A PM1 concentration estimation method and system
The PM1 concentration prediction model was constructed through the LGBM algorithm, and combined with data such as PM2.5 concentration, geographical location, meteorological conditions and aerosol optical thickness, the problem of low PM1 concentration estimation accuracy in the existing technology was solved, and high-precision PM1 concentration distribution estimation was achieved. , improves prediction accuracy and resolution.
Patent Information
- Application Number
- CN202210396971.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-04-15
- Publication Date
- 2024-12-17
- Estimated Expiration
- 2042-04-15
AI Technical Summary
In the prior art, the accuracy of PM1 concentration estimation is low, and the uneven distribution of the monitoring stations makes it impossible to perform spatiotemporal monitoring on a detailed scale.
Using machine learning method based on LGBM algorithm, the PM1 concentration prediction model is constructed, and the correlation between PM1 concentration and PM1 concentration influence indicators is used, and data such as PM2.5 concentration, geographical location, meteorological conditions and aerosol optical thickness are used to achieve high quality. Accurate PM1 concentration estimation.
The accuracy and resolution of PM1 concentration prediction are improved, so that the prediction results are consistent with the real value, with small errors and high accuracy, and can effectively realize high-precision estimation of the PM1 concentration distribution.
Smart Images
Figure CN114819312B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of environmental protection, and more specifically, relates to a method and system for estimating PM1 concentration. Background Art
[0002] Nowadays, with the rapid economic development, air pollution has also occurred in my country. Among all the pollutants emitted, aerosol particles have a particularly significant impact on air quality. Humans are paying more and more attention to the measurement of PM1. Studies have shown that exposure to particulate matter is more dangerous, and it is more harmful to human health than other air pollutants such as ozone and nitrogen oxides. PM1 stays longer than other particulate matter and is easier to enter the alveoli. Therefore, it is very important to estimate the PM1 concentration distribution with high precision for further prevention and control of air pollution.
[0003] However, there are few studies on PM1 concentration estimation, and many studies are based on the ground observation station network of PM1 to invert its spatiotemporal distribution. However, the uneven distribution of monitoring stations makes it impossible to monitor PM1 in a more detailed scale. How to achieve high-precision estimation of PM1 concentration is a problem that needs to be solved in this field. Summary of the invention
[0004] The present invention solves the problem of low accuracy of PM1 concentration estimation in the prior art by providing a PM1 concentration estimation method and system.
[0005] The present invention provides a method for estimating PM1 concentration, comprising the following steps:
[0006] Step 1: Collect and obtain data on PM1 concentration information and PM1 concentration impact indicators;
[0007] Step 2: Analyze the correlation between PM1 concentration and PM1 concentration influencing indicators to obtain PM1 concentration related indicators; construct a training set and a test set based on the PM1 concentration information and the data of the PM1 concentration related indicators;
[0008] Step 3: Construct a PM1 concentration prediction model based on the LGBM algorithm, taking PM1 concentration related indicators as input parameters of the model and PM1 concentration as output parameters of the model;
[0009] Step 4: train the PM1 concentration prediction model using the training set, test the trained PM1 concentration prediction model using the test set, and estimate PM1 concentration based on the trained PM1 concentration prediction model.
[0010] Preferably, in step 1, the PM1 concentration impact index includes PM 2.5concentration, geographical location, meteorological conditions and aerosol optical depth.
[0011] Preferably, the China multi-resolution emission inventory model obtained by WRF-CHEM is used to perform PM 2.5 The nationwide simulation results show that PM 2.5 Concentration data.
[0012] Preferably, the geographical location includes one or more indicators of the normalized difference vegetation index and the digital elevation model; the meteorological conditions include one or more indicators of surface relative humidity, air temperature, wind speed, surface pressure, and boundary layer height; the data of aerosol optical thickness comes from the Himawari-8 level 3 hourly data set, and uses data samples with a confidence level marked as "very good".
[0013] Preferably, in step 2, the correlation coefficient between each PM1 concentration influencing indicator and the PM1 concentration is calculated, and a judgment is made as to whether the correlation coefficient corresponding to each PM1 concentration influencing indicator meets a preset condition, and the PM1 concentration influencing indicator that meets the preset condition is used as a PM1 concentration-related indicator.
[0014] Preferably, step 3 also includes: optimizing basic parameters of the LGBM algorithm by a random parameter search method.
[0015] Preferably, the LGBM algorithm uses a gradient boosting decision tree, and the basic parameters of the LGBM algorithm include: optimal leaf weight, number of leaves, maximum depth of the tree, and feature sampling value.
[0016] In another aspect, the present invention provides a system for estimating PM1 concentration, comprising:
[0017] Data acquisition module, used to collect PM1 concentration information and PM1 concentration impact index data;
[0018] A correlation analysis module is used to analyze the correlation between PM1 concentration and PM1 concentration influencing indicators to obtain PM1 concentration related indicators, and to construct a training set and a test set based on the PM1 concentration information and the data of the PM1 concentration related indicators;
[0019] The model building module is used to build a PM1 concentration prediction model based on the LGBM algorithm, with PM1 concentration-related indicators as input parameters of the model and PM1 concentration as output parameters of the model;
[0020] A fitting prediction module, used to train the PM1 concentration prediction model through the training set, test the trained PM1 concentration prediction model through the test set, and realize PM1 concentration estimation based on the trained PM1 concentration prediction model;
[0021] The PM1 concentration estimation system is used to implement the steps in the above-mentioned PM1 concentration estimation method.
[0022] One or more technical solutions provided in the present invention have at least the following technical effects or advantages:
[0023] (1) The present invention designs a high-precision concentration prediction model based on the LGBM algorithm of machine learning to estimate PM1 concentration. The present invention determines PM1 concentration-related indicators based on the correlation between PM1 concentration and PM1 concentration influencing indicators, uses PM1 concentration-related indicators as input parameters of the PM1 concentration prediction model, and uses PM1 concentration as the output parameter of the model. After training and testing the model, PM1 concentration estimation is achieved based on the trained PM1 concentration prediction model. The present invention combines multiple PM1 concentration-related indicators to obtain a PM1 concentration prediction model, which can effectively improve the accuracy of PM1 concentration prediction.
[0024] (2) The present invention innovatively proposes to use PM 2.5 The concentration is used as an influencing indicator of PM1 concentration in the prediction of PM1 concentration. That is, the present invention innovatively uses the simulation information of WRF-CHEM of China's multi-resolution emission inventory as an important indicator affecting PM1 concentration, so that the prediction accuracy and resolution of PM1 concentration of the present invention are higher than other current prediction methods. The prediction results obtained by the present invention are consistent with the true value, with small errors and high accuracy.
[0025] (3) The present invention uses random search parameter optimization to construct a high-precision prediction model, which can further improve the estimation accuracy and efficiency of PM1 based on parameter optimization. BRIEF DESCRIPTION OF THE DRAWINGS
[0026] Figure 1 A schematic diagram of a flow chart of a method for estimating PM1 concentration provided in Example 1 of the present invention;
[0027] Figure 2 A schematic diagram of a process for training and testing a PM1 concentration prediction model in a PM1 concentration estimation method provided in Example 1 of the present invention;
[0028] Figure 3 This is a flow chart of the LGBM algorithm in a PM1 concentration estimation method provided in Example 1 of the present invention. DETAILED DESCRIPTION
[0029] In order to better understand the above technical solution, the above technical solution will be described in detail below in conjunction with the accompanying drawings and specific implementation methods.
[0030] Embodiment 1:
[0031] Example 1 provides a method for estimating PM1 concentration, see Figures 1 to 3 , including the following steps:
[0032] Step 1: Collect and obtain data on PM1 concentration information and PM1 concentration impact indicators.
[0033] Among them, the PM1 concentration impact indicators include PM 2.5 concentration, geographical location, meteorological conditions and aerosol optical depth.
[0034] Specifically, the China multi-resolution emission inventory model obtained through WRF-CHEM can be used to perform PM 2.5 The nationwide simulation results show that PM 2.5 Concentration data.
[0035] PM obtained 2.5 The data is simulated using the WRF-CHEM framework based on the China High Resolution Emission Inventory (MEIC), which provides emission data of anthropogenic pollutants and greenhouse gases, and provides annual emission data with a spatial resolution of 0.1 degrees. The data is gridded to meet the needs of atmospheric chemical transport models at multiple scales. By simulating this emission inventory using WRF-CHEM, a high-resolution national PM 2.5 The concentration distribution data.
[0036] PM1 refers to particles in the atmosphere with a dynamic diameter less than or equal to 1 micron. 2.5 Particles with a diameter of less than or equal to 2.5 microns, so the two are included in the relationship. In addition, because the two have homology, PM 2.5 The prior can greatly improve the accuracy of PM1 concentration prediction. 2.5 The concentration of PM1 is an important input parameter of the PM1 concentration prediction model, which can effectively improve the prediction accuracy.
[0037] The geographical location includes one or more indicators of the normalized difference vegetation index and the digital elevation model; the meteorological conditions include one or more indicators of surface relative humidity, air temperature, wind speed, surface pressure, and boundary layer height; the data of aerosol optical thickness comes from the Himawari-8 level 3 hourly data set, and uses data samples with a confidence level marked as "very good".
[0038] The geographical location data and meteorological condition data across the country can be obtained through satellite observations, the aerosol optical thickness data can be obtained through satellite remote sensing methods, and the PM1 concentration information can be obtained through ground observation stations.
[0039] For example, to obtain the PM1 concentration value at the actual monitoring site, it can be obtained from the aerosol observation network operated by the China Meteorological Administration, which consists of 73 available monitoring sites that simultaneously measure PM1, PM2. 2.5 and PM 10 The mass concentration is measured and recorded every 5 minutes. From July 2015 to December 2016, a valid sample of 32,123 hours of average PM1 concentration data was obtained, excluding calibration instrument failures, missing data, and abnormal data. In order to monitor other data around the actual monitoring site, historical data of aerosols, normalized difference vegetation index, and meteorological data (humidity, temperature, wind speed, boundary layer height, etc.) across the country can be obtained from the Himawari-8 satellite, MODIS (Moderate-resolution Imaging Spectroradiometer) satellite, and ERA-interior. These data are also saved in hours corresponding to the detection date.
[0040] Step 2: Analyze the correlation between PM1 concentration and PM1 concentration influencing indicators to obtain PM1 concentration related indicators; construct a training set and a test set based on the PM1 concentration information and the data of the PM1 concentration related indicators.
[0041] Specifically, the correlation coefficient between each PM1 concentration influencing indicator and the PM1 concentration is calculated, and it is judged whether the correlation coefficient corresponding to each PM1 concentration influencing indicator meets the preset conditions, and the PM1 concentration influencing indicator that meets the preset conditions is used as the PM1 concentration related indicator.
[0042] In specific implementation, the PM1 concentration is analyzed for correlation with the acquired meteorological conditions (e.g., surface relative humidity, air temperature), aerosol optical depth, geographic location (e.g., normalized difference vegetation index) and other indicators according to formula (1) to determine the indicators closely related to PM1. The calculation formula of the correlation coefficient is as follows:
[0043]
[0044] Among them, X i is the change of the ith indicator (i.e. parameter), Y is the change of PM1 concentration, var[X i ] indicates X i The variance of Y, var[Y] represents the variance of Y, Cov(X i ,Y) is X i Covariance with Y, r i (X i ,Y) is the i-th parameter (X i ) and PM1 concentration.
[0045] This embodiment determines the input parameters of machine learning Boosting by judging whether the correlation coefficient reaches the parameter within the required range.
[0046] Step 3: Construct a PM1 concentration prediction model based on the LGBM algorithm, take PM1 concentration related indicators as the input parameters of the model, and take PM1 concentration as the output parameter of the model.
[0047] The preferred solution also includes: optimizing the basic parameters of the LGBM algorithm by a random parameter search method. The LGBM algorithm uses a gradient boosting decision tree, and the basic parameters of the LGBM algorithm include: optimal leaf weight, number of leaves, maximum depth of the tree, feature sampling value, etc.
[0048] See also Figure 3 In this step, the input and output are set for the LGBM machine learning method, and the specific implementation method of using the random parameter search method to optimize the basic parameters of the LGBM algorithm is as follows:
[0049] The initialization phase of Boosting uses no-return sampling to randomly select a subset from the training sample, which is a dependent serial machine learning. 2.5 The output is a training model fitted by machine learning.
[0050] LGBM uses Histogram for differential acceleration and adopts Leaf-wise growth strategy. The LGBM algorithm needs to set the selection of tree structure, adjust the compensation of gradient descent, and set the appropriate maximum number of iterations. The algorithm optimizes a maximum depth limit parameter to prevent overfitting while ensuring high efficiency.
[0051] Use random search parameter adjustment to adjust the parameters of LGBM so that the algorithm parameters of machine learning are optimized. This parameter adjustment method randomly generates a series of random points within a certain interval of machine learning, and then calculates the values of the objective function and constraint function of these points. Then further compare these values, extract the random points that meet the constraints, and then compare them with the values of the objective function, delete the points with large differences, and retain the points with good comparison results, so as to obtain an approximate solution to the optimal solution of the parameters through continuous comparison. This method is based on the theory of probability theory. The more random points are generated, the closer the solution is to the optimal solution. This method optimizes the parameters of LGBM and improves its parameter performance.
[0052] In order to better understand the present invention, this step is further described below.
[0053] The random parameter search method is used to optimize the LGBM parameters. 2.5 The concentration, geographic location, meteorological conditions and aerosol optical depth data are used as input, and they are split into data sets and training sets according to a ten-to-one cross validation. The range of parameters such as the optimal leaf weight, number of leaves, maximum tree depth, feature sampling value, etc. is defined, and then the range parameters are fitted to the function value using the random parameter search method, and finally the best training parameter combination is returned.
[0054] The parameters of the LGBM machine learning model are adjusted, the optimal leaf weights are adjusted, and the limiting parameters are set to prevent the model from overfitting. The core parameters of LGBM are boosting, using the traditional gradient boosting decision tree; the number of leaves is 300, which prevents overfitting; the tree depth is 9, which improves the decision tree fitting performance; the feature sampling value is 0.92, so that some features are randomly selected for iteration in each iteration, improving the performance of iterative fitting.
[0055] Step 4: train the PM1 concentration prediction model using the training set, test the trained PM1 concentration prediction model using the test set, and estimate PM1 concentration based on the trained PM1 concentration prediction model.
[0056] See also Figure 2 The present invention divides the data into a training set and a test set, and performs ten-fold cross validation; based on the training set, the LGBM algorithm trains the fitting data to obtain a training model; the test set data is predicted according to the training model, the difference index between the actual value (i.e., the predicted value) and the true value is calculated, and the fitting result is evaluated.
[0057] When predicting, the PM2.5 data, meteorological data, geographic data and aerosol optical thickness of the PM1 concentration unknown point simulated by WRF-CHEM are input into the trained PM1 concentration prediction model, and the model outputs the predicted PM1 concentration. The present invention can estimate the distribution of PM1 concentration in the whole country and analyze the pollution degree of PM1 based on the estimated overall distribution.
[0058] Embodiment 2:
[0059] Embodiment 2 provides a PM1 concentration estimation system, comprising:
[0060] Data acquisition module, used to collect PM1 concentration information and PM1 concentration impact index data;
[0061] A correlation analysis module is used to analyze the correlation between PM1 concentration and PM1 concentration influencing indicators to obtain PM1 concentration related indicators, and to construct a training set and a test set based on the PM1 concentration information and the data of the PM1 concentration related indicators;
[0062] The model building module is used to build a PM1 concentration prediction model based on the LGBM algorithm, with PM1 concentration-related indicators as input parameters of the model and PM1 concentration as output parameters of the model;
[0063] The fitting prediction module is used to train the PM1 concentration prediction model through the training set, test the trained PM1 concentration prediction model through the test set, and realize PM1 concentration estimation based on the trained PM1 concentration prediction model.
[0064] The PM1 concentration estimation system provided in Example 2 is used to implement the steps in the PM1 concentration estimation method described in Example 1.
[0065] Finally, it should be noted that the above specific implementation methods are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to examples, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the spirit and scope of the technical solution of the present invention, which should be included in the scope of the claims of the present invention.
Claims
1. A method for estimating PM1 concentration, characterized in that: The following steps are involved: Step 1: Collect and obtain data on PM1 concentration information and PM1 concentration impact indicators; Among them, the PM1 concentration impact indicators include PM 2.5 concentration, geographic location, meteorological conditions and aerosol optical depth; the China Multi-resolution Emission Inventory Model obtained by WRF-CHEM was used to perform PM 2.5 The nationwide simulation results show that PM 2.5 Concentration data; Step 2: Analyze the correlation between PM1 concentration and PM1 concentration influencing indicators to obtain PM1 concentration related indicators; construct a training set and a test set based on the PM1 concentration information and the data of the PM1 concentration related indicators; Step 3: Construct a PM1 concentration prediction model based on the LGBM algorithm, taking PM1 concentration related indicators as input parameters of the model and PM1 concentration as output parameters of the model; Among them, the basic parameters of the LGBM algorithm are optimized by random parameter search method; Step 4: train the PM1 concentration prediction model using the training set, test the trained PM1 concentration prediction model using the test set, and estimate PM1 concentration based on the trained PM1 concentration prediction model.
2. The method for estimating PM1 concentration according to claim 1, characterized in that: The geographical location includes one or more indicators of the normalized difference vegetation index and the digital elevation model; the meteorological conditions include one or more indicators of surface relative humidity, air temperature, wind speed, surface pressure, and boundary layer height; the data of aerosol optical thickness comes from the Himawari-8 level 3 hourly data set, and uses data samples with a confidence level marked as "very good".
3. The method for estimating PM1 concentration according to claim 1, characterized in that: In step 2, the correlation coefficient between each PM1 concentration influencing indicator and the PM1 concentration is calculated, and it is judged whether the correlation coefficient corresponding to each PM1 concentration influencing indicator meets the preset conditions, and the PM1 concentration influencing indicator that meets the preset conditions is used as the PM1 concentration related indicator.
4. The method for estimating PM1 concentration according to claim 1, characterized in that: The LGBM algorithm uses a gradient boosting decision tree, and the basic parameters of the LGBM algorithm include: optimal leaf weight, number of leaves, maximum depth of the tree, and feature sampling value.
5. A PM1 concentration estimation system, characterized in that: include: Data acquisition module, used to collect PM1 concentration information and PM1 concentration impact index data; A correlation analysis module is used to analyze the correlation between PM1 concentration and PM1 concentration influencing indicators to obtain PM1 concentration related indicators, and to construct a training set and a test set based on the PM1 concentration information and the data of the PM1 concentration related indicators; The model building module is used to build a PM1 concentration prediction model based on the LGBM algorithm, with PM1 concentration-related indicators as input parameters of the model and PM1 concentration as output parameters of the model; A fitting prediction module, used to train the PM1 concentration prediction model through the training set, test the trained PM1 concentration prediction model through the test set, and realize PM1 concentration estimation based on the trained PM1 concentration prediction model; The PM1 concentration estimation system is used to implement the steps in the PM1 concentration estimation method as described in any one of claims 1-4.