PM2.5 chemical component concentration vertical profile inversion model method and system based on aerosol laser radar

Abstract

The invention discloses a PM2.5 chemical component concentration vertical profile inversion model method and system based on an aerosol laser radar, and the method comprises the following operation steps: collecting 532nm aerosol data as an input feature through the aerosol laser radar, carrying out the preprocessing of the aerosol data, inputting the preprocessed aerosol data into a deep learning model, and obtaining a PM2.5 chemical component concentration vertical profile inversion model; the PM2.5 chemical component characteristics are output; performing convolutional nonlinear fitting on the aerosol data through a neural convolutional network, and performing optimization of an attention mechanism and a long-short-term memory neural network on the aerosol data after nonlinear fitting to obtain a deep learning model; and screening an optimal solution by the deep learning model through a Bayesian optimization algorithm and a fast non-dominated sorting genetic algorithm, and normalizing the deep learning model to obtain a PM2.5 chemical component concentration vertical profile inversion model.

Description

Technical Field

[0001] The present invention relates to the field of PM2.5 identification, and in particular to a method and system for an aerosol laser radar-based PM2.5 chemical component concentration vertical profile inversion model. Background Art

[0002] Fine particulate matter (PM2.5) is one of the atmospheric pollutants that are controlled. It is a mixture of various chemical components, such as sulfate (SO42-), nitrate (NO3-), ammonium salt (NH4+), organic matter (OM) and black carbon (BC). These chemical components directly affect ground PM2.5 pollution in atmospheric physical and chemical processes such as chemical transformation, advection transport, turbulent diffusion and mixed deposition at different altitudes in the atmospheric boundary layer. Therefore, obtaining the concentration of chemical components at different altitudes in the atmosphere is an important basis for further precise prevention and control of PM2.5 pollution.

[0003] Aerosol lidar is a vertical detection tool for atmospheric pollution with quasi-real-time, rapid response capabilities, wide vertical coverage, and long-term continuous detection. It can provide PM2.5-related optical parameters in different altitude layers, such as backscattering coefficient, extinction coefficient, and depolarization ratio. The research team identified a variety of aerosol components, such as inorganic salts, water-soluble organic matter, black carbon, sand and dust, and sea salt from the various optical parameters obtained by aerosol lidar by developing a reasonable vertical inversion algorithm for aerosol components. However, these algorithms rely on differences in the optical and microphysical properties of aerosol components, and therefore cannot distinguish chemical components with similar properties, such as sulfates and nitrates with similar light scattering, particle size, and shape. The lack of a vertical profile inversion algorithm for PM2.5 chemical component concentrations limits the application of aerosol lidar in the vertical detection of PM2.5 chemical components.

[0004] Therefore, the present invention proposes a vertical profile inversion model of PM2.5 chemical component concentration based on aerosol lidar, which solves the limitation that the optical parameters of aerosol lidar cannot separate chemical components, and provides technical support for the refined three-dimensional monitoring of PM2.5 pollution in my country. Summary of the invention

[0005] The purpose of the present invention is to provide a method and system for a PM2.5 chemical component concentration vertical profile inversion model based on aerosol lidar, which solves the above-mentioned technical problems pointed out in the prior art.

[0006] The present invention provides a method for inverting a PM2.5 chemical component concentration vertical profile model based on an aerosol laser radar, comprising the following steps: Using an aerosol lidar to collect 532nm aerosol data as input features, preprocessing the aerosol data, inputting the preprocessed aerosol data into a deep learning model, and outputting PM2.5 chemical component features; Performing nonlinear convolution fitting on the aerosol data through a neural convolution network, optimizing the attention mechanism and long short-term memory neural network on the aerosol data after the nonlinear fitting, and obtaining a deep learning model; The deep learning model screens the optimal solution through a Bayesian optimization algorithm and a fast non-dominated sorting genetic algorithm, and normalizes the deep learning model to obtain a vertical profile inversion model of PM2.5 chemical component concentration.

[0007] Preferably, the 532nm aerosol data includes: backscattering coefficient characteristics, extinction coefficient characteristics and depolarization ratio characteristics, as well as meteorological element characteristics; The PM2.5 chemical component characteristics include: sulfate, nitrate, ammonium salt, organic matter and black carbon.

[0008] Preferably, the aerosol data is convolved and nonlinearly fitted by a neural convolution network, and the attention mechanism and long short-term memory neural network are optimized for the aerosol data after nonlinear fitting to obtain a deep learning model. The specific operation steps are as follows: The pre-processed aerosol data is input into a deep learning model, and a convolutional neural network is used to perform convolution calculation and nonlinear fitting on the aerosol data; Calculating the weight differences between the features in the aerosol data in the vertical profile inversion task of the PM2.5 chemical components through an attention mechanism; dynamically adjusting the weights of the features in the aerosol data according to the weight differences between the features to optimize the aerosol data; The time of aerosol data collection is bidirectionally transmitted in forward and backward directions through a long short-term memory neural network to capture the time series characteristics of aerosol data.

[0009] Preferably, the deep learning model screens the optimal solution through a Bayesian optimization algorithm and a fast non-dominated sorting genetic algorithm, and normalizes the deep learning model to obtain a PM2.5 chemical component concentration vertical profile inversion model. The specific operation steps are as follows: The deep learning model is coupled online by a Bayesian optimization algorithm, and an objective function is constructed for the aerosol data by a Bayesian optimization algorithm to optimize the optimal solution of the aerosol data; The optimal solution obtained by the deep learning model is normalized by a fast non-dominated sorting genetic algorithm to obtain a vertical profile inversion model of PM2.5 chemical component concentration.

[0010] Preferably, the optimal solution obtained by the deep learning model is normalized by a fast non-dominated sorting genetic algorithm to obtain a PM2.5 chemical component concentration vertical profile inversion model. The specific operation steps are as follows: Constructing multiple objective functions based on multiple statistical indicators of the PM2.5 chemical components, and respectively obtaining the optimal solutions corresponding to the multiple objective functions to obtain the Pareto optimal solution as the Pareto frontier; Setting a crowding distance for each Pareto optimal solution in the Pareto front; All Pareto optimal solutions in the corresponding Pareto front of each objective function are sorted from small to large according to the minimum value of the objective function; Assigning the maximum crowding distance to the optimal solution of the minimum value and the optimal solution of the maximum value in the order of the minimum value of the objective function from small to large; The sum of the calculated crowding distances of adjacent optimal solutions is calculated for the intermediate solutions in the order of the minimum value of the objective function from small to large; the sum of the calculated crowding distances of the adjacent optimal solutions is normalized to obtain the optimal solution of the deep learning model, and the vertical profile inversion model of PM2.5 chemical component concentration is obtained.

[0011] Preferably, the optimal solution obtained by the deep learning model is optimized by a fast non-dominated sorting genetic algorithm to obtain a PM2.5 chemical component concentration vertical profile inversion model, which also includes: Performing fast non-dominated sorting on the Pareto optimal solutions in the Pareto front, setting the Pareto optimal solutions to levels, and obtaining a first Pareto optimal solution level, a second Pareto optimal solution level, ... an nth Pareto optimal solution level; Allocating a fitness value to the Pareto optimal solution according to the level of the Pareto optimal solution; The Pareto frontier is adjusted according to the fitness value of the Pareto optimal solution. Generate The Pareto frontier of the offspring of ; Through the Pareto frontier of the parent and the Pareto frontier of the offspring Composition The Pareto front set ; The Pareto front set Perform fast non-dominated sorting again and set the Pareto front set The Pareto frontier in the set level, get the first level Pareto frontier level, the second level Pareto frontier level... the nth level Pareto frontier level; Determine the Pareto frontier The number of Pareto optimal solutions in the Pareto frontier level in the next generation is constructed ; The Pareto frontier Execute the above steps to generate a new recursive frontier set Find the Pareto optimal solution.

[0012] Preferably, the allocation of the fitness value is correlated with the level of the Pareto optimal solution.

[0013] Determine the Pareto front set The number of Pareto optimal solutions in the Pareto frontier level in the next generation is constructed , the specific steps are as follows: Determine the Pareto front set Are all the Pareto optimal solutions of the first level Pareto frontier equal to ; If so, then the Pareto front set All Pareto optimal solutions in the first level Pareto frontier hierarchy build the next generation of Pareto frontier ; If not, then the Pareto front set All Pareto optimal solutions in the first level Pareto frontier do not satisfy ; Combine all the Pareto optimal solutions in the second level Pareto frontier layer with all the Pareto optimal solutions in the first level Pareto frontier layer to continue to determine whether they are equal ; If so, then the Pareto front set All Pareto optimal solutions in the first-level Pareto frontier layer and all Pareto optimal solutions in the second-level Pareto frontier layer construct the next generation of Pareto frontier ; If not, continue to follow the Pareto front set Fast non-dominated sorting to construct the next generation of Pareto frontier , until the conditions are met.

[0014] Accordingly, the present application also provides a system for an inversion model of a PM2.5 chemical component concentration vertical profile based on an aerosol lidar, comprising: a preprocessing module; a deep learning training module; a normalization optimization module; The preprocessing module is used to collect 532nm aerosol data using aerosol lidar as input features, preprocess the aerosol data, input the preprocessed aerosol data into the deep learning model, and output PM2.5 chemical component features; The deep learning training module is used to perform nonlinear convolution fitting on the aerosol data through a neural convolution network, and optimize the attention mechanism and long short-term memory neural network on the aerosol data after nonlinear fitting to obtain a deep learning model; The normalization optimization module is used to screen the optimal solution of the deep learning model through the Bayesian optimization algorithm and the fast non-dominated sorting genetic algorithm, and normalize the deep learning model to obtain the vertical profile inversion model of PM2.5 chemical component concentration.

[0015] Compared with the prior art, the embodiments of the present invention have at least the following technical advantages: By analyzing the method and system for the vertical profile inversion model of PM2.5 chemical component concentration based on aerosol lidar provided by the present invention, it can be known that the input and output data types used by the deep learning model in specific applications include aerosol data with a wavelength of 532nm obtained from the aerosol lidar, and the aerosol data is preprocessed to remove noise and outliers, so that the PM2.5 chemical component characteristics output when the aerosol data is used as an input feature are more accurate; Furthermore, the preprocessed aerosol data is input into the deep learning model. CNN gradually conducts deep learning on the data set through local perception, sparse connection, and weight and bias sharing, thereby reducing the complexity of the neural network and preventing overfitting. The features in the aerosol data are weighted according to their importance through the attention mechanism, thereby optimizing the weight of the aerosol data and reducing the interference of irrelevant or redundant information. The sequence data is processed from two directions (forward and backward) simultaneously through the long short-term memory neural network. The long short-term memory neural network actually contains two layers (forward and backward) of LSTM: one processes the input sequence from forward to backward, and the other, on the contrary, processes from backward to forward, to capture the time series characteristics of the aerosol data, thereby capturing the time series characteristics of the aerosol data. Furthermore, the deep learning model uses the Bayesian optimization algorithm to construct an objective function for the aerosol data, with the aim of finding a hyperparameter combination (i.e., hyperparameters) that minimizes the objective function, thereby finding the best hyperparameters. By integrating multiple statistical indicators into the objective function, the optimization problem is transformed into a multi-objective optimization problem, aiming to simultaneously minimize the minimum values ​​of all objective functions and find the possibility of the best solution that meets specific needs as the Pareto front. Each Pareto optimal solution in the Pareto front is set with a crowding distance to measure the density of a solution in its Pareto front. All solutions in the current Pareto front are sorted from small to large according to the value of the objective function, because these optimal solutions represent extreme cases in the solution space, and increasing their selection probability helps explore more solution spaces. In addition to the minimum value of the objective function, the optimal solution with the minimum value and the optimal solution with the maximum value in the sorting from small to large are calculated with their crowding distances, and the sum of the adjacent optimal solutions before and after each intermediate solution in the sorting of its objective function value is found and normalized to obtain the final PM2.5 chemical component concentration vertical profile inversion model. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0017] Figure 1 A flowchart of a method for inverting a PM2.5 chemical component concentration vertical profile model based on an aerosol lidar provided in Example 1 of the present invention; Figure 2 A schematic flow chart of a method for inverting a PM2.5 chemical component concentration vertical profile model based on an aerosol lidar provided in Example 1 of the present invention; Figure 3 A deep learning model training flow chart of a method for inverting a PM2.5 chemical component concentration vertical profile based on an aerosol lidar provided in Example 1 of the present invention; Figure 4 A deep learning model normalization flow chart of a method for inverting a PM2.5 chemical component concentration vertical profile based on an aerosol lidar provided in Example 1 of the present invention; Figure 5 A first normalized flow chart of a fast non-dominated sorting genetic algorithm of a deep learning model of a method for inverting a PM2.5 chemical component concentration vertical profile based on an aerosol lidar provided in Example 1 of the present invention; Figure 6A second normalized flow chart of a fast non-dominated sorting genetic algorithm of a deep learning model of a method for inverting a vertical profile of PM2.5 chemical component concentration based on an aerosol lidar provided in Example 1 of the present invention; Figure 7 A flow chart of a system for inversion model of PM2.5 chemical component concentration vertical profile based on aerosol lidar provided in Example 2 of the present invention.

[0018] Marking: preprocessing module 10; deep learning training module 20; normalization optimization module 30. DETAILED DESCRIPTION

[0019] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0020] The present invention is further described in detail below through specific embodiments in conjunction with the accompanying drawings.

[0021] Embodiment 1 like Figure 1 As shown, Figure 2 As shown, the present invention proposes a method for inverting a PM2.5 chemical component concentration vertical profile model based on an aerosol lidar, comprising the following steps: S1: using aerosol lidar to collect 532nm aerosol data as input features, preprocessing the aerosol data, inputting the preprocessed aerosol data into a deep learning model, and outputting PM2.5 chemical component features; The 532nm aerosol data include: backscattering coefficient (σbsc, 532) characteristics, extinction coefficient characteristics (σext, 532) and depolarization ratio characteristics (σdep, 532), as well as meteorological element characteristics; The PM2.5 chemical composition characteristics include: sulfate (SO42-), nitrate (NO3-), ammonium salt (NH4+), organic matter (OM) and black carbon (BC); It should be noted that the input and output data types used by the deep learning model include aerosol data with a wavelength of 532nm obtained from aerosol lidar (i.e., including the scattering coefficient (σbsc, 532), extinction coefficient (σext, 532) and depolarization ratio (σdep, 532) of ground data (i.e., data collected on or near the surface), and vertical layer data (i.e., data collected from the surface to different altitudes) from ERA5 global reanalysis data) of various meteorological elements (meridional wind, zonal wind, temperature, relative humidity, specific humidity, vertical velocity and potential height)), and the input data are preprocessed; Aerosol data preprocessing includes: using Hampel filter to identify outliers; in a 12-hour time window, if the difference between a data point and the median value in the window is greater than three times the mean absolute deviation, the data point is considered an outlier; for the identified outliers and the existing missing data, linear interpolation is used to fill them; this means that according to the known data points before and after the missing value or outlier, a reasonable value of the missing or outlier position is estimated through a simple linear relationship to maintain the integrity and continuity of the data set; it is expressed by the linear interpolation formula: ;in, are the abnormal data points or missing data points that are filtered out. and for Two known data points before and after, namely ; Calculate outliers through Hampel filter, and further calculate interpolation to achieve the purpose of aerosol data preprocessing; The preprocessed aerosol data were input into the deep learning model for analysis, and the output type was PM2.5 chemical composition characteristics, which consisted of five different types of PM2.5 particle components (i.e., sulfate (SO42-), nitrate (NO3-), ammonium salt (NH4+), organic matter (OM), and black carbon (BC)); S2: performing nonlinear convolution fitting on the aerosol data through a neural convolution network, performing attention mechanism and long short-term memory neural network optimization on the aerosol data after nonlinear fitting, and obtaining a deep learning model; It should be noted that by combining CNN, ATT and BiLSTM, a deep learning model that can efficiently process aerosol data is constructed; CNN extracts local features, ATT enhances key features, and BiLSTM captures long-term dependencies in time series. The synergistic effect of the three significantly improves the performance of the model; aerosol data has the characteristics of high dimensionality, nonlinearity and time correlation, and a single model is difficult to cope with, so multi-module collaboration is required to cope with complex tasks.

[0022] S3: The deep learning model screens the optimal solution through the Bayesian optimization algorithm and the fast non-dominated sorting genetic algorithm, and normalizes the deep learning model to obtain the vertical profile inversion model of PM2.5 chemical component concentration.

[0023] It should be noted that the model was optimized for hyperparameters and multi-objective screening through Bayesian optimization and NSGA-II, and the stability of the model was improved through normalization, and finally the PM2.5 chemical component concentration vertical profile inversion model was obtained; Bayesian optimization improved the efficiency of hyperparameter selection, NSGA-II realized multi-objective optimization, and normalization improved the model training effect. Finally, the model was able to accurately invert the vertical distribution of PM2.5 chemical component concentrations; hyperparameter optimization and multi-objective screening are key steps to improve model performance, and normalization is the basic operation of deep learning model training. These steps together ensure the efficiency and accuracy of the model.

[0024] Specifically, Figure 3 As shown, in step S2, the aerosol data is convolved and nonlinearly fitted by a neural convolutional network, and the aerosol data after nonlinear fitting is optimized by an attention mechanism and a long short-term memory neural network to obtain a deep learning model. The specific operation steps are as follows: S21: the pre-processed aerosol data is input into a deep learning model, and a convolution neural network is used to perform convolution calculation and nonlinear fitting on the aerosol data; It should be noted that convolutional neural network (CNN) is one of the variants of multi-layer perceptron, which can effectively establish the mapping relationship between multi-dimensional input and output; unlike other neural networks, CNN gradually conducts deep learning of data sets through local perception, sparse connection, and weight and bias sharing, thereby reducing the complexity of neural networks and preventing overfitting (that is, CNN uses local perception, sparse connection and weight sharing to reduce the number of parameters and model complexity while maintaining efficient learning ability, which is suitable for processing high-dimensional data); CNN has good computational efficiency and robustness, and can deeply mine the local features of multi-dimensional data sets to perform tasks such as regression or classification. In this study, CNN is used to process time series data of different variables; The convolutional layer of the convolutional neural network (CNN) is responsible for detecting features in the input data, while the ReLU activation function increases the nonlinearity of the model, helping the model to better fit the data. The pooling layer further simplifies the representation and reduces the model parameters. The pooling layer is responsible for nonlinear downsampling. The pooling layer reduces the input size by taking the maximum value or average value, reducing the amount of calculation and the number of parameters, while trying to retain key features to prevent overfitting. The calculation formula is: ;in, and are the output and input of the convolutional layer, respectively. is the value after convolution calculation, and They are the weight matrix and bias term of the convolution kernel respectively; S22: calculating the weight differences between the features in the aerosol data in the vertical profile inversion task of the PM2.5 chemical components through an attention mechanism; dynamically adjusting the weights of the features in the aerosol data according to the weight differences between the features to optimize the aerosol data; It should be noted that in the task of analyzing the vertical distribution of PM2.5 chemical components, different meteorological parameters and aerosol data features (i.e., backscattering coefficient (σbsc, 532) features, extinction coefficient features (σext, 532) and depolarization ratio features (σdep, 532), as well as meteorological element features) have different degrees of influence on the results; the attention mechanism can automatically and flexibly determine which features are most important for prediction and give these features higher weights accordingly; The importance weight of each feature is generated through the fully connected layer of the convolutional neural network (CNN). According to the difference in the importance weight of each feature, the task requirements dynamically emphasize the weight of some features and weaken the weight of other features. The purpose is to improve the quality of the final feature representation and make it more focused on the information that is critical to the task, so as to dynamically adjust the weight of the features and optimize the features in the aerosol data; the attention mechanism (ATT) is used to help the model dynamically assign weights to different features, so that the model can pay more attention to the features that are more important to the current task, while reducing the interference of irrelevant or redundant information, which is particularly important when processing time series data with multiple variables and high heterogeneity, thereby optimizing aerosol data; S23: bidirectionally transmitting the time of aerosol data collection in a forward and backward direction through a long short-term memory neural network to capture the time series characteristics of the aerosol data; sequentially training the input aerosol data through the time series characteristics to obtain a deep learning model; It should be noted that the long short-term memory neural network (LSTM) is a variant of the recurrent neural network (RNNs). BiLSTM processes sequence data from two directions (forward and backward) at the same time. BiLSTM actually contains two layers (forward and backward) LSTM: one processes the input sequence from forward to backward, and the other does the opposite, processing from backward to forward. The purpose of this is to combine information from two perspectives to enhance the performance of the model; LSTM is mainly composed of cell state, forget gate, input gate, output gate and activation function; cell state is the core of LSTM, responsible for transmitting information throughout the sequence; forget gate determines which information should be discarded; input gate determines which new information should be stored; output gate controls the content of the final output; activation function helps regulate these processes; cell state acts as a path for information to flow through the entire network; forget gate determines which information should be forgotten based on the current input and the hidden state of the previous moment; input gate determines which new information should be added to the cell state; output gate generates the output of the current moment based on the cell state; the formula for bidirectional transmission in the forward and backward directions is: ; ; ; ; ; ; ; in, , , , , and They are the forget gate output, input gate output, candidate cell state, cell state, cell state at time t-1, and output gate output at time t; and are the output values ​​at time t and t-1 respectively; and are the weight matrices of the forget gate, input gate, cell state, and output gate at time t-1; is the input value at time t; and are the input weight matrices of the forget gate, input gate, cell state, and output gate at time t-1 respectively; and are the bias items of the forget gate, input gate, cell state and output gate at time t-1 respectively; is the final output value of BiLSTM at time t, which is the forward output value and backward output The combination of LSTM can effectively manage long-term dependencies by using mechanisms such as forget gates, input gates, and output gates, and BiLSTM further enhances this by processing sequence data bidirectionally to capture the time series characteristics of aerosol data; the input aerosol data is trained sequentially using the time series characteristics to obtain a deep learning model; Specifically, Figure 4 As shown, in step S3, the deep learning model screens the optimal solution through the Bayesian optimization algorithm and the fast non-dominated sorting genetic algorithm, and normalizes the deep learning model to obtain the PM2.5 chemical component concentration vertical profile inversion model. The specific operation steps are as follows: S31: online coupling the deep learning model through a Bayesian optimization algorithm, and constructing an objective function for the aerosol data through the Bayesian optimization algorithm to optimize the optimal solution of the aerosol data; It should be noted that the hyperparameters of the deep learning model CNN-ATT-BiLSTM determine the neural network structure, initialization process, training process, testing process, and convergence performance, which directly affect the inversion results of the normalized PM2.5 chemical component concentration vertical profiles; The Bayesian optimization algorithm is used to optimize the hyperparameters of the CNN-ATT-BiLSTM model. Bayesian optimization is an optimization method based on a probability model. It guides the search by constructing the posterior distribution of the objective function. It can find the global optimal solution in fewer iterations, reduce computational costs and improve model performance. Hyperparameters are regarded as decision variables, and the goal is to minimize the objective function (usually the loss function of the model), so the objective function is constructed. This process is converted into a mathematical optimization problem, the purpose of which is to find the hyperparameter combination that minimizes the objective function, so as to find the best hyperparameters. The calculation formula is: ; ; in, is a decision vector consisting of d hyperparameters, is the decision space, i.e. the optimization range of the hyperparameters, represents the input data of the deep learning model, is the optimal hyperparameter combination vector; is the objective function, is the actual observed value, that is, the objective function value of the known sampling point; is the likelihood distribution, is the prior probability distribution, that is, The estimate, is the marginal likelihood distribution, is the posterior probability distribution, i.e. The confidence coefficient of In this way, Bayesian optimization can effectively explore the hyperparameter space and find the optimal hyperparameter combination, thereby improving the performance of deep learning models (such as CNN-ATT-BiLSTM) on specific tasks, such as the inversion of vertical profiles of PM2.5 chemical component concentrations; this method not only reduces the computational resources required to evaluate the loss function, but also improves the accuracy and generalization ability of the model; S32: normalizing the optimal solution obtained by the deep learning model through a fast non-dominated sorting genetic algorithm, thereby obtaining an inversion model for the vertical profile of PM2.5 chemical component concentration; It should be noted that multiple statistical indicators (i.e., correlation coefficient, root mean square error, regression coefficient, etc. (i.e., correlation coefficient measures the consistency between the predicted value and the true value; root mean square error measures the magnitude of the prediction error)) are used to construct multiple objective functions for PM2.5 chemical components through a fast non-dominated sorting genetic algorithm. The optimal solution of the minimum and maximum values ​​of the objective function is assigned an infinite crowding distance. The optimal solution of the deep learning model is optimized by calculating the crowding distance, and the vertical profile inversion model of PM2.5 chemical component concentration is obtained. Specifically, Figure 5 As shown, in step S32, the optimal solution obtained by the deep learning model is normalized by a fast non-dominated sorting genetic algorithm to obtain a PM2.5 chemical component concentration vertical profile inversion model. The specific operation steps are as follows: S321: construct multiple objective functions based on multiple statistical indicators of the PM2.5 chemical components, and respectively obtain the optimal solutions corresponding to the multiple objective functions to obtain the Pareto optimal solution as the Pareto frontier; It should be noted that the present invention combines the fast non-dominated sorting genetic algorithm (NSGA-II) with the deep learning model to construct a PM2.5 chemical component concentration vertical profile inversion model based on aerosol lidar; the purpose of normalizing the deep learning model is to improve the quality of the final output; evaluating the quality of the PM2.5 chemical component concentration vertical profile inverted by aerosol lidar often requires considering multiple statistical indicators, including correlation coefficient, root mean square error, regression coefficient, etc. (i.e., the correlation coefficient measures the consistency between the predicted value and the true value; the root mean square error measures the size of the prediction error), and these statistical indicators reflect the expression of different aspects of the model (i.e., fast response and accurate results); By integrating multiple statistical indicators into the objective function, the optimization problem is transformed into a multi-objective optimization problem, aiming to minimize the minimum values ​​of all objective functions at the same time; by integrating multiple statistical indicators into the objective function, the optimization problem is transformed into a multi-objective function optimization problem, aiming to minimize the minimum values ​​of all objective functions at the same time; the diversity of the Pareto front means more choices, which increases the possibility of finding the best solution that meets specific needs (that is, the range of the Pareto front, which can also be called the diversity of the Pareto optimal solution, is similar to the diversity of the "population" in the concept of ecology. The greater the diversity, the greater the choice of the Pareto optimal solution; therefore, the Pareto optimal solution problem is usually solved using an evolutionary algorithm, that is, like the evolution of species, iteratively evolves multiple objective optimization solutions to obtain a set of solutions with the greatest "population advantage", gradually optimizes the solution set, and finds the best solution set); S322: Setting a crowding distance for each Pareto optimal solution in the Pareto front; All Pareto optimal solutions in the corresponding Pareto front of each objective function are sorted from small to large according to the minimum value of the objective function; Assigning the maximum crowding distance to the optimal solution of the minimum value and the optimal solution of the maximum value in the order of the minimum value of the objective function from small to large; It should be noted that, at the beginning, it is assumed that the crowding distance of each Pareto optimal solution is 0. This step is to prepare the basis for subsequent calculations; the crowding distance is used to measure the density of a solution in its Pareto front, that is, how many other solutions are around the solution. A larger crowding distance means that there are fewer solutions around the solution, so it is more representative in the population and helps to maintain the diversity of the population; for each objective function, all solutions in the current Pareto front are sorted from small to large according to the value of the objective function; the optimal solutions with the minimum and maximum objective function values ​​are assigned infinite crowding distances, because these optimal solutions represent the extreme cases of the solution space, and increasing their selection probability helps to explore more solution space; S323: For the intermediate solutions in the ascending order of the minimum value of the objective function (i.e., the optimal solutions in the intermediate part except the optimal solution of the minimum value and the optimal solution of the maximum value in the ascending order of the minimum value of the objective function), the sum of the calculated crowding distances between the adjacent optimal solutions is calculated; the sum of the calculated crowding distances between the adjacent optimal solutions is normalized to obtain the optimal solution of the deep learning model, and the PM2.5 chemical component concentration vertical profile inversion model is obtained; It should be noted that for each objective function in the sorting, the crowding distances of all intermediate solutions (i.e., the second to the second to last solutions) except for the boundary solutions (i.e., the optimal solution of the minimum value and the optimal solution of the maximum value in the sorting of the minimum value of the objective function from small to large) are calculated; Find the optimal solutions before and after each intermediate solution in the order of its objective function value; The sum of the crowding distances between the intermediate solution and its adjacent optimal solutions on the objective function is calculated, and normalized (usually by dividing by the difference between the maximum and minimum values ​​of the objective function) to obtain the final PM2.5 chemical component concentration vertical profile inversion model; Specifically, Figure 6 As shown, in step S32, the optimal solution obtained by the deep learning model is optimized by a fast non-dominated sorting genetic algorithm to obtain a PM2.5 chemical component concentration vertical profile inversion model, which also includes: S321': performing fast non-dominated sorting on the Pareto optimal solutions in the Pareto front, setting the Pareto optimal solutions to levels, and obtaining a first Pareto optimal solution level, a second Pareto optimal solution level, ... an nth Pareto optimal solution level; It should be noted that the Pareto front is sorted by the fast non-dominated sorting genetic algorithm, and all solutions are divided into different levels (the first Pareto optimal solution level, the second Pareto optimal solution level...the nth Pareto optimal solution level) according to their performance in the multi-objective optimization problem; the first level contains the best non-dominated solution set, the second level is second, and so on. Through non-dominated sorting, it is possible to clearly identify which solutions are the best in the current population (that is, solutions at different levels have different priorities; the first Pareto optimal solution level is the best and should be retained first, while solutions at lower levels are relatively inferior; providing a basis for subsequent selection operations to ensure that better solutions are given priority when generating the next generation of populations); S322': assigning a fitness value to the Pareto optimal solution according to the level of the Pareto optimal solution; The allocation of the fitness value is correlated with the level of the Pareto optimal solution (i.e., the fitness value of the solution at the earlier level of the Pareto optimal solution is higher); It should be noted that the fitness value is allocated according to the level where the Pareto optimal solution is located. The earlier the level, the higher the fitness value of the Pareto optimal solution (that is, solutions at different levels have different priorities; the first Pareto optimal solution level is the best and should be retained first, while solutions at lower levels are relatively inferior); a numerical evaluation standard is provided for each Pareto optimal solution to facilitate subsequent genetic operations (such as selection, crossover, and mutation); the fitness value is an important basis for selection operations, and the Pareto optimal solution with a high fitness value has a higher probability of being selected for genetic operations, thereby transmitting its excellent characteristics; through the allocation of fitness values, the diversity and quality of solutions in the population can be balanced to avoid premature convergence to the local optimal solution.

[0025] S323': Adjust the Pareto frontier according to the fitness value of the Pareto optimal solution Generate (That is, the Pareto optimal solution is )’s Pareto frontier ; Through the Pareto frontier of the parent and the Pareto frontier of the offspring Composition The Pareto front set (i.e., the Pareto front set The Pareto front containing the parent The Pareto frontier of the offspring , there are multiple Pareto optimal solutions for Pareto frontiers); It should be noted that based on the parent population (i.e., the Pareto frontier ) is the Pareto optimal solution, and a new offspring population is generated through genetic operations (i.e., the Pareto frontier of the offspring ), the size is still ; Ensure that each generation of the population (i.e., the Pareto frontier of the offspring ) is constant, which helps the algorithm run stably; through genetic operations (selection, crossover, mutation), while maintaining excellent solutions, explore new potential solution spaces and improve global search capabilities; combined with the Pareto frontier of the parent generation and the Pareto frontier of the offspring , that is, composed of (That is, the size of the parent generation's Pareto front is , the Pareto frontier of the offspring The size is also , so ) Pareto front set , prevent the loss of high-quality solutions and enhance the robustness of the algorithm; S324': The Pareto front set Perform fast non-dominated sorting again and set the Pareto front set The Pareto frontier in the set of levels is obtained to obtain the first level Pareto frontier level, the second level Pareto frontier level ... the nth level Pareto frontier level (that is, the step S324' is the same as the step S321', except that the step S324' is a Pareto frontier set Fast non-dominated sorting of multiple Pareto fronts in hierarchical order); It should be noted that the Pareto frontier of the merged parent generation The Pareto frontier of the offspring The Pareto front set (Size is ) Perform fast non-dominated sorting, divide it into multiple levels, and obtain new Pareto frontier levels such as the first level and the second level for the next step of screening; By combining the combined population (i.e., the Pareto front set ) are reordered, and the best and most diverse solutions can be selected to form the next generation population, ensuring the quality and diversity of the population; ensuring that each generation of population contains the best solutions, avoiding the loss of high-quality solutions due to genetic operations, and improving the convergence speed and stability of the algorithm; S321' and S324' stratify the solutions through fast non-dominated sorting, clearly understand the priority relationship between them, and provide a basis for subsequent operations; S325': Determine the Pareto front set The number of Pareto optimal solutions in the Pareto frontier level in the next generation is constructed ; It should be noted that the number of Pareto optimal solutions in the current first-level Pareto frontier is equal to the population size. , if it meets the requirements, then build the next generation of Pareto frontier If not, then the first level Pareto front level, the second level Pareto front level, and the nth level Pareto front level are sequentially sorted according to the level of fast non-dominated sorting to determine whether the number of Pareto optimal solutions is equal to the population size. , until it reaches the population size , building the next generation of Pareto frontiers until; S326': On the Pareto frontier Execute the above steps S322' to S325' to generate a new Pareto front set Finding Pareto optimal solutions; It should be noted that the entire process from fitness value allocation (S322') to population construction (S325') is repeatedly executed to continuously optimize and generate new Pareto optimal solutions. Through multiple iterations, the optimal solution set is gradually approached to improve the overall quality and diversity of the population. Multiple rounds of iterations help the algorithm gradually converge to the global optimal solution while avoiding falling into the local optimum. Each iteration re-evaluates and adjusts the population structure to ensure that the algorithm can flexibly respond to complex optimization problems, thereby obtaining the final Pareto optimal solution. Specifically, in step S326', the Pareto frontier set is determined. The number of Pareto optimal solutions in the Pareto frontier level in the next generation is constructed The specific operation steps are as follows: S3261': Determine the Pareto frontier set Are all the Pareto optimal solutions of the first level Pareto frontier equal to ; If so, then the Pareto front set All Pareto optimal solutions in the first level Pareto frontier hierarchy construct the next generation of Pareto frontier ; If not, then the Pareto front set All Pareto optimal solutions in the first level Pareto frontier do not satisfy (That is, all Pareto optimal solutions in the first level Pareto frontier are not ); It should be noted that the number of Pareto optimal solutions in the current first-level Pareto frontier is equal to the population size. If the number of Pareto optimal solutions at the first level of the Pareto frontier is exactly equal to , it is directly used as the next generation population (i.e., the Pareto frontier ); otherwise, further processing is required; the population must remain fixed in size in each generation , to maintain the stability and predictability of the algorithm; the first level Pareto front level contains the best non-dominated solution set, and giving priority to these Pareto optimal solutions can ensure the quality of the next generation population; S3262': Combine all the Pareto optimal solutions in the second level Pareto frontier layer with all the Pareto optimal solutions in the first level Pareto frontier layer to continue to determine whether they are equal ; If so, then the Pareto front set All Pareto optimal solutions in the first-level Pareto frontier layer and all Pareto optimal solutions in the second-level Pareto frontier layer construct the next generation of Pareto frontier ; If not, continue to follow the Pareto front set The next generation of Pareto frontiers is constructed by fast non-dominated sorting (i.e., the first level Pareto frontier level, the second level Pareto frontier level, ... the nth level Pareto frontier level) , until the conditions are met (that is, until the first level Pareto front level, the second level Pareto front level, or the Pareto optimal solution added to the third level Pareto front level meets the size of ); It should be noted that if the number of Pareto optimal solutions at the first level of the Pareto frontier is insufficient , then add the number of Pareto optimal solutions at the second level Pareto frontier level, and continue to determine whether the total number (i.e., the number of Pareto optimal solutions at the first level Pareto frontier level and the number of Pareto optimal solutions at the second level Pareto frontier level) reaches ; If it is still insufficient, continue to add the number of Pareto optimal solutions of the third level Pareto frontier level, the fourth level Pareto frontier level, etc. until it is satisfied requirements; By gradually adding the Pareto optimal solutions of the lower levels, we can ensure that as many high-quality Pareto optimal solutions as possible are retained while meeting the population size; even if it is not possible to completely fill the Pareto frontier of the next generation with the Pareto optimal solutions of the first level Pareto frontier, The number of Pareto optimal solutions can also be increased by adding suboptimal solutions to maintain the diversity and exploration ability of the population.

[0026] Embodiment 2 like Figure 7 As shown, accordingly, the present invention also proposes a system for PM2.5 chemical component concentration vertical profile inversion model based on aerosol laser radar, including: a preprocessing module 10; a deep learning training module 20; a normalization optimization module 30; The preprocessing module 10 is used to collect 532nm aerosol data as input features using an aerosol laser radar, preprocess the aerosol data, input the preprocessed aerosol data into a deep learning model, and output PM2.5 chemical component features; The deep learning training module 20 is used to perform nonlinear convolution fitting on the aerosol data through a neural convolution network, and optimize the attention mechanism and long short-term memory neural network on the aerosol data after nonlinear fitting to obtain a deep learning model; The normalization optimization module 30 is used for the deep learning model to screen the optimal solution through the Bayesian optimization algorithm and the fast non-dominated sorting genetic algorithm, and normalize the deep learning model to obtain the PM2.5 chemical component concentration vertical profile inversion model.

[0027] In summary, the method and system for the vertical profile inversion model of PM2.5 chemical component concentration based on aerosol lidar proposed in the example of the present invention can be known that, through the input and output data types used by the deep learning model, the input includes aerosol data with a wavelength of 532nm of the laser acquired from the aerosol lidar, the aerosol data is preprocessed, the noise and outliers are removed, so that the PM2.5 chemical component characteristics output when the aerosol data is used as the input feature are more accurate; Furthermore, the preprocessed aerosol data is input into the deep learning model. CNN gradually conducts deep learning on the data set through local perception, sparse connection, and weight and bias sharing, thereby reducing the complexity of the neural network and preventing overfitting. The features in the aerosol data are weighted according to their importance through the attention mechanism, thereby optimizing the weight of the aerosol data and reducing the interference of irrelevant or redundant information. The sequence data is processed from two directions (forward and backward) simultaneously through the long short-term memory neural network. The long short-term memory neural network actually contains two layers (forward and backward) of LSTM: one processes the input sequence from forward to backward, and the other, on the contrary, processes from backward to forward, to capture the time series characteristics of the aerosol data, thereby capturing the time series characteristics of the aerosol data. Furthermore, the deep learning model uses the Bayesian optimization algorithm to construct an objective function for the aerosol data, with the aim of finding a combination of hyperparameters (i.e., hyperparameters) that minimizes the objective function, thereby finding the best hyperparameters. By integrating multiple statistical indicators into the objective function, the optimization problem is transformed into a multi-objective optimization problem, aiming to simultaneously minimize the minimum values ​​of all objective functions and find the possibility of the best solution that meets specific needs as the Pareto front. A crowding distance is set for each Pareto optimal solution in the Pareto front to measure the density of a solution in its Pareto front. All solutions in the current Pareto front are sorted from small to large according to the value of the objective function, because these optimal solutions represent extreme cases in the solution space, and increasing their selection probability helps explore more solution space. In addition to the optimal solution of the minimum value in the order of the minimum value of the objective function and the optimal solution of the maximum value, their crowding distances are calculated, and the sum of the adjacent optimal solutions before and after each intermediate solution in the order of its objective function value is found and normalized to obtain the final PM2.5 chemical component concentration vertical profile inversion model.

[0028] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. A person skilled in the art may modify the technical solutions described in the above embodiments, or replace part or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for inverting a PM2.5 chemical component concentration vertical profile based on an aerosol lidar, characterized in that: The steps are as follows: Using an aerosol lidar to collect 532nm aerosol data as input features, preprocessing the aerosol data, inputting the preprocessed aerosol data into a deep learning model, and outputting PM2.5 chemical component features; Performing nonlinear convolution fitting on the aerosol data through a neural convolution network, performing attention mechanism and long short-term memory neural network optimization on the aerosol data after nonlinear fitting, and obtaining a deep learning model; The deep learning model screens the optimal solution through a Bayesian optimization algorithm and a fast non-dominated sorting genetic algorithm, and normalizes the deep learning model to obtain a vertical profile inversion model of PM2.5 chemical component concentration.

2. According to claim 1, a method for inversion model of PM2.5 chemical component concentration vertical profile based on aerosol lidar is characterized in that: The 532nm aerosol data include: backscattering coefficient characteristics, extinction coefficient characteristics and depolarization ratio characteristics, as well as meteorological element characteristics; The PM2.5 chemical component characteristics include: sulfate, nitrate, ammonium salt, organic matter and black carbon.

3. The method for inverting a PM2.5 chemical component concentration vertical profile based on an aerosol lidar according to claim 2, characterized in that: The aerosol data is convolved and nonlinearly fitted by a neural convolutional network, and the attention mechanism and long short-term memory neural network are optimized for the aerosol data after nonlinear fitting to obtain a deep learning model. The specific operation steps are as follows: The pre-processed aerosol data is input into a deep learning model, and a convolutional neural network is used to perform convolution calculation and nonlinear fitting on the aerosol data; Calculating the weight differences between the features in the aerosol data in the vertical profile inversion task of the PM2.5 chemical components through an attention mechanism; dynamically adjusting the weights of the features in the aerosol data according to the weight differences between the features to optimize the aerosol data; The time of aerosol data collection is bidirectionally transmitted in forward and backward directions through a long short-term memory neural network to capture the time series characteristics of aerosol data.

4. The method of the PM2.5 chemical component concentration vertical profile inversion model based on aerosol lidar according to claim 3 is characterized in that: The deep learning model screens the optimal solution through the Bayesian optimization algorithm and the fast non-dominated sorting genetic algorithm, and normalizes the deep learning model to obtain the PM2.5 chemical component concentration vertical profile inversion model. The specific operation steps are as follows: The deep learning model is coupled online by a Bayesian optimization algorithm, and an objective function is constructed for the aerosol data by a Bayesian optimization algorithm to optimize the optimal solution of the aerosol data; The optimal solution obtained by the deep learning model is normalized by a fast non-dominated sorting genetic algorithm to obtain a vertical profile inversion model of PM2.5 chemical component concentration.

5. The method for inverting the PM2.5 chemical component concentration vertical profile based on aerosol lidar according to claim 4 is characterized in that: The optimal solution obtained by the deep learning model is normalized by a fast non-dominated sorting genetic algorithm to obtain a PM2.5 chemical component concentration vertical profile inversion model. The specific operation steps are as follows: Constructing multiple objective functions based on multiple statistical indicators of the PM2.5 chemical components, and respectively obtaining the optimal solutions corresponding to the multiple objective functions to obtain the Pareto optimal solution as the Pareto frontier; Setting a crowding distance for each Pareto optimal solution in the Pareto front; All Pareto optimal solutions in the corresponding Pareto front of each objective function are sorted from small to large according to the minimum value of the objective function; Assigning the maximum crowding distance to the optimal solution of the minimum value and the optimal solution of the maximum value in the order of the minimum value of the objective function from small to large; The sum of the calculated crowding distances of adjacent optimal solutions is calculated for the intermediate solutions in the order of the minimum value of the objective function from small to large; the sum of the calculated crowding distances of the adjacent optimal solutions is normalized to obtain the optimal solution of the deep learning model, and the vertical profile inversion model of PM2.5 chemical component concentration is obtained.

6. The method for inverting a PM2.5 chemical component concentration vertical profile based on an aerosol lidar according to claim 5, characterized in that: The optimal solution obtained by the deep learning model is optimized by a fast non-dominated sorting genetic algorithm to obtain a PM2.5 chemical component concentration vertical profile inversion model, which also includes: Performing fast non-dominated sorting on the Pareto optimal solutions in the Pareto front, setting the Pareto optimal solutions to levels, and obtaining a first Pareto optimal solution level, a second Pareto optimal solution level, ... an nth Pareto optimal solution level; Allocating a fitness value to the Pareto optimal solution according to the level of the Pareto optimal solution; The Pareto frontier is adjusted according to the fitness value of the Pareto optimal solution. Generate The Pareto frontier of the offspring of ; Through the Pareto frontier of the parent and the Pareto frontier of the offspring Composition The Pareto front set ; The Pareto front set Perform fast non-dominated sorting again and set the Pareto front set The Pareto frontier in the set level, get the first level Pareto frontier level, the second level Pareto frontier level... the nth level Pareto frontier level; Determine the Pareto frontier The number of Pareto optimal solutions in the Pareto frontier level in the next generation is constructed ; The Pareto frontier Execute the above steps to generate a new recursive frontier set Find the Pareto optimal solution.

7. The method for inverting a PM2.5 chemical component concentration vertical profile based on an aerosol lidar according to claim 6, characterized in that: The allocation of the fitness value is correlated with the level of the Pareto optimal solution.

8. The method for inverting a PM2.5 chemical component concentration vertical profile based on an aerosol lidar according to claim 7, characterized in that: Determine the Pareto front set The number of Pareto optimal solutions in the Pareto frontier level in the next generation is constructed , the specific steps are as follows: Determine the Pareto front set Are all the Pareto optimal solutions of the first level Pareto frontier equal to ; If so, then the Pareto front set All Pareto optimal solutions in the first level Pareto frontier hierarchy build the next generation of Pareto frontier ; If not, then the Pareto front set All Pareto optimal solutions in the first level Pareto frontier do not satisfy ; Combine all the Pareto optimal solutions in the second level Pareto frontier layer with all the Pareto optimal solutions in the first level Pareto frontier layer to continue to determine whether they are equal ; If so, then the Pareto front set All Pareto optimal solutions in the first-level Pareto frontier layer and all Pareto optimal solutions in the second-level Pareto frontier layer construct the next generation of Pareto frontier ; If not, continue to follow the Pareto front set Fast non-dominated sorting to construct the next generation of Pareto frontier , until the conditions are met.

9. A system for inversion model of PM2.5 chemical component concentration vertical profile based on aerosol lidar, characterized in that: include: Preprocessing module; Deep learning training module; Normalization optimization module; The preprocessing module is used to collect 532nm aerosol data using aerosol lidar as input features, preprocess the aerosol data, input the preprocessed aerosol data into the deep learning model, and output PM2.5 chemical component features; The deep learning training module is used to perform nonlinear convolution fitting on the aerosol data through a neural convolution network, and optimize the attention mechanism and long short-term memory neural network on the aerosol data after nonlinear fitting to obtain a deep learning model; The normalization optimization module is used to screen the optimal solution of the deep learning model through the Bayesian optimization algorithm and the fast non-dominated sorting genetic algorithm, and normalize the deep learning model to obtain the vertical profile inversion model of PM2.5 chemical component concentration.

Images