Atmospheric PM2.5 source analysis and monthly scale prediction method and system based on artificial intelligence

Abstract

The invention discloses an atmospheric PM2.5 source analysis and monthly scale prediction method and system based on artificial intelligence, and the method comprises the steps: collecting the data information of the atmospheric PM2.5 pollutant component concentration X and the meteorological element M day by day, and carrying out the preprocessing of the data information; on the basis of an Auto-encoder algorithm, an atmospheric pollutant source analysis model based on the Auto-encoder is constructed, the preprocessed atmospheric PM2.5 pollutant component concentration is input into the model, and the relative contribution of a pollution source is obtained; and constructing a KAN-Transformer model, respectively inputting the preprocessed meteorological element data and the obtained relative contribution of the pollution source into the model, and carrying out multi-step prediction to obtain an absolute value of the monthly-scale atmospheric PM2.5 pollutant component concentration. According to the invention, the complex non-linear relationship among the atmospheric pollution source components can be captured, and the accuracy of PM2.5 component source analysis is improved.

Description

Technical Field

[0001] The present invention relates to the field of atmospheric pollutant source analysis and forecasting technology, and specifically to an atmospheric PM2.5 source analysis and monthly scale forecasting method and system based on artificial intelligence. Background Art

[0002] The sources of atmospheric PM2.5 are very complex, involving not only the direct emission of particulate matter (primary pollution), but also the emission and secondary effects of various gaseous pollutants. The current common atmospheric pollution source analysis methods, such as the orthogonal factor matrix (PMF5.0) receptor model proposed by the U.S. Environmental Protection Agency, mainly rely on linear characteristic decomposition and cannot capture the nonlinear relationship between atmospheric pollution source components. Studies have also pointed out that local PM2.5 pollution is highly correlated with meteorological conditions, especially large-scale short-term climate systems, such as El Niño and MJO, which can significantly affect the medium- and short-term PM2.5 pollution situation through the overall atmospheric diffusion conditions, and play an important role in future PM2.5 pollution prediction and government policy adjustments. Previous work mainly focused on single forecasting of atmospheric pollutant concentrations, lacking forecasts and assessments of various atmospheric pollutant components, and lacking the ability to localize forecasts of pollutant concentrations at the monthly scale. Summary of the invention

[0003] The purpose of the present invention is to provide an artificial intelligence-based atmospheric PM2.5 source analysis and monthly scale prediction method and system, based on an artificial intelligence algorithm Auto-encoder with a higher degree of nonlinearity, to conduct PM2.5 source analysis research on the target area, determine the component sources of PM2.5, quantify the contribution of each pollution source, and further construct an improved deep learning model KAN-Transformer to predict the various pollution source components of PM2.5, which is of great significance for carrying out particulate pollution prevention and control work and providing a basis for atmospheric management policies.

[0004] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0005] The atmospheric PM2.5 source analysis and monthly scale prediction method based on artificial intelligence includes the following steps:

[0006] S1. Collect data information on atmospheric PM2.5 pollutant component concentration X and meteorological elements M on a daily basis and pre-process the data information.

[0007] S2. Based on the Auto-encoder algorithm, an Auto-encoder-based atmospheric pollutant source analysis model is constructed, and the pre-treated atmospheric PM2.5 pollutant component concentrations are input into the model to obtain the relative contribution set C of the pollution sources.

[0008] S3. Construct a KAN-Transformer (Kolmogorov-Arnold Networks-Transformer) model, input the meteorological element data preprocessed in step S1 and the relative contribution of pollution sources obtained in step S2 into the model respectively, and obtain the absolute value of the atmospheric PM2.5 pollutant component concentration on a monthly scale through multi-step prediction.

[0009] Furthermore, in step S1, preprocessing the data information includes the following sub-steps:

[0010] S101. Meteorological elements include wind speed and direction, temperature and relative humidity. Quality control is performed on data information to monitor whether abnormal values ​​appear. When abnormal values ​​appear, they are replaced with the valid values ​​of the previous moment. For wind speed data, abnormal values ​​are data below 0m / s or above 40m / s; for wind direction data, abnormal values ​​are data below 0° or above 360°; for temperature data, abnormal values ​​are data below -30℃ or above 50℃; for relative humidity, abnormal values ​​are data below 0 or above 100%.

[0011] S102. Normalize the data information after quality control. The specific formula is:

[0012]

[0013] Among them, z max and z min Respectively represent the maximum and minimum values ​​in the original data information, z norm represents the normalized data information, and z represents the original data information.

[0014] The normalized PM2.5 pollutant component concentration data is recorded as X norm , the normalized meteorological data is recorded as M norm .

[0015] Furthermore, in step S2, obtaining the relative contribution of the pollution source includes the following sub-steps:

[0016] S201. The air pollutant source analysis model based on Auto-encoder includes an encoder and a decoder.

[0017] S202, arranging the concentrations of the pre-treated atmospheric PM2.5 pollutant components in time series to obtain an observation sample set X with s features norm , X norm ={x1,x2,...,x s}, x sRepresents the s-th feature; the observed samples are input into the atmospheric pollutant source parsing model. After dimensionality reduction by the encoder, the s-dimensional features in the atmospheric PM2.5 pollutant component concentrations are reduced in dimension to obtain the feature vector P, P = {p1, p2,..., p k}, p k represents the k-th feature vector value, and k < s; the feature vector P is processed by the decoder to obtain the reconstructed set of observed samples X c .

[0018] S203. Calculate the error between the set of observed samples X norm and the reconstructed set of observed samples X c . Use iterative optimization of the reconstruction error to obtain the trained atmospheric pollutant source parsing model.

[0019] S204. Input the set of observed samples X norm into the trained atmospheric pollutant source parsing model. After dimensionality reduction by the encoder, obtain the corresponding dimensionality-reduced features.

[0020] S205. Standardize the dimensionality-reduced features to obtain the standardized dimensionality-reduced feature vector C. The specific formula is:

[0021]

[0022] where c i represents the i-th standardized feature vector value, and p i represents the i-th feature vector value in the dimensionality-reduced features;

[0023] C = {c1, c2,..., c i ,..., c k}, c k represents the k-th standardized feature vector value, which is the set of relative contributions of the k typical pollution sources causing the PM2.5 pollutant component concentrations.

[0024] Furthermore, in step S3, obtaining the absolute value of the monthly-scale atmospheric PM2.5 pollutant component concentrations includes the following sub-steps:

[0025] S301. The KAN-Transformer model includes a linear mapping, position encoding, N KAN-Transformer modules, and a fully connected layer; where the KAN-Transformer module includes two regularization layers, a self-attention mechanism, a residual structure, and a KAN network.

[0026] S302. Combine the set of relative contributions of pollution sources C at time t with the preprocessed meteorological element data M normThe information is input into the KAN-Transformer model, and after linear mapping and position encoding, the processed encoded information is obtained, and the processed encoded information is input into N KAN-Transformer modules.

[0027] For a single KAN-Transformer module, the first regularization layer is used to regularize the processed encoded information. The regularized result is passed through the self-attention mechanism and three sets of randomly initialized weights, namely, query weights W Q , key weight W K and numerical weight W V , we get three features, namely: Q = W Q I, K = W K I and V = W V I, where Q represents the query matrix, K represents the key matrix, V represents the numerical matrix, and I represents the identity matrix, and then the self-attention F corresponding to I is obtained. Indicates custom dimension d n The square root of K T Represents the transpose of K; a residual structure is constructed between the regular layer and the self-attention layer, and the processed encoded information is added through the residual structure and the self-attention F to obtain the result of the residual structure.

[0028] The result of the residual structure is processed by the second regularization layer and then extracted by the KAN network. The KAN network is set to a three-layer structure, with u inputs in the first layer, v neuron nodes in the second layer, and y outputs in the third layer. The output result of the second layer can be obtained by using u×v single variable functions φ i,j Linear summation is performed to represent and serve as the input of the third layer. The output of the third layer is v×y single variable functions Φ i,j The linear summation of u×v single variable functions φ i,j and v×y univariate functions Φ i,j , and obtain the nonlinear relationship between the first layer input and the third layer output.

[0029] N KAN-Transformer modules are spliced ​​together to obtain the relative contribution set of pollution sources at time t and the feature extraction results of the preprocessed meteorological element data.

[0030] S303: Input the feature extraction result in step S302 into the fully connected network to obtain the relative contribution C' of the pollution source and the total concentration S' of PM2.5 at time t+1.

[0031] S304. Multiply the relative contribution C' of the pollution source at time t+1 by the total concentration S' of PM2.5 to obtain the absolute concentration X' of the atmospheric PM2.5 components caused by each pollution source at time t+1, and repeat steps S302 and S303 to obtain the actual PM2.5 concentration emission value of each pollution source within the future set time.

[0032] Furthermore, the present invention also proposes an artificial intelligence-based atmospheric PM2.5 source analysis and monthly scale prediction system, including:

[0033] The data preprocessing module is used to collect data information on the concentration of atmospheric PM2.5 pollutant components and meteorological elements on a daily basis and to preprocess the data information.

[0034] The pollution source relative contribution acquisition module is used to build an atmospheric pollutant source analysis model based on the Auto-encoder algorithm, input the atmospheric PM2.5 pollutant component concentrations into the model, and obtain the relative contribution of the pollution sources.

[0035] The predicted value acquisition module is used to construct the KAN-Transformer model. The preprocessed meteorological element data and the obtained relative contribution of pollution sources are respectively input into the model, and the absolute value of the atmospheric PM2.5 pollutant component concentration on a monthly scale is obtained through multi-step prediction.

[0036] Furthermore, the present invention also proposes an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the steps of the above-mentioned artificial intelligence-based atmospheric PM2.5 source analysis and monthly-scale prediction method are implemented.

[0037] Furthermore, the present invention also proposes a computer-readable storage medium, which stores a computer program. When the computer program is run by a processor, it executes the artificial intelligence-based atmospheric PM2.5 source analysis and monthly-scale prediction method.

[0038] Compared with the prior art, the present invention adopts the above technical solution and has the following technical effects:

[0039] The present invention constructs a KAN-Transformer model to predict the PM2.5 component concentration on a monthly scale from the perspective of the contribution of PM2.5 component concentration caused by atmospheric PM2.5 pollution sources, thereby quantifying the monthly scale impact of each atmospheric PM2.5 pollution source, capturing the complex nonlinear relationship between atmospheric pollution source components, improving the accuracy of PM2.5 component source analysis, and obtaining more accurate and reliable quarterly and monthly forecasts of PM2.5 concentrations. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is an overall implementation flow chart of the present invention.

[0041] Figure 2 It is a schematic diagram of the atmospheric pollutant source analysis model of the present invention.

[0042] Figure 3 It is a schematic diagram of the absolute value of the atmospheric PM2.5 pollutant component concentration predicted by the present invention on a monthly scale. DETAILED DESCRIPTION

[0043] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present invention.

[0044] To achieve the above objectives, the present invention proposes an atmospheric PM2.5 source analysis and monthly scale prediction method based on artificial intelligence, such as Figure 1 As shown, the specific steps are as follows:

[0045] S1. Collect data information on atmospheric PM2.5 pollutant component concentration X and meteorological elements M on a daily basis and pre-process the data information; the specific contents are as follows:

[0046] S101. Meteorological elements include wind speed and direction, temperature and relative humidity. Quality control is performed on data information to monitor whether abnormal values ​​appear. When abnormal values ​​appear, they are replaced with the valid values ​​of the previous moment. For wind speed data, abnormal values ​​are data below 0m / s or above 40m / s; for wind direction data, abnormal values ​​are data below 0° or above 360°; for temperature data, abnormal values ​​are data below -30℃ or above 50℃; for relative humidity, abnormal values ​​are data below 0 or above 100%.

[0047] S102. Normalize the data information after quality control. The specific formula is:

[0048]

[0049] Among them, z max and z min Respectively represent the maximum and minimum values ​​in the original data information, z norm represents the normalized data information, and z represents the original data information.

[0050] The normalized PM2.5 pollutant component concentration data is recorded as X norm , the normalized meteorological data is recorded as M norm .

[0051] S2. Based on the Auto-encoder algorithm, construct an atmospheric pollutant source analysis model based on Auto-encoder, and input the preprocessed concentration of atmospheric PM2.5 pollutant components into this model to obtain the relative contribution set C of the pollution sources; as Figure 2 shown, the specific content is as follows:

[0052] S201. The atmospheric pollutant source analysis model based on Auto-encoder includes an encoder and a decoder.

[0053] S202. Arrange the preprocessed concentration of atmospheric PM2.5 pollutant components in time series, such as sample ①, sample ②, etc., to obtain an observation sample set X with s features norm , X norm = {x1, x2,..., x s}, x s represents the s-th feature. Input the observation samples into the atmospheric pollutant source analysis model. After dimensionality reduction by the encoder, reduce the s-dimensional features in the concentration of atmospheric PM2.5 pollutant components to obtain a feature vector P, P = {p1, p2,..., p k}, p k represents the k-th feature vector value, and k < s; the feature vector P is processed by the decoder to obtain the reconstructed observation sample set X c .

[0054] S203. Calculate the root mean square error between the observation sample set X norm and the reconstructed observation sample set X c , and use iterative optimization of the reconstruction error to obtain the trained atmospheric pollutant source analysis model.

[0055] S204. Input the observation sample set X norm into the trained atmospheric pollutant source analysis model. After dimensionality reduction by the encoder, obtain the corresponding dimensionality-reduced features.

[0056] S205. Standardize the dimensionality-reduced features to obtain the standardized dimensionality-reduced feature vector C. The specific formula is:

[0057]

[0058] where c i represents the i-th standardized feature vector value, and p i represents the i-th feature vector value in the dimensionality-reduced features;

[0059] C = {c1, c2,..., c i ,..., c k}, c kIt represents the kth eigenvector value after standardization, which is the relative contribution set of k typical pollution sources that cause the concentration of PM2.5 pollutant components.

[0060] S3, constructing a KAN-Transformer (Kolmogorov-Arnold Networks-Transformer) model, inputting the meteorological element data preprocessed in step S1 and the relative contribution of the pollution source obtained in step S2 into the model respectively, and obtaining the absolute value of the atmospheric PM2.5 pollutant component concentration on a monthly scale through multi-step prediction; Figure 3 The specific contents are as follows:

[0061] S301, KAN-Transformer model includes linear mapping, position encoding, N KAN-Transformer modules and fully connected layer; the KAN-Transformer module includes two regularization layers, self-attention mechanism, residual structure and KAN network.

[0062] S302: compare the relative contribution set C of the pollution sources (such as pollution source A, pollution source B, etc.) at time t with the preprocessed meteorological element data M norm The information is input into the KAN-Transformer model, and after linear mapping and position encoding, the processed encoded information is obtained, and the processed encoded information is input into N KAN-Transformer modules.

[0063] For a single KAN-Transformer module, the first regularization layer is used to regularize the processed encoded information so that the sample data in the network obeys the same distribution and improves the convergence efficiency of the model. The regularized result is passed through the self-attention mechanism and three sets of randomly initialized weights are used, namely, the query weight W Q , key weight W K and numerical weight W V , we get three features, namely: Q = W Q I, K = W K I and V = W V I, where Q represents the query matrix, K represents the key matrix, V represents the numerical matrix, and I represents the identity matrix, and then the self-attention F corresponding to I is obtained. Indicates custom dimension d n The square root of K TRepresents the transpose of K; a residual structure is constructed between the regular layer and the self-attention layer, and the processed encoded information is added through the residual structure and the self-attention F to obtain the result of the residual structure, which helps to maintain the continuity of information transmission in the model and improve the representation and generalization capabilities of the model.

[0064] The result of the residual structure is processed by the second regularization layer and then extracted by the KAN network. The KAN network is a neural network model based on the Kolmogorov-Arnold theorem. Its core idea is: for any multivariate continuous function, it can be expressed as a combination of a finite number of single variable functions (such as exponential functions, quadratic functions, etc.) and addition; the KAN network is set to a three-layer structure, the first layer has u inputs, the second layer has v neuron nodes, and the third layer has y outputs. The output result of the second layer can be expressed using u×v single variable functions φ i,j Linear summation is performed to represent and serve as the input of the third layer. The output of the third layer is v×y single variable functions Φ i,j The linear summation of u×v single variable functions φ i,j and v×y univariate functions Φ i,j , and the nonlinear relationship between the first layer input and the third layer output is obtained. It can be seen that the KAN network uses the sum of multiple sets of univariate functions to fit the statistical relationship between the network input and output. Its activation function is not fixed on the network node, but is located on the edge of the network (i.e., weight) as a learnable univariate function; using the KAN network to replace the original feedforward neural network can make the model achieve better generalization ability with a smaller number of parameters and have stronger interpretability.

[0065] N KAN-Transformer modules are spliced ​​together to obtain the relative contribution set of pollution sources at time t and the feature extraction results of the preprocessed meteorological element data.

[0066] S303: Input the feature extraction result in step S302 into the fully connected network to obtain the relative contribution C' of the pollution source and the total concentration S' of PM2.5 at time t+1.

[0067] S304. Multiply the relative contribution C' of the pollution source at time t+1 by the total concentration S' of PM2.5 to obtain the absolute concentration X' of the atmospheric PM2.5 components caused by each pollution source at time t+1, and repeat steps S302 and S303 to obtain the actual PM2.5 concentration emission values ​​of each pollution source in the next 30 days.

[0068] The embodiment of the present invention also proposes an atmospheric PM2.5 source analysis and monthly scale prediction system based on artificial intelligence, including a data preprocessing module, a pollution source relative contribution acquisition module, a prediction value acquisition module and a computer program that can be run on a processor. It should be noted that each module in the above system corresponds to the specific steps of the method provided in the embodiment of the present invention, and has the corresponding functional modules and beneficial effects of the execution method. For technical details not described in detail in this embodiment, please refer to the method provided in the embodiment of the present invention.

[0069] The embodiment of the present invention further provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. It should be noted that when the processor executes the computer program, the specific steps of the method provided in the embodiment of the present invention are corresponding to the specific steps of the method provided in the embodiment of the present invention, and the processor has the functional modules and beneficial effects corresponding to the execution method. For technical details not described in detail in this embodiment, please refer to the method provided in the embodiment of the present invention.

[0070] The embodiment of the present invention further provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program. It should be noted that when the computer program is executed by the processor, it corresponds to the specific steps of the method provided in the embodiment of the present invention, and has the functional modules and beneficial effects corresponding to the execution method. For technical details not described in detail in this embodiment, please refer to the method provided in the embodiment of the present invention.

[0071] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. An artificial intelligence-based atmospheric PM2.5 source analysis and monthly scale prediction method, characterized in that: include: S1. Collect data information of atmospheric PM2.5 pollutant component concentration X and meteorological elements M on a daily basis and pre-process the data information; S2. Based on the Auto-encoder algorithm, an Auto-encoder-based atmospheric pollutant source analysis model is constructed, and the concentration of the pre-treated atmospheric PM2.5 pollutant components is input into the model to obtain the relative contribution set C of the pollution sources; S3. Construct a KAN-Transformer model, input the meteorological element data preprocessed in step S1 and the relative contribution of pollution sources obtained in step S2 into the model respectively, and obtain the absolute value of the atmospheric PM2.5 pollutant component concentration on a monthly scale through multi-step prediction.

2. The method for atmospheric PM2.5 source analysis and monthly scale prediction based on artificial intelligence according to claim 1, characterized in that: In step S1, preprocessing the data information includes the following sub-steps: S101. Meteorological elements include wind speed and direction, temperature and relative humidity. Quality control is performed on data information to monitor whether abnormal values ​​appear. When abnormal values ​​appear, they are replaced with the valid values ​​at the previous moment. For wind speed data, abnormal values ​​are data below the minimum wind speed or above the maximum wind speed. For wind direction data, abnormal values ​​are data below the minimum wind direction or above the maximum wind direction. For temperature data, abnormal values ​​are data below the minimum temperature or above the maximum temperature. For relative humidity, abnormal values ​​are data below the minimum relative humidity or above the maximum relative humidity. S102. Normalize the data information after quality control. The specific formula is: Among them, z max and z min Respectively represent the maximum and minimum values ​​in the original data information, z norm represents the normalized data information, and z represents the original data information; The normalized PM2.5 pollutant component concentration data is recorded as X norm , the normalized meteorological data is recorded as M norm .

3. The method for atmospheric PM2.5 source analysis and monthly scale prediction based on artificial intelligence according to claim 2, characterized in that: In step S1, for wind speed data, the abnormal value is data below 0m / s or above 40m / s; for wind direction data, the abnormal value is data below 0° or above 360°; for temperature data, the abnormal value is data below -30°C or above 50°C; for relative humidity, the abnormal value is data below 0 or above 100%.

4. The method for atmospheric PM2.5 source analysis and monthly scale prediction based on artificial intelligence according to claim 1, characterized in that: In step S2, obtaining the relative contribution of the pollution source includes the following sub-steps: S201. The air pollutant source analysis model based on Auto-encoder includes an encoder and a decoder; S202. Arrange the concentrations of the pre - processed atmospheric PM2.5 pollutant components in a time series to obtain an observation sample set X with s features norm , X norm = {x1, x2,..., x s}, where x s represents the s - th feature; input the observation samples into the atmospheric pollutant source analysis model. After the dimensionality reduction processing by the encoder, the s - dimensional features in X norm are reduced in dimension to obtain a feature vector P, P = {p1, p2,..., p k}, where p k represents the value of the k - th feature vector, and k < s; the feature vector P is processed by the decoder to obtain the reconstructed observation sample set X c ; S203, calculate the observation sample set X norm and the reconstructed observation sample set X c The error between them is reconstructed by iterative optimization to obtain the trained atmospheric pollutant source analysis model; S204, observe the sample set X norm The input is fed into the trained atmospheric pollutant source analysis model, and the corresponding dimension reduction features are obtained after the dimensionality reduction processing of the encoder. S205, standardize the dimension reduction feature to obtain a standardized dimension reduction feature vector. The specific formula is: Among them, c i represents the standardized eigenvector value; p i Represents the i-th eigenvector value in the dimensionality reduction feature; All eigenvector values ​​​​compose the standardized reduced-dimensionality eigenvector C, C={c1,c2,...,c i ,...,c k },c k It represents the kth eigenvector value after standardization, which is the relative contribution set of k typical pollution sources that cause the concentration of PM2.5 pollutant components.

5. The method for atmospheric PM2.5 source analysis and monthly scale prediction based on artificial intelligence according to claim 1, characterized in that: In step S3, obtaining the absolute value of the atmospheric PM2.5 pollutant component concentration on a monthly scale includes the following sub-steps: S301, KAN-Transformer model includes linear mapping, position encoding, N KAN-Transformer modules and fully connected layer; the KAN-Transformer module includes two regularization layers, self-attention mechanism, residual structure and KAN network; N KAN-Transformer modules are connected in sequence. S302: The relative contribution set C of the pollution sources at time t is compared with the pre-processed meteorological element data M norm Input into the KAN-Transformer model, and after linear mapping and position encoding processing, the processed encoding information is obtained, and the processed encoding information is input into the first KAN-Transformer module. The output of the previous KAN-Transformer module is used as the input of the next KAN-Transformer module; For the first KAN-Transformer module, the first regularization layer is used to regularize the processed encoded information. The regularized result is passed through the self-attention mechanism and three sets of randomly initialized weights are used, namely, the query weight W Q , key weight W K and numerical weight W V , we get three features, namely: Q = W Q I, K = W K I and V = W V I, where Q represents the query matrix, K represents the key matrix, V represents the numerical matrix, and I represents the identity matrix, and then the self-attention F corresponding to I is obtained. Indicates custom dimension d n The square root of K T represents the transpose of K; a residual structure is constructed between the regular layer and the self-attention layer, and the processed encoded information is added through the residual structure and the self-attention F to obtain the result of the residual structure; The result of the residual structure is processed by the second regularization layer and then extracted by the KAN network. The KAN network is set to a three-layer structure. The first layer has u inputs and the second layer has v neuron nodes. The output result of the second layer is u×v single variable functions φ i,j Linear summation is performed to represent it and used as the input of the third layer. The output of the third layer is v×y single variable functions Φ i,j The linear summation of u×v single variable functions φ i,j and v×y univariate functions Φ i,j , we get the nonlinear relationship between the first layer input and the third layer output, and then get the y outputs of the third layer; N KAN-Transformer modules are spliced ​​together to obtain the relative contribution set of pollution sources at time t and the feature extraction results of the preprocessed meteorological element data; S303, inputting the feature extraction result in step S302 into the fully connected network to obtain the relative contribution C' of the pollution source and the total concentration S' of PM2.5 at time t+1; S304. Multiply the relative contribution C' of the pollution source at time t+1 by the total concentration S' of PM2.5 to obtain the absolute concentration X' of the atmospheric PM2.5 components caused by each pollution source at time t+1; repeat steps S302 and S303 to obtain the actual PM2.5 concentration emission value of each pollution source within the future set time.

6. A system for the artificial intelligence-based atmospheric PM2.5 source analysis and monthly scale prediction method applied to claim 1, characterized in that: include: The data preprocessing module is used to collect data information on the concentration of atmospheric PM2.5 pollutants and meteorological elements on a daily basis and preprocess the data information; The pollution source relative contribution acquisition module is used to build an atmospheric pollutant source analysis model based on the Auto-encoder algorithm, input the atmospheric PM2.5 pollutant component concentrations into the model, and obtain the relative contribution of the pollution source; The predicted value acquisition module is used to construct the KAN-Transformer model. The preprocessed meteorological element data and the obtained relative contribution of pollution sources are respectively input into the model, and the absolute value of the atmospheric PM2.5 pollutant component concentration on a monthly scale is obtained through multi-step prediction.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of any one of the artificial intelligence-based atmospheric PM2.5 source analysis and monthly-scale prediction methods of claims 1 to 5 are implemented.

8. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the atmospheric PM2.5 source analysis and monthly scale prediction method based on artificial intelligence described in any one of claims 1 to 5 is executed.

Images