A real-time source analysis method for VOCs

Abstract

The invention relates to a VOCs real-time source analysis method. The method comprises the following steps: performing species screening, data interpolation below a detection limit and missing value interpolation preprocessing on monitoring data; organizing the preprocessed monitoring data into input data and incorporating them into a PMF model for analysis; calculating and obtaining a real source contribution value; taking the preprocessed monitoring data and the real source contribution value obtained by PMF calculation as sample data sets, and randomly dividing them into three parts: a training data set, a verification data set and a test data set; the training data set and the verification data set are used for back propagation neural network model training; the test data set is used for model accuracy verification; and establishing a functional relationship model between monitoring data and source contribution, thereby performing real-time source analysis on VOCs monitoring data obtained in real time, improving the timeliness of VOCs source analysis results from a lag of several days to real time, and providing support for ozone pollution emergency response.

Description

Technical Field

[0001] The present invention relates to the application field of VOCs pollution prevention and control technology, and in particular to a VOCs real-time source analysis method. Background Art

[0002] With the acceleration of urbanization, the emission density of ozone (O3) precursors such as volatile organic compounds (VOCs) and nitrogen oxides (NOx) remains high, and the O3 pollution problem has become increasingly prominent. Studies have shown that VOCs are important precursors of O3 and secondary organic aerosols, and studies have shown that ozone pollution in most Chinese cities belongs to VOCs control areas. Therefore, conducting VOCs source analysis research is of great significance for formulating accurate and efficient air pollution prevention and control strategies and achieving coordinated control of O3 and PM2.5.

[0003] At present, VOCs source analysis methods include source analysis methods based on receptor models and source analysis methods based on air quality models. Among them, the VOCs source analysis method based on the receptor model does not require detailed meteorological forecast information and pollution source emission inventories, and is therefore more widely used. The most important model used in the receptor model method is the PMF model. In recent years, with the popularization of high-time-resolution VOCs online monitoring instruments, high-time-resolution VOCs monitoring data have been widely used in the research field of VOCs pollution characteristics, sources, and ozone causes. Therefore, many VOCs online source analysis studies based on the PMF model have been carried out at home and abroad.

[0004] The process of VOCs online source analysis method based on PMF model can be simply summarized as follows: organize VOCs online monitoring data within a certain monitoring period into an input data set, incorporate it into the PMF model for factor analysis, and interpret the obtained factor spectrum to identify the main VOCs emission sources and quantify the source contribution. The VOCs source analysis method based on PMF model can realize the source analysis of VOCs historical data over a period of time, but there is currently a lack of technology for real-time source analysis of VOCs, which makes it difficult to meet the timeliness requirements of environmental management in ozone pollution emergency work. Therefore, there is an urgent need to develop VOCs real-time source analysis technology to meet timeliness requirements. Summary of the invention

[0005] The purpose of the present invention is to address the problem that the existing VOCs online source analysis based on the PMF model has a lag and cannot meet the timeliness requirements of ozone pollution emergency work. A VOCs real-time source analysis method is proposed to achieve source analysis of real-time monitored VOCs data, thereby quantifying the contribution of various emission sources to VOCs in real time, and providing support for ozone pollution prevention and control.

[0006] To achieve the above object, the present invention provides a VOCs real-time source analysis method, which comprises the following steps:

[0007] Step S1, automatically obtaining hourly monitoring data of VOCs monitoring equipment through a python program;

[0008] Step S2, preprocessing the monitoring data by species screening, interpolation of data below the detection limit, and interpolation of missing values;

[0009] Step S3, organizing the pre-processed monitoring data into input data and incorporating it into the PMF model for analysis, and calculating the real source contribution value;

[0010] Step S4, taking the pre-processed monitoring data and the real source contribution value obtained by PMF calculation as sample data sets, randomly dividing them into three parts: training data set, verification data set and test data set. The training data set and verification data set are used for training the BP neural network model (back propagation neural network model), and the test data set is used for model accuracy verification to obtain a trained BP neural network model;

[0011] Step S5, after preprocessing the VOCs real-time monitoring data, input it into the trained BP neural network model to predict the VOCs source analysis results.

[0012] Preferably, in step S1, the monitoring data is derived from the monitoring values ​​of an online monitoring instrument for VOCs components, including common species such as alkanes, alkenes, alkynes, aromatic hydrocarbons, halogenated hydrocarbons and OVOCs; the known source contribution is derived from the source contribution value obtained by source parsing the receptor data using a PMF model.

[0013] Preferably, in step S2, the pre-processing process for species screening includes:

[0014] The species obtained from VOCs monitoring were screened, and species with strong source indicators, species with higher volume fractions or mass concentrations, and species with relatively weak atmospheric photochemical reaction activity were selected. Then, for species with concentrations lower than the instrument detection limit, the sample concentration was replaced by 1 / 2*MDL (instrument detection limit), and missing values ​​were replaced by 1 / 2*MDL.

[0015] Preferably, in step S3, the process of calculating the real source contribution using the PMF model includes:

[0016] The preprocessed monitoring data are organized into input data sets and incorporated into the PMF model for analysis. Different numbers of factors are set and the rationality of the results of different factor numbers is compared to determine the optimal number of factors. The optimal number of factors is then used for model calculations, the obtained factors are source identified, and the contribution of each source is quantified.

[0017] Preferably, the method for collating the input data set comprises:

[0018] In addition to the preprocessed monitoring data, the uncertainty data corresponding to the monitoring data also needs to be calculated. A unified uncertainty coefficient of 0.3 is multiplied by the VOCs component concentration monitoring value to obtain the uncertainty corresponding to the concentration monitoring value. The VOCs component monitoring data and the uncertainty data together constitute the input data set.

[0019] Preferably, in step S4, the sample data set is randomly divided into 10 parts using a "ten-fold" cross-validation method, 8 of which are used as training data and 1 as verification data. The training data and verification data are used for model training, and the other is used as test data for model accuracy verification.

[0020] Preferably, the topological structure of the BP neural network model consists of an input layer, a hidden layer and an output layer, the number of nodes in the input layer is the number of species screened out during the receptor data preprocessing process, the number of nodes in the output layer is the number of source classes contributed by the known sources, and the number of neurons in the hidden layer is determined by trial and error;

[0021] The hidden layer uses a nonlinear activation function Sigmoid, and the output layer uses a nonlinear activation function softmax.

[0022] Preferably, the back propagation neural network model training adopts the Adam function, sets the maximum number of iterations to 300 times, and automatically determines the weight coefficients and thresholds of the final model through iterative calculation to complete the model training.

[0023] Preferably, the step of verifying the model accuracy includes:

[0024] The Pearson correlation coefficient r between the predicted value and the true value, the mean absolute error AAE and the root mean square error RMSE were used as the evaluation criteria of the model to calculate the calculation accuracy of the contribution to the VOCs source.

[0025] Based on the above technical solution, the advantages of the present invention are:

[0026] The VOCs real-time source analysis method of the present invention solves the problem that the existing VOCs source analysis technology has obvious lag in obtaining the source analysis results. The existing VOCs source analysis technology incorporates the VOCs monitoring data accumulated for a certain period of time into the PMF model for analysis to obtain the VOCs source analysis results within the period. Since the accumulation of data and the PMF model calculation require a lot of time, the results obtained are often greatly lagged behind the data monitoring. The VOCs real-time source analysis method of the present invention can perform real-time source analysis on the VOCs monitoring data obtained in real time, so that the timeliness of the VOCs source analysis results is improved from a lag of several days to real time, which can provide support for ozone pollution emergency response. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of this application. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0028] Figure 1 It is a flow chart of a VOCs real-time source analysis method;

[0029] Figure 2 A step diagram of a VOCs real-time source analysis method;

[0030] Figure 3 A result diagram of an embodiment of a VOCs real-time source analysis method;

[0031] Figure 4 This is a diagram for verifying the accuracy of an embodiment of a VOCs real-time source analysis method. DETAILED DESCRIPTION

[0032] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments.

[0033] The present invention provides a VOCs real-time source analysis method, such as Figures 1 to 4 As shown, a preferred embodiment of the present invention is shown.

[0034] Since the existing VOCs online source analysis technology based on the PMF model requires the accumulation of a certain amount of VOCs monitoring data to be included in the PMF model for analysis, it is necessary to manually diagnose and screen the PMF model results in order to obtain the source analysis results corresponding to the VOCs monitoring data. Therefore, it has a lag and cannot meet the timeliness requirements of ozone pollution emergency work.

[0035] The present invention proposes a VOCs real-time source analysis method to meet the timeliness requirements. Specifically, Figure 1 , Figure 2As shown, the VOCs real-time source analysis method includes the following steps:

[0036] Step S1, automatically obtain the hourly monitoring data of the VOCs monitoring equipment through the Python program.

[0037] Step S2, preprocessing the monitoring data by species screening, interpolation of data below the detection limit, and interpolation of missing values.

[0038] Step S3, organize the pre-processed monitoring data into input data and incorporate it into the PMF model for analysis to calculate the real source contribution value.

[0039] Step S4, taking the preprocessed monitoring data and the real source contribution value obtained by PMF calculation as sample data sets, randomly dividing them into three parts: training data set, verification data set and test data set. The training data set and verification data set are used for BP neural network model training, and the test data set is used for model accuracy verification to obtain a trained BP neural network model.

[0040] Step S5, after preprocessing the VOCs real-time monitoring data, input it into the trained BP neural network model to calculate the corresponding VOCs source analysis results.

[0041] Preferably, in step S1, the monitoring data is derived from the monitoring values ​​of an online monitoring instrument for VOCs components, including common species such as alkanes, alkenes, alkynes, aromatic hydrocarbons, halogenated hydrocarbons and OVOCs; the known source contribution is derived from the source contribution value obtained by source parsing the receptor data using a PMF model.

[0042] The monitoring data preprocessing process is as follows: the species obtained from VOCs monitoring are screened, and the component screening considers the following three principles: 1) Select species with strong source indicators; 2) Select species with high volume fraction or mass concentration; 3) Select species with relatively weak atmospheric photochemical reaction activity. Afterwards, for species with concentrations lower than the instrument detection limit, the sample concentration is replaced by 1 / 2 of the instrument detection limit, and for missing values, 1 / 2 of the instrument detection limit is used to replace;

[0043] In step S3, the process of calculating the real source contribution by the PMF model includes: organizing the preprocessed monitoring data into an input data set, incorporating it into the PMF model for analysis, setting different numbers of factors and comparing the rationality of the results of different numbers of factors, establishing the optimal number of factors, and then using the optimal number of factors for model calculations, performing source identification on the obtained factors, and quantifying the contribution of each source.

[0044] Preferably, the method for collating the input data set includes: in addition to the preprocessed monitoring data, uncertainty data corresponding to the monitoring data needs to be calculated, and a unified uncertainty coefficient of 0.3 is multiplied by the VOCs component concentration monitoring value to obtain the uncertainty corresponding to the concentration monitoring value. The VOCs component monitoring data and the uncertainty data together constitute the input data set.

[0045] In step S4, the sample data set is randomly divided into 10 parts using a "ten-fold" cross-validation method, 8 of which are used as training data and 1 as verification data. The training data and verification data are used for model training, and the other is used as test data for model accuracy verification.

[0046] Preferably, the back propagation neural network model (BP neural network model) is built based on Python's TensorFlow mathematical system. The topological structure of the BP neural network consists of three layers: an input layer, a hidden layer, and an output layer. The number of nodes in the input layer is the number of species screened out during the receptor data preprocessing process, the number of nodes in the output layer is the number of source classes contributed by the known source, and the number of neurons in the hidden layer is determined by trial and error, that is, repeatedly trying different numbers of nodes to determine the optimal number of neurons. The hidden layer uses the nonlinear activation function Sigmoid, and the output layer uses the nonlinear activation function softmax. The model training uses the Adam (Adaptive Moment Estimation method Adaptive Moment Estimation) function, sets the maximum number of iterations to 300 times, and automatically determines the weight coefficients and thresholds through iterative calculations to complete the model training.

[0047] More preferably, the step of verifying the model accuracy comprises:

[0048] The Pearson correlation coefficient r between the predicted value and the true value, the mean absolute error AAE and the root mean square error RMSE were used as the evaluation criteria of the model to calculate the calculation accuracy of the contribution to the VOCs source.

[0049] The VOCs real-time source analysis method of the present invention utilizes training data (VOCs component monitoring data and corresponding real source contributions) to train a BP neural network, thereby being able to perform real-time source analysis on the VOCs monitoring data obtained in real time, thereby increasing the timeliness of the VOCs source analysis results from a lag of several days to real time, and providing support for ozone pollution emergency response.

[0050] To further explain the VOCs real-time source analysis method of the present invention, an example is given below to illustrate it.

[0051] This example uses a python program to automatically obtain historical VOCs monitoring data and real-time monitoring data. The VOCs component online monitoring data comes from the monitoring of the VOCs component online monitoring instrument, and the monitoring items generally include common alkanes, alkenes, alkynes, aromatic hydrocarbons, halogenated hydrocarbons and OVOCs.

[0052] The VOCs component monitoring data collected above are subjected to preprocessing processes such as species screening, interpolation of data below the detection limit, and interpolation of missing values ​​to construct the PMF input data set. Species screening needs to comprehensively consider the following three principles to select suitable components from the various components monitored by VOCs. The principles include: 1) Selecting species with strong source indicators; 2) Selecting species with higher volume fractions or mass concentrations; 3) Selecting species with relatively weak atmospheric photochemical reaction activity. Interpolation of data below the detection limit is to replace the monitoring values ​​of the screened components whose concentrations are lower than the instrument detection limit with 1 / 2 of the instrument detection limit. Interpolation of missing values ​​is to replace the missing concentrations of the screened components with 1 / 2 of the instrument detection limit.

[0053] The above pre-treated VOCs component monitoring data is organized into input data and incorporated into the PMF model for analysis to obtain the real source contribution value. The input data must include the pre-processed monitoring data and uncertainty data. The uncertainty data is calculated by multiplying the VOCs component concentration monitoring value by a unified uncertainty coefficient of 0.3. The above input data is incorporated into the PMF model, different numbers of factors are set and the rationality of the results of different numbers of factors is compared to establish the optimal number of factors. The optimal number of factors is selected for model calculation, the pollution sources are identified for the obtained factors, and the contribution values ​​of each source are quantified.

[0054] The above VOCs component monitoring data and the known source contributions calculated by the PMF model are randomly divided into three parts: training data set, verification data set and test data set. The training data set and verification data set are used for BP neural network model training, and the test data set is used for model accuracy verification.

[0055] The BP neural network is built based on Python's TensorFlow mathematical system, and the model structure includes an input layer, a hidden layer, and an output layer. The known source contribution is the output parameter of the BP model, and the VOCs component monitoring data is the input parameter of the model. Therefore, the number of nodes in the input layer is the number of species screened out during the receptor data preprocessing process, and the number of nodes in the output layer is the number of source classes contributed by the known source. The number of neurons in the hidden layer is determined by trial and error, that is, repeatedly trying different numbers of nodes to determine the optimal number of neurons. The hidden layer uses the nonlinear activation function Sigmoid, and the output layer uses the nonlinear activation function softmax. The model training uses the Adam (Adaptive Moment Estimation method Adaptive Moment Estimation) function, sets the maximum number of iterations to 300 times, and automatically determines the weight coefficients and thresholds through iterative calculations of the final model to complete the model training.

[0056] After data preprocessing, the VOCs real-time monitoring data can be input into the above-mentioned trained model to calculate the corresponding source analysis results.

[0057] The results of the embodiment of the VOCs real-time source analysis method of the present invention are shown in Figure 3 shown. Figure 4 The Pearson correlation coefficient r between the simulated value and the true value of the test data set is between 0.62 and 0.93, the AAE is between 1.9 and 3.9%, and the RMSE is between 0.03 and 0.06, indicating that the calculation results of the source analysis method proposed in the present invention are accurate, indicating the feasibility of the present invention.

[0058] Finally, it should be noted that the above embodiments 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 the preferred embodiments, ordinary technicians in the field should understand that the specific implementation methods of the present invention can still be modified or some technical features can be replaced by equivalents without departing from the spirit of the technical solution of the present invention, which should be included in the scope of the technical solution for protection of the present invention.

Claims

1. A VOCs real-time source analysis method, characterized by: The VOCs real-time source analysis method comprises the following steps: Step S1, automatically obtaining hourly monitoring data of VOCs monitoring equipment through a python program; Step S2, pre-processing the monitoring data by species screening, interpolation of data below the detection limit, and interpolation of missing values, wherein the pre-processing process of species screening includes: screening the species obtained by VOCs monitoring, selecting species with strong source indicators, selecting species with high volume fractions or mass concentrations, and selecting species with relatively weak atmospheric photochemical reaction activity, and then replacing the sample concentration with 1 / 2*MDL for species concentrations below the instrument detection limit, and replacing the missing values ​​with 1 / 2*MDL, wherein MDL represents the instrument detection limit; Step S3, arranging the pre-processed monitoring data into an input data set and incorporating it into the PMF model for analysis, and calculating the real source contribution value. The input data set arranging method includes: in addition to the pre-processed monitoring data, it is also necessary to calculate the uncertainty data corresponding to the monitoring data, and use a unified uncertainty coefficient of 0.3 to multiply the VOCs component concentration monitoring value to obtain the uncertainty corresponding to the concentration monitoring value. The VOCs component monitoring data and the uncertainty data together constitute the input data set. The process of calculating the real source contribution by the PMF model includes: arranging the pre-processed monitoring data into an input data set, incorporating it into the PMF model for analysis, setting different factor numbers and comparing the rationality of the results of different factor numbers, establishing the optimal factor number, and then using the optimal factor number for model calculation, performing source identification on the obtained factors, and quantifying the contribution of each source; Step S4, taking the pre-processed monitoring data and the real source contribution value obtained by PMF calculation as sample data sets, randomly dividing them into three parts: training data set, verification data set and test data set, wherein the training data set and verification data set are used for back propagation neural network model training, and the test data set is used for model accuracy verification, so as to obtain a trained BP neural network model; the back propagation neural network model training adopts Adam function, sets the maximum number of iterations to 300 times, and automatically determines the weight coefficient and threshold value of the final model through iterative calculation to complete the model training; Step S5, after preprocessing the VOCs real-time monitoring data, input it into the trained BP neural network model to calculate the corresponding VOCs source analysis results.

2. The VOCs real-time source analysis method according to claim 1, characterized in that: In step S1, the monitoring data is derived from the monitoring values ​​of the VOCs component online monitoring instrument, including common alkanes, alkenes, alkynes, aromatic hydrocarbons, halogenated hydrocarbons and OVOCs species.

3. The VOCs real-time source analysis method according to claim 1, characterized in that: In step S4, the sample data set is randomly divided into 10 parts using a "ten-fold" cross-validation method, 8 of which are used as training data and 1 as verification data. The training data and verification data are used for model training, and the other is used as test data for model accuracy verification.

4. The VOCs real-time source analysis method according to claim 1, characterized in that: The topological structure of the back propagation neural network model consists of an input layer, a hidden layer and an output layer. The number of nodes in the input layer is the number of species screened out during the receptor data preprocessing process, the number of nodes in the output layer is the number of source classes with known source contributions, and the number of neurons in the hidden layer is determined by trial and error; the known source contribution comes from the source contribution value obtained by source parsing the receptor data using the PMF model; The hidden layer uses a nonlinear activation function Sigmoid, and the output layer uses a nonlinear activation function softmax.

5. The VOCs real-time source analysis method according to claim 1, characterized in that: The steps of model accuracy verification include: The Pearson correlation coefficient r between the predicted value and the true value, the mean absolute error AAE and the root mean square error RMSE were used as the evaluation criteria of the model to calculate the calculation accuracy of the contribution to the VOCs source.

Images