Atmospheric black carbon source analysis method and system fusing multi-dimensional elements and physical constraints

Abstract

The invention discloses a multi-dimensional element and physical constraint fused atmospheric black carbon source analysis method and system, and the method comprises the steps: obtaining a ratio of BC total mass concentration to a primary light absorption coefficient, and calculating a primary BC light absorption coefficient according to the BC total mass concentration and the ratio; carrying out pretreatment on the primary BC light absorption coefficient; sending the preprocessed primary BC light absorption coefficient and the obtained environmental data into a pre-trained neural network model to obtain light absorption coefficients, BC total mass concentration and BC mass concentration discharged by each discharge source under different wavebands output by the neural network model; when the neural network model is constructed, an activation function and a loss function are defined, and training is completed. The method is based on the atmospheric environment data, combines the physical process of atmospheric BC quality and light absorption coefficient change, overcomes the defect that a traditional neural network model lacks physical significance, and is suitable for BC source analysis tasks in different regions and different atmospheric environments.

Description

Technical Field

[0001] This invention relates to the field of pollutant source apportionment technology, and in particular to a method and system for apportioning atmospheric black carbon sources by integrating multidimensional factors and physical constraints. Background Technology

[0002] Black carbon (BC) plays a crucial role in regulating the global climate system. As an effective light-absorbing substance, BC warms the atmosphere by directly absorbing solar radiation, profoundly impacting the Earth's energy balance and ecosystems. Identifying the potential sources of BC is essential for effectively managing BC pollution and developing effective control strategies. However, current BC optical source apportionment models, such as the Aethalometer Model (AE model), can only distinguish between two sources, failing to meet the source apportionment needs of regions with complex emission sources. Furthermore, the AE model does not consider the light absorption effect of secondary brown carbon when performing BC source apportionment. Studies show that secondary brown carbon can absorb 15% (660nm)–27% (370nm) of the total light absorption, and the light absorption contribution of secondary brown carbon varies significantly among different emission sources, ranging from 2% to 23%. Therefore, failing to account for the light absorption effect of secondary brown carbon may lead to misinterpretations of BC optical source apportionment results.

[0003] Source apportionment methods for black carbon (BC) can be broadly classified into four categories. The first category is acceptor model apportionment (such as the chemical mass balance model CMB and the orthogonal matrix factorization model PMF); the second category is isotope source apportionment; the third category is model simulation; and the fourth category is source apportionment based on online optical measurements (such as the AE model).

[0004] Chemical mass balance (CMB) models require local emission source spectrum data to analyze the sources of carbon dioxide (BC). However, the application of CMB is significantly limited due to the lack of high-quality local source spectrum data in many cities. Orthogonal matrix factorization (PMF) models decompose the total atmospheric BC concentration into a factor spectrum matrix and a source contribution matrix to distinguish different sources and their contributions to BC concentration. However, this method is based solely on chemical mass balance analysis and relies on comprehensive chemical component data; insufficient data volume and component variety limit its application. Furthermore, it is usually necessary to compare the correlation and diurnal variation characteristics of the analyzed source contributions with indicator substances and to robustly validate the factor spectra to ensure the reliability of the analysis results. Therefore, before final confirmation of the analysis results, multiple iterations of analysis, validation, and comparison are typically required until all indicators are reasonable. Isotope source analysis has the highest accuracy, but its time resolution is low due to limitations in sample loading, making it difficult to provide high-temporal-resolution analysis results. Therefore, this method is not suitable for source tracing analysis of pollution events, and its application in pollution source control is also limited. Model simulations can provide high temporal resolution source apportionment results for anthropogenic sources, but the simulation results often deviate significantly from actual observations. The main factors contributing to this deviation include: simulation errors related to meteorological conditions; uncertainties in emission inventory data; and simplification assumptions inherent in the models themselves. For example, there are estimation errors in the transport, diffusion, and deposition rates of BC. In contrast, BC source apportionment based on online optical measurements offers even higher temporal resolution and is more suitable for tracking the dynamic evolution of BC sources. However, this method can only distinguish a limited number of pollution source categories, typically differentiated between fossil fuel and biomass combustion emissions, or liquid fossil fuel and solid fuel emissions. This limitation makes it difficult to further classify anthropogenic BC sources into more refined categories, such as further distinguishing between vehicle emissions, biomass combustion, and coal combustion.

[0005] Therefore, a more comprehensive online source apportionment method is urgently needed to fill this research gap. Furthermore, the application of machine learning techniques has provided some improvements to BC source apportionment, enhancing the accuracy of the results to a certain extent. However, traditional machine learning methods primarily rely on data-driven approaches. While they can reveal numerical relationships between variables, they lack physical interpretability, making it difficult to ensure that the apportionment results conform to real physical laws. Therefore, machine learning methods alone cannot directly determine the scientific validity of the apportionment results; verification with external observational data is still necessary to ensure their accuracy and reliability. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a method and system for analyzing the sources of atmospheric black carbon by integrating multi-dimensional factors and physical constraints, thereby solving the problems mentioned in the background art.

[0007] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A method for apportioning atmospheric black carbon sources that integrates multidimensional factors and physical constraints includes: Step 1: Obtain the ratio of the total mass concentration of BC to the primary absorption coefficient, and calculate the primary BC absorption coefficient based on the ratio of the total mass concentration of BC. Step 2: Preprocess the first BC absorption coefficient; Step 3: Input the preprocessed BC absorption coefficient and the acquired environmental data into the pre-trained neural network model to obtain the absorption coefficient, total BC mass concentration and BC mass concentration of each emission source under different bands output by the neural network model. The neural network model defines activation and loss functions during construction. Then, it takes the BC absorption coefficient and environmental data as inputs and the absorption coefficient, total BC mass concentration, and BC mass concentration emitted from each emission source at different wavelengths as outputs to complete the training.

[0008] Furthermore, the construction of the neural network model includes: An input layer is established based on the input, and the input layer is connected to a fully connected hidden layer. A dropout layer is added to each hidden layer, and the dropout rate, activation function and weights in the neural network layer are set. A parallel output layer is established to output the AAE layer controlled by each emission source, the Mass layer of BC emitted by each emission source, and the MAE layer in the 880nm band, respectively. At the same time, the activation function of the AAE layer is determined according to the different absorption angstrom index of each emission source of BC. The output layer is then connected to the hidden layer. A light absorption coefficient layer and a total mass concentration (BC) layer were constructed in the 370nm and 880nm wavelength bands, and then connected to the output layer. Establish loss functions for the total mass concentration of BC and the mass concentration of BC emitted from each emission source, respectively. Then, the neural network is trained using the BC absorption coefficient and environmental data as inputs, and the absorption coefficients, total BC mass concentrations, and BC mass concentrations emitted from various emission sources at different wavelengths as outputs. During training, the neural network model is corrected based on the loss value of the loss function until training is complete. Output the trained neural network model.

[0009] Furthermore, the expression for the activation function of the AAE layer is as follows: scaled_output = min_val + (max_val - min_val) ×sigmoid(x) In the formula, scaled_output is the activation function of the AAE layer, min_val is the calculated value of the lower limit of the emission source absorption angstrom index, max_val is the calculated value of the upper limit of the emission source absorption angstrom index, and sigmoid(x) is the sigmoid activation function.

[0010] Furthermore, the loss function of the total mass concentration of BC as follows:

[0011] In the formula, where For loss function The scaling value, For the Pearson correlation coefficient, we have:

[0012] In the formula, where and They are respectively a source number The mass concentration of each emission source and the predicted mass concentration of BC emitted from that emission source. and These are the average mass concentrations of each emission source and the predicted average mass concentration of BC emitted by that emission source, respectively, where n is the number of data samples used to calculate the gradient and update the model parameters in each iteration.

[0013] Furthermore, the various emission sources include coal combustion emissions, biomass combustion emissions, and motor vehicle emissions.

[0014] Furthermore, the loss function for the mass concentration of BC emitted from each emission source is as follows:

[0015]

[0016] + +

[0017] In the formula, , as well as Both are loss functions. , The absorption coefficients are for the 880nm and 370nm wavelength bands, respectively. , , The mass absorption cross section of BC for coal combustion emissions, the mass absorption cross section of BC for biomass combustion emissions, and the mass absorption cross section of BC for motor vehicle emissions are given. , , These are the mass emissions of BC from coal combustion, the mass emissions of BC from biomass combustion, and the mass emissions of BC from motor vehicles. , , These are the ranges for the angstrom absorption index of coal combustion, the angstrom absorption index of biomass combustion, and the angstrom absorption index of motor vehicles, respectively. The total mass concentration of BC.

[0018] Furthermore, during the training of the neural network model, an adaptive learning rate optimization algorithm is used to set the learning rate, and a callback function is used to dynamically adjust the learning rate during the training process.

[0019] Furthermore, in step 2, the first BC absorbance coefficient is preprocessed, including handling missing values ​​and handling outliers.

[0020] Furthermore, the processing of missing values ​​includes: Missing data is filled in by means of the mean, median, interpolation, or by removing missing values. The processing of outliers includes: Outliers are identified and processed using statistical methods.

[0021] Based on the same inventive concept, this invention also proposes an atmospheric black carbon source apportionment system that integrates multidimensional factors and physical constraints, comprising: The data acquisition module acquires the ratio of the total mass concentration of BC to the primary absorption coefficient, and calculates the primary BC absorption coefficient based on the ratio of the total mass concentration of BC. The data processing module preprocesses the first BC absorption coefficient; The model calculation module inputs the preprocessed BC absorption coefficient and the acquired environmental data into a pre-trained neural network model to obtain the absorption coefficient, total BC mass concentration, and BC mass concentration of each emission source under different wavelengths output by the neural network model. The neural network model defines activation and loss functions during construction. Then, it takes the BC absorption coefficient and environmental data as inputs and the absorption coefficient, total BC mass concentration, and BC mass concentration emitted from each emission source at different wavelengths as outputs to complete the training.

[0022] Compared with the prior art, the present invention has the following advantages: This invention provides a method and system for apportioning atmospheric black carbon sources by integrating multidimensional factors and physical constraints. Based on atmospheric environmental data (including meteorological factors and air pollutants), and combined with the physical processes of atmospheric brown carbon (BC) mass and absorption coefficient changes, it overcomes the deficiency of traditional neural network models lacking physical meaning, thus improving the scientific rigor and applicability of model predictions. Compared with traditional receptor models, this invention effectively eliminates the perturbation of BC absorption coefficient caused by secondary brown carbon formation, making the apportionment results more accurate. Simultaneously, it fully considers the influence of atmospheric environmental factors on BC absorption capacity and mass concentration, eliminating the need to rely on emission source spectra or repeated analysis to verify BC source contributions. Compared with traditional machine learning models, this invention, by embedding physical processes and combining the absorption Å index range of BC from different emission sources, makes the model output physically interpretable, avoiding the limitations of "black box models" and providing a high-temporal-resolution and interpretable analytical tool for atmospheric BC source apportionment. Furthermore, the model is highly flexible; the input features and activation functions can be adjusted according to specific data and the absorption Å index range of different emission sources, making it suitable for BC source apportionment tasks in different regions and under different atmospheric environments.

[0023] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0024] Figure 1 This is a flowchart illustrating a method for apportioning atmospheric black carbon sources that integrates multidimensional factors and physical constraints.

[0025] Figure 2 This is a schematic diagram of the neural network model of the present invention.

[0026] Figure 3 This is a schematic diagram of the detection of the analysis results of the present invention.

[0027] Figure 4 This is a schematic diagram illustrating the model analysis of the present invention.

[0028] Figure 5 This is a schematic diagram of a system for apportioning the sources of atmospheric black carbon that integrates multidimensional factors and physical constraints. Detailed Implementation

[0029] like Figure 1 As shown, the present invention provides a method for analyzing the sources of atmospheric black carbon by integrating multidimensional factors and physical constraints, comprising: Step 1: Obtain the ratio of the total mass concentration of BC to the primary absorption coefficient, and calculate the primary BC absorption coefficient based on the ratio of the total mass concentration of BC. Step 2: Preprocess the first BC absorption coefficient; Step 3: Input the preprocessed BC absorption coefficient and the acquired environmental data into the pre-trained neural network model to obtain the absorption coefficient, total BC mass concentration and BC mass concentration of each emission source under different bands output by the neural network model. The neural network model defines activation and loss functions during construction. Then, it takes the BC absorption coefficient and environmental data as inputs and the absorption coefficient, total BC mass concentration, and BC mass concentration emitted from each emission source at different wavelengths as outputs to complete the training.

[0030] In this invention, the emission sources include coal combustion emissions, biomass combustion emissions, and motor vehicle emissions.

[0031] This invention proposes an innovative method for source apportionment of atmospheric brown carbon (BC) based on atmospheric environmental data (including meteorological factors and air pollutants) and the physical processes underlying changes in atmospheric BC mass and absorption coefficient. This method overcomes the deficiency of traditional neural network models in lacking physical meaning, improving the scientific rigor and applicability of model predictions. Compared to traditional receptor models, this invention effectively eliminates the perturbation of BC absorption coefficient caused by secondary brown carbon formation, resulting in more accurate apportionment results. Simultaneously, it fully considers the influence of atmospheric environmental factors on BC absorption capacity and mass concentration, eliminating the need to rely on emission source spectra or repeated analysis to verify BC source contributions. Compared to traditional machine learning models, this invention, by embedding physical processes and incorporating the absorption Å index range of BC from different emission sources, makes the model output physically interpretable, avoiding the limitations of "black box models" and providing a high-temporal-resolution and interpretable analytical tool for atmospheric BC source apportionment. Furthermore, the model is highly flexible; the input features and activation functions can be adjusted according to specific data and the absorption Å index range of different emission sources, making it suitable for BC source apportionment tasks in different regions and under different atmospheric conditions.

[0032] In step 1, the ratio of the total mass concentration of BC to the primary absorption coefficient is obtained, and the primary BC absorption coefficient is calculated based on the ratio of the total mass concentration of BC.

[0033] Hourly data of pollutant concentrations, total biomass concentration (BC), and absorbance coefficients at different wavelengths will be stored in a tabular file. Other environmental data can also be acquired as features based on specific needs, provided the resolution matches the BC total mass concentration and absorbance coefficient. Mass concentration data of indicative substances for coal combustion emissions, biomass combustion emissions, and vehicle emissions will be obtained. The indicative substance data will be determined based on the specific source type being analyzed. In this study, the selected indicators for coal combustion emissions are its coal-fired organic matter emissions concentration (CCOA), biomass combustion organic matter emissions concentration (BBOA), and vehicle emissions organic matter concentration (HOA).

[0034] Simultaneously, obtain or calculate the ratio of the first absorption coefficient to the BC mass concentration. The absorption coefficient of primary emission BC is calculated using the following formula:

[0035] in for The absorption coefficient of the next emission of BC in the band. for Total absorption coefficient of particulate matter in the specified band. This represents the mass concentration of black carbon.

[0036] This invention obtains or calculates the ratio of the primary absorbance coefficient to the BC mass concentration (…). This effectively eliminates the disturbance to the BC absorption coefficient caused by the formation of secondary brown carbon.

[0037] In step 2, the first BC absorption coefficient is preprocessed.

[0038] This invention provides two processing methods: handling missing values ​​and handling outliers.

[0039] The missing value handling involves filling in the missing data using methods such as mean, median, interpolation, or deletion. Missing values ​​are filled using methods such as mean, median, interpolation, or deletion. In this invention, since missing values ​​are relatively few, if there are missing values ​​for two consecutive time steps, the missing samples are deleted; if there is a missing value for a single time step, the average of the previous and next time steps is used for interpolation. Outlier handling involves identifying and processing outliers statistically. Outlier handling uses statistical methods (such as the 3σ principle) to identify and process outliers.

[0040] After data preprocessing, the data is scaled, that is, normalized. The Min-Max normalization method is used to standardize the input features in the dataset, reducing the impact of differences in units and orders of magnitude on the model.

[0041] After processing the data, we input it into the model. In step 3, the preprocessed BC absorption coefficient and the acquired environmental data are fed into a pre-trained neural network model to obtain the absorption coefficient, total BC mass concentration, and BC mass concentration of each emission source under different wavelengths output by the neural network model.

[0042] The core content of this invention is the neural network model, and its construction process includes: An input layer is established based on the input, and the input layer is connected to a fully connected hidden layer. A dropout layer is added to each hidden layer, and the dropout rate, activation function and weights in the neural network layer are set. A parallel output layer is established to output the AAE layer controlled by each emission source, the Mass layer of BC emitted by each emission source, and the MAE layer in the 880nm band, respectively. At the same time, the activation function of the AAE layer is determined according to the different absorption angstrom index of each emission source of BC. The output layer is then connected to the hidden layer. A light absorption coefficient layer and a total mass concentration (BC) layer were constructed in the 370nm and 880nm wavelength bands, and then connected to the output layer. Establish loss functions for the total mass concentration of BC and the mass concentration of BC emitted from each emission source, respectively. Then, the neural network is trained using the BC absorption coefficient and environmental data as inputs, and the absorption coefficients, total BC mass concentrations, and BC mass concentrations emitted from various emission sources at different wavelengths as outputs. During training, the neural network model is corrected based on the loss value of the loss function until training is complete. Output the trained neural network model.

[0043] In this invention, the architecture diagram of the neural network model is as follows: Figure 2 As shown. Before constructing, we first define the activation function and the loss function.

[0044] Activation function: Three activation functions are established for the main emission sources of BC (coal combustion emissions, biomass combustion, and vehicle emissions). The activation range is controlled by the Absorbed Angle Index (AAE) of each emission source BC. The activation range of the activation functions for the three representative sources of the model is set according to the AAE of different emission sources BC. The calculation of the activation range is as follows, that is, the expression of the activation function of the AAE layer is as follows: scaled_output = min_val + (max_val - min_val) ×sigmoid(x) In the formula, scaled_output is the activation function of the AAE layer, min_val is the calculated value of the lower limit of the emission source absorption angstrom index, max_val is the calculated value of the upper limit of the emission source absorption angstrom index, and sigmoid(x) is the sigmoid activation function.

[0045] The formulas for calculating min_val and max_val are as follows:

[0046] Where value is the min_val and max_val value, and AAE is the upper or lower limit of the absorption angstrom index of each source.

[0047] Loss function: The concentrations of the obtained indicator substances from each source were standardized, converting the characteristic values ​​into a distribution with a mean of 0 and a standard deviation of 1. The standardization formula is as follows:

[0048] Among them Standardized values, The original value, This is the average value. The standard deviation is denoted as .

[0049] A model was established using the Pearson correlation coefficient analysis method to predict the correlation between the mass concentrations of BC from each source and the mass concentrations of indicator substances from each source. The Pearson correlation coefficient was calculated. The calculation formula is as follows:

[0050] Where r is the Pearson correlation coefficient. and These are the mass concentration of the i-th indicator from a certain source and the predicted mass concentration of BC emitted from that source, respectively. and These are the average mass concentration of a source indicator and the average mass concentration of the predicted emission BC from that source, respectively, where n is the number of data samples used to calculate the gradient and update the model parameters in each iteration.

[0051] The loss function for the mass concentration of each source BC in the model is established using the following formula:

[0052] Where loss is the loss function that needs to be minimized. The scaling factor for the loss function is 100 (default), and r is the Pearson correlation coefficient. The mean absolute error (MAE) is used as the loss function for the absorption coefficients at 370 nm and 880 nm of total BC concentration, and its calculation is as follows:

[0053] Where n is the number of samples. This represents the true value of the i-th sample. Let be the predicted value for the i-th sample.

[0054] Therefore, the loss function of the total mass concentration of BC as follows:

[0055] In the formula, where For loss function The scaling value, For the Pearson correlation coefficient, we have:

[0056] In the formula, where and They are respectively a source number The mass concentration of each emission source and the predicted mass concentration of BC emitted from that emission source. and These are the average mass concentrations of each emission source and the predicted average mass concentration of BC emitted by that emission source, respectively, where n is the number of data samples used to calculate the gradient and update the model parameters in each iteration.

[0057] At this point, we begin to build the model, establishing an input layer based on the number of features. The input layer is connected to a fully connected hidden layer, and a dropout layer is added to each hidden layer. We set the dropout rate, activation function, and initialize the weights in the neural network layers for the dropout layer.

[0058] A parallel output layer is established, outputting the AAE layer (3 neurons) controlled by each emission source, the Mass layer (3 neurons) for coal combustion emissions BC, biomass emissions BC, and motor vehicle emissions BC, and the MAE layer (3 neurons) at the 880nm wavelength, connecting to the hidden layer. The activation functions of the AAE layer, Mass layer, and MAE layer ensure that the output of each neuron is non-negative. The activation function of the AAE layer is as described above.

[0059] The absorption coefficient layer and the total mass concentration layer of BC in the 70nm and 880nm bands each have 1 neuron, representing the absorption coefficient of BC in the 370nm and 880nm bands and the total mass concentration of BC, respectively, and are connected to the output layer.

[0060] Loss functions are selected for the absorption coefficients and total BC mass concentrations at 370nm and 880nm wavelengths, as well as the model-predicted BC mass concentrations from the three sources. The custom loss functions from step 6 are used for the three source BC mass concentrations. Initial weights are then assigned to these loss functions. The custom loss functions, i.e., the loss functions for the BC mass concentrations emitted from each emission source, are as follows:

[0061]

[0062] + +

[0063] In the formula, , as well as These are all loss functions of the model, constraining the absorption coefficient at 880nm, the absorption coefficient at 370nm, and the simulated total mass of BC, respectively. , The absorption coefficients are for the 880nm and 370nm wavelength bands, respectively. , , The mass absorption cross section of BC for coal combustion emissions, the mass absorption cross section of BC for biomass combustion emissions, and the mass absorption cross section of BC for motor vehicle emissions are given. , , These are the mass emissions of BC from coal combustion, the mass emissions of BC from biomass combustion, and the mass emissions of BC from motor vehicles. , , These are the ranges for the angstrom absorption index of coal combustion, the angstrom absorption index of biomass combustion, and the angstrom absorption index of motor vehicles, respectively. The total mass concentration of BC.

[0064] During the training of the neural network model, an adaptive learning rate optimization algorithm is used to set the learning rate, and a callback function is used to dynamically adjust the learning rate during the training process.

[0065] In other words, this invention optimizes the neural network by: (a) adjusting the learning rate: the model uses an adaptive learning rate optimization algorithm (Adam), setting the learning rate, and dynamically adjusting the learning rate during training using a callback function. (b) adjusting the training batch size: trying different batch sizes from small to large to train the model. (c) adjusting the training epochs: trying different epochs from small to large to find a suitable epoch size. To avoid overfitting, the training epochs are supervised using an early stopping function. (d) verifying the accuracy of the model's analytical results: verifying the accuracy of the model's analytical results based on external validation. The reliability and accuracy of the model can be seen from the loss curves of the training and validation sets to determine whether the model is overfitting or underfitting. Simultaneously, the model's performance can be verified by comparing the simulated and actual values ​​of the absorbance coefficient and BC concentration in the 370nm and 880nm wavelength bands, considering correlation and error.

[0066] In this invention, after the model is trained, the absorption coefficient, total BC mass concentration, and BC mass concentration of coal combustion emissions, biomass combustion, and motor vehicle emissions are output in the 370nm and 880nm bands.

[0067] The feasibility of the method of the present invention will be demonstrated below through a specific practice.

[0068] (1) The environmental data used in this example are the concentration data of organic aerosols, NO, NOx, CO, SO2, and O3, with an hourly resolution. The organic aerosol concentration data comes from the analysis data of the Aerosol Chemical Components Online Monitor (ACSM), and the gaseous pollutant data comes from the site data of the national control station. The BC concentration data, 370nm and 880nm absorbance data come from the black carbon analyzer data (model AE33, Magee Scientific, Berkeley, CA, USA), with an hourly resolution. The coal combustion emission source indicator is its emitted organic matter concentration (CCOA), the biomass combustion emitted organic matter concentration (BBOA), and the motor vehicle emitted organic matter concentration (HOA), which come from the analysis data of ACSM.

[0069] (2) Using the particulate matter absorbance data measured by AE33, the ratio of the primary absorbance coefficient to the BC mass concentration was obtained by the least squares (MRS) method. This ratio was used to correct the measured particulate matter absorbance coefficient to remove the influence of secondary brown carbon, thus obtaining the true absorbance coefficient of BC. Then, a high-dimensional dataset was constructed using the true absorbance coefficient of BC and the data from (1), and aligned with time.

[0070] (3) For high-dimensional datasets, data prediction is performed. In this example, since there are few missing values, if there are two consecutive missing time steps, the missing samples are deleted. If there is a single missing time step, the average of the previous and next time steps is used for interpolation to fill the missing data. At the same time, outlier handling is performed on all features using the 3x standard deviation (3σ) principle, and data exceeding 3σ are removed. A second check for missing data is performed to ensure data integrity.

[0071] (4) After confirming that the dataset is complete and has no missing data, the min-max normalization method is used to scale the feature values ​​of all features to scale the range of these features to [0, 1], removing the influence of units and orders of magnitude. The input features in this example are a total of 7, including environmental data (aerosol concentration, NO, NOx, CO, SO2, O3) obtained in environment (1) and time (h).

[0072] (5) Ranges were set for the AAE values ​​of the three sources BC based on literature review and prior knowledge. In this example, the AAE value range for coal combustion emissions BC is 1.2–3, the AAE value range for biomass combustion emissions BC is 1.2–2.5, and the AAE value for motor vehicle emissions BC is 0.8–1.1. Upper and lower limits were established based on the AAE values ​​of the three sources. , , The activation function output by the three units has the following output value: .

[0073] (6) For the indicator substances selected for each emission source in the example, including: coal combustion emission source indicator is its coal-fired organic matter concentration (CCOA), biomass combustion emission organic matter concentration (BBOA), and motor vehicle emission organic matter concentration (HOA), the model will automatically standardize each source indicator through a custom loss function, converting the feature values ​​into a distribution with a mean of 0 and a standard deviation of 1. The definition of the custom activation function in the model is based on the Pearson correlation calculation formula and is completed through the function. Among them, BBOA, CCOA, and HOA are respectively input as y_true, and the model predicts , , The input will be y_pred. Each bach_size is calculated, and y_pred and y_true are standardized accordingly. Finally, the result is obtained... The formula calculates the Pearson correlation coefficient between y_pred and y_true.

[0074] (7) In this case, an input layer with 7 neurons (a) was established, which is connected to 6 fully connected hidden layers. The neuron data of the 6 hidden layers are (64, 32, 16, 16, 16, 8). A dropout layer is added to each hidden layer, and the dropout rate of the dropout layer is 0.2. Except for the last layer, the activation function of the other fully connected layers is tanh, and the activation function of the last layer is softplus (b). The weights in the neural network layers are initialized to he_uniform, and samples are drawn from a uniform distribution for weight initialization. (c) A parallel output layer with 9 neurons of 1 each is established to output the coal combustion emission BC, biomass emission BC, motor vehicle emission BC, and the mass absorption cross section of the coal combustion emission BC, biomass emission BC, and motor vehicle emission BC at 370nm and 880nm, respectively. , , Connect (b). The activation function of the BC mass absorption cross section and BC mass concentration output layer at 880nm is softplus, ensuring that the output of each neuron is non-negative. , , The activation function is the one defined in step 5. (d) Establish three custom layers, each with 1 neuron, representing the absorption coefficient of BC at 370nm and 880nm wavelengths and the total mass concentration of BC, respectively, and connect them to (c). Through custom... The model's predicted emissions BC (coal combustion emissions BC), BC (biomass emissions BC), and BC (motor vehicle emissions BC) are optimized. In this example, the scale in the loss function is set to 100. Other outputs are... The loss functions were optimized. Then, initial weights were assigned to these loss functions. In this study, the initial weights were set as follows: the weights for the absorption coefficients at 370nm and 880nm were 1 and 1, respectively; the total mass concentration of BC was 5; and the weights for the mass concentrations of the three source BCs were 5, 5, and 1, respectively.

[0075] (8) Adjustment of learning rate: In this case, the adaptive learning rate optimization algorithm (Adam) was used, with learning rates set to 0.01, 0.005, 0.001, 0.0005, and 0.0003. A callback function was used to dynamically adjust the learning rate during training. Multiple experiments showed that the initial learning rate of 0.001 yielded the best results in this invention, and subsequent adjustments to the learning rate were made automatically using the callback function. The monitored metric was the loss function of the validation set. If there was no significant decrease within 10 rounds, the learning rate was adjusted, with the new learning rate being 0.5 times the previous one, and the minimum learning rate being 1e-6. Adjustment of training batch size: Batch sizes were set to 8, 16, 32, 64, and 128, respectively, and the model was trained accordingly. It was found that setting 16 performed best. Adjustment of training epochs: An early stopping mechanism was used to determine the training epochs by observing changes in the validation set loss function, with a tolerance of 50 steps. To avoid model overfitting, the training epochs were supervised using an early stopping function. After multiple experiments, it was found that around 250 epochs performed best. Verification of the accuracy of the model analysis results: Verify the accuracy of the model analysis results based on external validation.

[0076] (9) Output the mass concentrations of BC from the three sources. The absorption coefficients at 370 nm and 880 nm, the total BC concentration, and the R² of their observed values ​​were compared to analyze the resolution results. The results were R² = 0.91 (370 nm), 0.85 (880 nm), and 0.8 (total BC mass concentration), respectively. This indicates that the model can effectively resolve the BC mass concentrations from vehicle emissions, biomass emissions, and coal combustion emissions.

[0077] like Figure 3 As shown, the horizontal axis of a, b, and c represents the simulated values, and the vertical axis represents the observed values. It can be seen that the R² values ​​for the absorbance coefficients at 370nm and 880nm and the mass concentration of BC obtained from both the model prediction and actual observation are high (both reaching 0.8 or above). Therefore, it can be concluded that the model can accurately reflect the absorbance coefficient and the mass concentration of BC. The values ​​are 21.6 Mm⁻¹, 4.3 Mm⁻¹, and 1 μgm⁻³, respectively.

[0078] like Figure 4 As shown, the model analysis process is as follows: The blue curve (loss curve for training data) and the green curve (loss curve for validation data) demonstrate that the model exhibits neither overfitting nor underfitting. The convergence and overlap of both curves indicate that the model's loss curve has stabilized and decreased synchronously, suggesting that the model has neither fallen into underfitting (high loss on the training set) nor overfitting (significantly higher loss on the validation set than the training set) due to excessive complexity. This demonstrates the model's strong generalization ability and robustness.

[0079] This invention effectively eliminates the disturbance to the BC absorption coefficient caused by secondary brown carbon formation, making the analysis results more accurate. 2. It fully considers the influence of atmospheric environmental factors on the BC absorption capacity and mass concentration, eliminating the need to rely on emission source spectra or repeated analysis to verify the BC source contribution. Source spectra are affected by combustion conditions, fuel type, etc., and lack universality. Traditional AE models (relying on optical data and physical formulas) can only resolve two sources. China's emission sources are complex; generally, BC comes from coal combustion, vehicle emissions, and biomass combustion, exceeding two emission sources. Therefore, AE models have limitations. Embedding physical processes ensures that the analysis results satisfy physical laws and are not merely calculations based on statistical methods.

[0080] In this invention, after deducting secondary brown carbon, it can also be obtained by collecting filter membrane samples and analyzing and calculating using a 2015 OC / EC analyzer.

[0081] Based on the same inventive concept, this invention also provides an atmospheric black carbon source apportionment system that integrates multi-dimensional factors and physical constraints, such as... Figure 5 As shown, it includes: The data acquisition module acquires the ratio of the total mass concentration of BC to the primary absorption coefficient, and calculates the primary BC absorption coefficient based on the ratio of the total mass concentration of BC. The data processing module preprocesses the first BC absorption coefficient; The model calculation module inputs the preprocessed BC absorption coefficient and the acquired environmental data into a pre-trained neural network model to obtain the absorption coefficient, total BC mass concentration, and BC mass concentration of each emission source under different wavelengths output by the neural network model. The neural network model defines activation and loss functions during construction. Then, it takes the BC absorption coefficient and environmental data as inputs and the absorption coefficient, total BC mass concentration, and BC mass concentration emitted from each emission source at different wavelengths as outputs to complete the training.

[0082] The above description is merely a preferred embodiment of the present invention and does not constitute any limitation on the present invention. Any simple modifications, alterations, or equivalent structural changes made to the above embodiments based on the technical essence of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for apportioning the sources of atmospheric black carbon by integrating multidimensional factors and physical constraints, characterized in that, include: Step 1: Obtain the ratio of the total mass concentration of BC to the primary absorption coefficient, and calculate the primary BC absorption coefficient based on the total mass concentration of BC and the ratio; Step 2: Preprocess the first BC absorption coefficient; Step 3: Input the preprocessed BC absorption coefficient and the acquired environmental data into the pre-trained neural network model to obtain the absorption coefficient, total BC mass concentration and BC mass concentration of each emission source under different bands output by the neural network model. The neural network model defines activation and loss functions during construction. Then, it takes the BC absorption coefficient and environmental data as inputs and the absorption coefficient, total BC mass concentration, and BC mass concentration emitted from each emission source at different wavelengths as outputs to complete the training.

2. The method for analyzing the sources of atmospheric black carbon according to claim 1, which integrates multi-dimensional factors and physical constraints, is characterized in that... The construction of the neural network model includes: An input layer is established based on the input, and the input layer is connected to a fully connected hidden layer. A dropout layer is added to each hidden layer, and the dropout rate, activation function and weights in the neural network layer are set. A parallel output layer is established to output the AAE layer controlled by each emission source, the Mass layer of BC emitted by each emission source, and the MAE layer in the 880nm band, respectively. At the same time, the activation function of the AAE layer is determined according to the different absorption angstrom index of each emission source of BC. The output layer is then connected to the hidden layer. A light absorption coefficient layer and a total mass concentration (BC) layer were constructed in the 370nm and 880nm wavelength bands, and then connected to the output layer. Establish loss functions for the total mass concentration of BC and the mass concentration of BC emitted from each emission source, respectively. Then, the neural network is trained using the BC absorption coefficient and environmental data as inputs, and the absorption coefficients, total BC mass concentrations, and BC mass concentrations emitted from various emission sources at different wavelengths as outputs. During training, the neural network model is corrected based on the loss value of the loss function until training is complete. Output the trained neural network model.

3. The method for analyzing the sources of atmospheric black carbon according to claim 2, which integrates multi-dimensional factors and physical constraints, is characterized in that... The expression for the activation function of the AAE layer is as follows: scaled_output = min_val + (max_val - min_val) ×sigmoid(x) In the formula, scaled_output is the activation function of the AAE layer, min_val is the calculated value of the lower limit of the emission source absorption angstrom index, max_val is the calculated value of the upper limit of the emission source absorption angstrom index, and sigmoid(x) is the sigmoid activation function.

4. The method for analyzing the sources of atmospheric black carbon according to claim 2, which integrates multi-dimensional factors and physical constraints, is characterized in that... The loss function of the total mass concentration of BC as follows: In the formula, where For loss function The scaling value, For the Pearson correlation coefficient, we have: In the formula, where and They are respectively a source number The mass concentration of each emission source and the predicted mass concentration of BC emitted from that emission source. and These are the average mass concentrations of each emission source and the predicted average mass concentration of BC emitted by that emission source, respectively, where n is the number of data samples used to calculate the gradient and update the model parameters in each iteration.

5. The method for analyzing the sources of atmospheric black carbon according to claim 2, which integrates multi-dimensional factors and physical constraints, is characterized in that... The emission sources include coal combustion emissions, biomass combustion emissions, and motor vehicle emissions.

6. The method for apportioning atmospheric black carbon sources by integrating multi-dimensional factors and physical constraints according to claim 5, characterized in that, The loss function for the mass concentration of BC emitted from each emission source is as follows: + + In the formula, , as well as Both are loss functions. , The absorption coefficients are for the 880nm and 370nm wavelength bands, respectively. , , The mass absorption cross section of BC for coal combustion emissions, the mass absorption cross section of BC for biomass combustion emissions, and the mass absorption cross section of BC for motor vehicle emissions are given. , , These are the mass emissions of BC from coal combustion, the mass emissions of BC from biomass combustion, and the mass emissions of BC from motor vehicles. , , These are the ranges for the angstrom absorption index of coal combustion, the angstrom absorption index of biomass combustion, and the angstrom absorption index of motor vehicles, respectively. The total mass concentration of BC.

7. The method for analyzing the sources of atmospheric black carbon according to claim 1, which integrates multi-dimensional factors and physical constraints, is characterized in that... During the training of the neural network model, an adaptive learning rate optimization algorithm is used to set the learning rate, and a callback function is used to dynamically adjust the learning rate during the training process.

8. The method for analyzing the sources of atmospheric black carbon according to claim 1, characterized in that, In step 2, the first BC absorption coefficient is preprocessed, including handling missing values ​​and handling outliers.

9. A method for analyzing the sources of atmospheric black carbon that integrates multidimensional factors and physical constraints according to claim 8, characterized in that, The processing of missing values ​​includes: Missing data is filled in by means of the mean, median, interpolation, or by removing missing values. The processing of outliers includes: Outliers are identified and processed using statistical methods.

10. A system for apportioning the sources of atmospheric black carbon that integrates multidimensional factors and physical constraints, characterized in that, include: The data acquisition module acquires the ratio of the total BC mass concentration to the primary absorption coefficient, and calculates the primary BC absorption coefficient based on the ratio of the total BC mass concentration. The data processing module preprocesses the first BC absorption coefficient; The model calculation module inputs the preprocessed BC absorption coefficient and the acquired environmental data into the pre-trained neural network model to obtain the absorption coefficient, total BC mass concentration and BC mass concentration of each emission source under different wavelengths output by the neural network model. The neural network model defines activation and loss functions during construction. Then, it takes the BC absorption coefficient and environmental data as inputs and the absorption coefficient, total BC mass concentration, and BC mass concentration emitted from each emission source at different wavelengths as outputs to complete the training.

Images