deviation of the output hourly concentration prediction value is corrected to obtain the final atmospheric pollutant concentration prediction result.

[0006] As a preferred embodiment of the present invention, in step S1, the formula for the improved polynomial-limit learning machine hybrid filling method is: in, Let be the fill value for the j-th pollutant at time i; For time series indexing, The coefficients are those of a second-order polynomial. Output matrix of ELM hidden layer; Output weights for ELM; This is the correlation coefficient.

[0007] As a preferred embodiment of the present invention, in step S2, the formula for the improved correlation coefficient method is: in, Let be the improved correlation coefficient between the j-th pollutant and the k-th feature; As weight; The correlation between the time feature and the association pair.

[0008] As a preferred embodiment of the present invention, in step S3, the normalization formula for the adaptive data transformation is: in, These are the eigenvalues ​​after the Yeo-Johnson transform; These are the Yeo-Johnson transform parameters; For adjustment factors; Let be the improved correlation coefficient between the j-th feature and the target pollutant.

[0009] As a preferred embodiment of the present invention, in step S4, the formula for the improved radial basis function kernel is: in, For kernel parameters; Let be the prediction weight of the k-th tree in the random forest; This is the fusion coefficient.

[0010] As a preferred embodiment of the present invention, in step S4, the output formula of the hybrid model is: in, To improve the predicted values ​​of RBF kernel SVR, These are the predicted values ​​for ELM; The weighting coefficient is calculated using the following formula: .

[0011] As a preferred embodiment of the present invention, in step S5, the objective function of the improved particle swarm optimization algorithm is: in, , The weight is determined by the maximum value of the correlation coefficient between the feature and the target pollutant.

[0012] As a preferred embodiment of the present invention, in step S5, the formula for peak deviation correction is: in, The predicted values ​​of the optimized mixture model; , The mean and standard deviation of the forecast values ​​for the previous two hours; Peak correction factor .

[0013] In summary, the present invention has the following beneficial effects: Firstly, to address the issue of insufficient handling of missing and outlier values ​​during data preprocessing, the improved polynomial-ELM hybrid imputation method and adaptive outlier detection mechanism significantly enhance the integrity and consistency of multi-source data, providing a high-quality and highly reliable input data foundation for subsequent modeling.

[0014] Secondly, by introducing a time-weighted improved correlation feature screening method and combining dynamic lag and rolling feature extraction, the spatiotemporal correlation expression ability between features and pollutant concentration is effectively enhanced, and the model’s feature sensitivity and interpretability to complex pollution processes are improved.

[0015] Third, a hybrid prediction model integrating improved RBF kernel SVR and ELM was constructed. By combining random forest weights and dynamic output weighting mechanism, the advantages of different machine learning models were fully utilized, significantly enhancing the model's generalization ability and overall prediction stability.

[0016] Fourth, the adaptive PSO optimization algorithm is used to dynamically optimize the key parameters of the hybrid model, and the model is trained with weighted MAE-RMSE as the target, which effectively improves the prediction accuracy of the model in the fluctuation of pollutant concentration, especially during peak periods, and the peak capture capability is significantly improved. Attached Figure Description

[0017] Figure 1 A flowchart illustrating a high-precision atmospheric pollutant concentration prediction method based on machine learning, provided as an embodiment of the present invention. Detailed Implementation

[0018] The present application will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present application, but do not limit the present application in any way. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of the present application. These all fall within the protection scope of the present application.

[0019] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0020] Unless otherwise defined, all technical and scientific terms used in this specification have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the scope of this application. The term "and / or" as used in this specification includes any and all combinations of one or more of the associated listed items.

[0021] Furthermore, the technical features involved in the various embodiments of this application described below can be combined with each other as long as they do not conflict with each other.

[0022] This disclosure aims to address the shortcomings of existing air pollutant concentration prediction methods, such as insufficient data preprocessing targeting, inadequate utilization of feature correlations, weak generalization ability of single models, and limited peak prediction accuracy. Therefore, this disclosure proposes a high-precision air pollutant concentration prediction method based on machine learning. It employs a closed-loop technical scheme involving multi-source data correlation preprocessing, dynamic feature engineering, construction of an improved RBF kernel SVR-ELM hybrid model, and adaptive PSO optimization. This achieves high-precision hourly concentration prediction for different types of air pollutants, providing reliable data support for air pollution control and public health protection.

[0023] Please refer to Figure 1 , Figure 1A flowchart of the high-precision atmospheric pollutant concentration prediction method based on machine learning according to an embodiment of this disclosure is shown. The overall process mainly includes the following five steps: S1. Multi-source data acquisition and correlation preprocessing.

[0024] S1.1 Data Acquisition: Acquire multi-source data for the target area, including: Pollutant data: Hourly concentration data; Meteorological data: temperature (T), humidity (RH), wind speed (WS), wind direction (WD), air pressure (P); Time data: year, month, day, hour, season, weekend identifier.

[0025] S1.2, Missing Value Filling: A modified polynomial-ELM hybrid filling method is adopted, combining pollutant hysteresis characteristics with meteorological correlation. The formula is as follows: (1) in, Let be the fill value for the j-th pollutant at time i; For time series indexing, The coefficients are second-order polynomials (fitted by least squares method). The output matrix of the ELM hidden layer (the input consists of the first two hysteresis characteristics of the pollutant and the meteorological data of the same period). Output weights for ELM, and solve using the Moore-Penrose generalized inverse: ; The correlation coefficient is determined by the correlation between the pollutant and meteorological data. .

[0026] S1.3 Adaptive Outlier Detection: Combining the improved IQR method with hysteresis feature verification, the formula is as follows: (2) in, Let j be the rolling average of the j-th pollutant over the previous hour. It is a lag-1 characteristic; Let be the standard deviation of the concentration of the j-th pollutant; mark it as an outlier ( The data is replaced by the fill formula in S1.2.

[0027] S2, Relational Feature Engineering.

[0028] S2.1 Dynamic Feature Extraction.

[0029] Time features: Converting hours and months into trigonometric function features: Construct seasonal features S∈{1,2,3,4} and weekend identifiers W∈{0,1}; Lag characteristics: Based on ACF / PACF analysis, the optimal lag order k (k=1,2) for each pollutant was determined, and lag characteristics were extracted. ; Scrolling features: Constructing dynamic window scrolling mean The window w is determined by time characteristics: w=24 on weekdays and w=12 on weekends, with a seasonal adjustment factor. .

[0030] S2.2 Improved Relevance Feature Screening: Introduce time-weighted and outlier-corrected Pearson correlation coefficients, as shown in the following formula: (3) in, Let be the improved correlation coefficient between the j-th pollutant and the k-th feature; The weight is determined by the proportion of outliers in S1.3. ; The correlation between the time feature and the association pair.

[0031] reserve Based on the characteristics, construct the feature matrix. Where n is the number of samples and m is the number of features after screening.

[0032] S3, Adaptive Data Transformation.

[0033] Combining the Yeo-Johnson transform with min-max normalization, the transform parameters are dynamically adjusted by the improved correlation coefficient of S2, as shown in the following formula: (4) in, The parameters for the Yeo-Johnson transform are obtained from the maximum likelihood estimation. For adjustment factor ( ); Let be the improved correlation coefficient between the j-th feature and the target pollutant.

[0034] After normalization, the data range is [0, 1.2], which enhances the weight of highly correlated features.

[0035] S4, Hybrid Machine Learning Model Construction.

[0036] A hybrid model combining the improved RBF kernel SVR and ELM is constructed, with the model parameters correlated with the normalized results of S3, as detailed below: S4.1 Improved RBF kernel function: Incorporating the ensemble weights of random forest, the formula is as follows: (5) in, The kernel parameters (initial values ​​are determined by the variance of the normalized features): ); Let be the prediction weight of the k-th tree in the random forest. , Let be the mean square error of the k-th tree; The fusion coefficient is... .

[0037] S4.2 Hybrid Model Output: Combining the prediction results of SVR and ELM, the weights are dynamically adjusted by the training errors of both, as shown in the following formula: (6) in, To improve the predicted values ​​of RBF kernel SVR, These are the predicted values ​​for ELM; These are the weighting coefficients. RMSE is the root mean square error of the training set.

[0038] S5. Model optimization and concentration prediction.

[0039] S5.1 Improved PSO Optimization Algorithm: Using weighted MAE-RMSE as the objective function, optimize the key parameters of the mixture model. The objective function is as follows: (7) in, , The weights are determined by the importance of the features in S2. ).

[0040] S5.2 Final Concentration Prediction: The optimized model is applied to the test set to output hourly predicted atmospheric pollutant concentrations, and then... Correction for peak deviation: (8) in, The predicted values ​​of the optimized mixture model; , The mean and standard deviation of the forecast values ​​for the previous two hours; Peak correction factor .

[0041] Example: Step 1: Collect multi-source data for a city, including data from 12 monitoring stations. Hourly concentration data, as well as concurrent meteorological data (temperature, humidity, wind speed, wind direction, air pressure) and time data.

[0042] Step 2: Fill in the missing data for each pollutant using Formula 1. Set the number of hidden layer nodes in the ELM to 20, and the correlation coefficient... Based on the correlation between pollutants and humidity Outlier detection: Outliers are detected using Formula 2. Outliers accounted for approximately 3%, and the data integrity reached 99.8% after replacement.

[0043] Step 3: Extract time features ( Seasonal (S), weekend (W), lag characteristics (k=1, 2), and rolling characteristics (weekday w=24, weekend w=12). The improved correlation coefficient is calculated using Formula 3, retaining... The features are such that the final feature matrix has a dimension of n×28 (m=28).

[0044] Step 4: Perform adaptive transformation using Equation 4, Yeo-Johnson transform parameters Obtained through maximum likelihood estimation After normalization, the data range is [0, 1.2].

[0045] Step 5: Set up a random forest with 500 trees and calculate the weight of each tree. The initial parameters for the improved RBF kernel SVR are: C∈[1, 100], ε∈[0.001, 0.1], γ=1 / 28; the number of ELM hidden layer nodes L∈[10, 50]. The parameters are optimized using the improved PSO algorithm, and the objective function weights are... The number of iterations was 50, and the optimized parameters were: C=3.2, ε=0.035, γ=0.042, L=30. Hybrid model weights The mixed prediction value is obtained through Formula 6, and the peak value is then corrected through Formula 8.

[0046] In summary, the test set prediction results show that the predicted concentrations of each pollutant are... All are greater than 0.85, among which By using a peak correction mechanism to calibrate the prediction results, the accuracy of hourly concentration predictions is further improved, and the R-value of the prediction results for each pollutant is increased. 2 All values ​​are better than 0.85, meeting the operational forecasting requirements of high accuracy and high reliability, and providing effective technical support for air pollution prevention and control and public health protection.

[0047] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A high-precision method for predicting atmospheric pollutant concentrations based on machine learning, characterized in that, The method includes the following steps: S1. Multi-source data acquisition and correlation preprocessing: Collect multi-source data containing pollutant concentration, meteorological factors and time information in the target area. Use an improved polynomial-extreme learning machine hybrid imputation method to fill missing values. Combine the improved IQR method and hysteresis feature verification to perform adaptive outlier detection and replacement. S2. Relationship feature engineering: Perform dynamic feature extraction, including converting time information into trigonometric function features, constructing lag features and rolling features, and using an improved correlation coefficient method that introduces time weights and outlier correction for feature selection, and constructing a feature matrix; S3. Adaptive Data Transformation: Adaptive data transformation is performed on the filtered feature data by combining Yeo-Johnson transformation and min-max normalization. The transformation parameters are dynamically adjusted by the improved correlation coefficient between the feature and the target pollutant. S4. Hybrid Machine Learning Model Construction: Construct a hybrid prediction model based on support vector regression and extreme learning machine with improved radial basis function kernel. The radial basis function kernel incorporates the ensemble weights of random forest. The final output of the hybrid model is the weighted sum of the predictions of the two models. The weights are dynamically determined by their respective errors on the training set. S5. Model Optimization and Concentration Prediction: An improved particle swarm optimization algorithm is adopted, with the weighted average absolute error and root mean square error as objective functions, to optimize the key parameters of the hybrid model. The optimized model is used for prediction, and the peak deviation of the output hourly concentration prediction value is corrected to obtain the final atmospheric pollutant concentration prediction result.

2. The high-precision atmospheric pollutant concentration prediction method based on machine learning according to claim 1, characterized in that, In S1, the formula for the improved polynomial-limit learning machine hybrid filling method is: in, Let be the fill value for the j-th pollutant at time i; For time series indexing, The coefficients are those of a second-order polynomial. Output matrix of ELM hidden layer; Output weights for ELM; This is the correlation coefficient.

3. The high-precision atmospheric pollutant concentration prediction method based on machine learning according to claim 1, characterized in that, In S2, the formula for the improved correlation coefficient method is: in, Let be the improved correlation coefficient between the j-th pollutant and the k-th feature; As weight; The correlation between the time feature and the association pair.

4. The high-precision atmospheric pollutant concentration prediction method based on machine learning according to claim 1, characterized in that, In step S3, the normalization formula for the adaptive data transformation is: in, These are the eigenvalues ​​after the Yeo-Johnson transform; These are the Yeo-Johnson transform parameters; For adjustment factors; Let be the improved correlation coefficient between the j-th feature and the target pollutant.

5. The high-precision atmospheric pollutant concentration prediction method based on machine learning according to claim 1, characterized in that, In S4, the formula for the improved radial basis function kernel is: in, For kernel parameters; Let be the prediction weight of the k-th tree in the random forest; This is the fusion coefficient.

6. The high-precision atmospheric pollutant concentration prediction method based on machine learning according to claim 1, characterized in that, In S4, the output formula of the hybrid model is: in, To improve the predicted values ​​of RBF kernel SVR, These are the predicted values ​​for ELM; The weighting coefficient is calculated using the following formula: .

7. The high-precision atmospheric pollutant concentration prediction method based on machine learning according to claim 1, characterized in that, In step S5, the objective function of the improved particle swarm optimization algorithm is: in, , The weight is determined by the maximum value of the correlation coefficient between the feature and the target pollutant.

8. The high-precision atmospheric pollutant concentration prediction method based on machine learning according to claim 1, characterized in that, In step S5, the formula for peak deviation correction is: in, The predicted values ​​of the optimized mixture model; , The mean and standard deviation of the forecast values ​​for the previous two hours; Peak correction factor .

Images