PM2.5 concentration prediction method, device and system and storage medium

Through the VMD-LSTM model optimized by variational modal decomposition and gray wolf optimization algorithm, and the CNN-LSTM model based on the maximum information coefficient, the problems of low accuracy and neglected spatial factors in the prediction of PM2.5 concentration are solved. A more efficient PM2.5 concentration prediction was achieved.

CN118917475AInactive Publication Date: 2024-11-08GUANGXI UNIVERSITY OF TECHNOLOGY

Patent Information

Application Number
CN202410990217.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-23
Publication Date
2024-11-08
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

In the prediction of PM2.5 concentration, traditional convolutional neural networks have low prediction accuracy due to the complexity of data fluctuations and change information, and ignore the factors influencing the spatial influence of PM2.5 concentration.

Method used

Variable modal decomposition (VMD) combined with particle swarm optimization algorithm is used to extract multiple eigenmodulo functions of PM2.5 concentration sequence and input them into long short-term memory network (LSTM) for prediction, and optimize weights using Gray Wolf Optimization Algorithm ; At the same time, a spatial characteristic factor based on the maximum information coefficient was constructed, combined with air pollutants and meteorological factors, and input the CNN-LSTM model for multi-character fusion prediction.

Benefits of technology

The accuracy of PM2.5 concentration prediction is improved, the difficulty of fitting the deep learning model is reduced, the consideration of spatial influencing factors is enhanced, and the prediction accuracy and reliability are improved.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118917475A_ABST
    Figure CN118917475A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of environmental protection, and particularly discloses a PM2.5 concentration prediction method, device and system and a storage medium, a PM2.5 concentration sequence is decomposed by using a variational mode decomposition method, parameters in VMD are optimized by applying a particle swarm optimization algorithm, and the parameters are determined by combining an observation center frequency, so that the PM2.5 concentration is predicted. The VMD can deeply extract features and information of the sequence at different frequencies, so that the fitting difficulty of a subsequent deep learning model is reduced, and the efficiency of the model is improved; a plurality of eigenmode functions obtained through decomposition are input into a long short-term memory network for fitting to obtain a predicted value of each component, in order to further improve the prediction precision of the model, a grey wolf optimization algorithm is introduced to reconstruct and optimize the weight of each component, the contribution of each component to a prediction result is continuously corrected, and the prediction precision of the model is improved. And correspondingly multiplying the weight by each component prediction result, adding, and outputting a prediction value of the VMD-LSTM-GWO combined model, thereby solving the problems of complexity of PM2.5 concentration data fluctuation change information and low prediction precision of a traditional model.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of environmental protection technology, and in particular to a PM 2.5 Concentration prediction method, device, system and storage medium. Background Art

[0002] In air quality monitoring, the Air Quality Index (AQI) is a key quantitative parameter used to describe the state of the atmosphere in detail. Referring to the Technical Provisions on Ambient Air Quality Index (AQI) (Trial) promulgated by the Ministry of Environmental Protection of my country in 2012, the air quality sub-index for a specific pollutant item P is calculated using the following formula.

[0003]

[0004] In the formula, IAQI P is the air quality index of pollutant item P; C p is the mass concentration value of pollutant item P; BP Hi For C p High position with similar pollutant concentration limit; BP LO For C p Lower value of similar pollutant concentration limit; IAQI LO For BP LO Corresponding air quality index; IAQI Hi Is with BP Hi The corresponding air quality index.

[0005] The air quality index is calculated using the following formula.

[0006] AQI=max{IAQI 1 ,IAQI 2 ,IAQI 3 ,...IAQI n}.

[0007] In the formula, IAQI represents the air quality index; n represents the pollutant item.

[0008] PM 2.5 The concentration indicates the concentration of dust or floating dust with a diameter less than or equal to 2.5 μm in the ambient air. 2.5 Concentration is often used to measure air quality. PM 2.5 The prediction of concentration is of great significance to the control of air quality. 2.5In concentration prediction, CNN is a feedforward neural network model that integrates deep structures, and its characteristic is the application of convolution operation mechanism. As a deep learning framework, CNN aims to significantly reduce the need for raw data preprocessing steps and show superior performance compared to traditional algorithms. CNN, inspired by the biological visual system, creatively integrates the convolution principle into the neural network architecture and has been widely used and verified in many fields. The concept of CNN architecture can be traced back to 1989. Since then, the performance of CNN has been significantly enhanced and continuously improved through gradient optimization technology. A complete CNN covers multiple core components, including but not limited to input layer, convolution layer, application of activation function, introduction of pooling layer, construction of fully connected layer, and setting of loss function. The key is to achieve the most accurate feature extraction possible by carefully combining these network layer structures. In this process, the design and configuration of each layer are crucial. They work together to enable CNN to efficiently process input data and extract feature information useful for subsequent tasks.

[0009] When using traditional convolutional neural networks for PM 2.5 When predicting concentration, PM 2.5 The complexity of concentration data fluctuations leads to low prediction accuracy of traditional models. Summary of the invention

[0010] In order to solve the above problems, the present invention provides a PM 2.5 Concentration prediction method, device, system and storage medium to solve PM 2.5 The complexity of concentration data fluctuations leads to low prediction accuracy of traditional models.

[0011] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0012] According to a first embodiment of the present invention, there is provided a PM 2.5 A concentration prediction method, the method comprising:

[0013] About PM 2.5 The concentration value time series data is subjected to variational mode decomposition to obtain multiple decomposition sequences, which are eigenmode functions. When performing variational mode decomposition, the envelope entropy in the variational mode decomposition is used as the first objective function, and the minimum envelope entropy is used as the target value of the first objective function. The particle swarm optimization algorithm is used to determine the two parameters k and α in the variational mode decomposition, where the parameter k is used to define the PM 2.5The number of modes selected when performing variational mode decomposition of concentration value time series data is used to determine the spectrum coverage of each decomposition sequence with reference to α; the multiple decomposition sequences are fed into the LSTM network model to obtain the predicted value of each decomposition sequence; the predicted value of each decomposition sequence is compared with the PM 2.5 The mean square error between the actual concentration values ​​is used as the objective function, and the minimum mean square error is used as the target value. The weight value of each decomposition sequence is determined by the gray wolf optimization algorithm; the weight value of each decomposition sequence is multiplied by the predicted value and then accumulated to obtain PM 2.5 Predicted concentration.

[0014] According to the second technical solution of the present invention, there is provided a PM 2.5 A concentration prediction device, the device comprising a processor, wherein the processor is configured to:

[0015] About PM 2.5 The concentration value time series data is subjected to variational mode decomposition to obtain multiple decomposition sequences, which are eigenmode functions. When performing variational mode decomposition, the envelope entropy in the variational mode decomposition is used as the first objective function, and the minimum envelope entropy is used as the target value of the first objective function. The particle swarm optimization algorithm is used to determine the two parameters k and α in the variational mode decomposition, where the parameter k is used to define the PM 2.5 The number of modes selected when performing variational mode decomposition of concentration value time series data is used to determine the spectrum coverage of each decomposition sequence with reference to α; the multiple decomposition sequences are fed into the LSTM network model to obtain the predicted value of each decomposition sequence; the predicted value of each decomposition sequence is compared with the PM 2.5 The mean square error between the actual concentration values ​​is used as the objective function, and the minimum mean square error is used as the target value. The weight value of each decomposition sequence is determined by the gray wolf optimization algorithm; the weight value of each decomposition sequence is multiplied by the predicted value and then accumulated to obtain PM 2.5 Predicted concentration.

[0016] According to the third technical solution of the present invention, there is provided a PM 2.5 A concentration prediction system, the system comprising: a memory for storing a computer program; and a processor for executing the computer program to implement the method as described above.

[0017] According to a fourth technical solution of the present invention, a non-transitory computer-readable storage medium storing instructions is provided. When the instructions are executed by a processor, the method described above is executed.

[0018] The present invention has at least the following beneficial effects:

[0019] 1. The present invention proposes a VMD-LSTM-GWO combined prediction model based on time series. The model first uses the variational mode decomposition method to 2.5The concentration sequence is decomposed, and the particle swarm optimization algorithm is used to optimize the parameters in VMD, and the parameters are determined by combining the observation of the central frequency. VMD can deeply extract the characteristics and information of the sequence at different frequencies, thereby reducing the difficulty of fitting the subsequent deep learning model and improving the efficiency of the model; the several intrinsic mode functions obtained by decomposition are input into the long short-term memory network for fitting to obtain the predicted value of each component. In order to further improve the prediction accuracy of the model, the gray wolf optimization algorithm is used to reconstruct and optimize the weights of each component, and the contribution of each component to the prediction result is continuously corrected. The weights are multiplied and added with the corresponding prediction results of each component, and the predicted value of the VMD-LSTM-GWO combined model is output to solve PM 2.5 The complexity of concentration data fluctuation information and the low prediction accuracy of traditional models.

[0020] 2. The present invention also constructs a PM based on the maximum information coefficient 2.5 The spatial characteristic factors of concentration are used to calculate the PM2.5 concentration between air quality monitoring stations in a certain study area. 2.5 The maximum information coefficient of the concentration is used as the element of the spatial weight matrix, which is combined with the PM 2.5 The spatial characteristic factor is obtained by multiplying the data set consisting of the concentration, which is further used as the PM 2.5 The concentration of PM2.5 is an influencing factor. It is combined with other air pollutant factors and meteorological factors and input into the CNN-LSTM model for fitting prediction. The predicted value of the CNN-LSTM combined model based on multi-feature fusion is output to solve the problem of ignoring PM2.5 in existing prediction methods. 2.5 The problem of spatial influencing factors of concentration. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 A data preprocessing flow chart according to an embodiment of the present invention is shown;

[0022] Figure 2 A framework diagram of a VMD-LSTM-GWO model according to an embodiment of the present invention is shown;

[0023] Figure 3 The PM data for the study area from 2021.01.01 to 2023.09.30 every three hours according to an embodiment of the present invention is shown. 2.5 Time series graph of concentration values;

[0024] Figure 4 A schematic diagram of visualization of VMD decomposition results according to an embodiment of the present invention is shown;

[0025] Figure 5 The internal structure diagram of the LSTM network according to an embodiment of the present invention is shown;

[0026] Figure 6 It shows various IMF forecast result diagrams according to an embodiment of the present invention;

[0027] Figure 7 A comparison diagram between the prediction results of the VMD-LSTM-GWO model according to an embodiment of the present invention and the true value is shown;

[0028] Figure 8 A comparison diagram of prediction results of different models and true values ​​according to an embodiment of the present invention is shown;

[0029] Fig. 9 A CNN-LSTM model framework diagram based on multi-feature fusion according to an embodiment of the present invention is shown;

[0030] Fig.10 The PM diagram of each site according to an embodiment of the present invention is shown. 2.5 Concentration correlation coefficient;

[0031] Fig.11 A comparison diagram of the prediction results of the CNN-LSTM model based on multi-feature fusion according to an embodiment of the present invention and the actual value is shown;

[0032] Fig.12 A comparison diagram of the CNN model prediction results and the true values ​​based on different input features according to an embodiment of the present invention is shown;

[0033] Fig.13 A comparison diagram of LSTM model prediction results and true values ​​based on different input features according to an embodiment of the present invention is shown;

[0034] Fig.14 A comparison chart of the prediction results of the CNN-LSTM model based on different input features and the true values ​​according to an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0035] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict.

[0036] The specific implementation of the present invention is further described in detail below in conjunction with the drawings and examples.

[0037] The embodiment of the present invention provides a PM 2.5 Concentration prediction methods to achieve PM2.5 Concentration prediction, this example uses two models to achieve PM 2.5 One of them is the VMD-LSTM-GWO combined prediction model based on time series, which is used to solve PM 2.5 The VMD-LSTM-GWO combined prediction model first uses the variational mode decomposition method to predict PM 2.5 The concentration sequence is decomposed, and the particle swarm optimization algorithm is used to optimize the parameters in VMD, and the parameters are determined by observing the central frequency. VMD can deeply extract the characteristics and information of the sequence at different frequencies, thereby reducing the difficulty of fitting the subsequent deep learning model and improving the efficiency of the model; several intrinsic mode functions obtained by decomposition are input into the long short-term memory network for fitting to obtain the predicted value of each component. In order to further improve the prediction accuracy of the model, the gray wolf optimization algorithm is used to reconstruct and optimize the weights of each component, and the contribution of each component to the prediction result is continuously corrected. The weight is multiplied and added with the corresponding prediction result of each component to output the predicted value of the VMD-LSTM-GWO combined model.

[0038] The other is a CNN-LSTM combined prediction model based on multi-feature fusion to solve the problem of ignoring PM in existing prediction methods. 2.5 The problem of spatial influencing factors of concentration. First, calculate the PM between air quality monitoring stations in a certain area. 2.5 The maximum information coefficient of the concentration is used as the element of the spatial weight matrix, which is combined with the PM 2.5 The spatial characteristic factor is obtained by multiplying the data set consisting of the concentration, which is further used as the PM 2.5 The concentration is an influencing factor, combined with other air pollutant factors and meteorological factors, and input into the CNN-LSTM model for fitting and prediction, and the predicted value of the CNN-LSTM combined model based on multi-feature fusion is output.

[0039] The above two prediction models correspond to one that can achieve PM 2.5 Methods for concentration prediction.

[0040] This embodiment first implements PM 2.5 The relevant data required for the concentration prediction method are based on the geographical environment and climatic conditions of the study area. The specific information of 12 air quality monitoring stations and 1 ground meteorological station is introduced at the same time. The air quality and meteorological data of each monitoring station are collected, and the meaning of its indicators is briefly introduced. Necessary descriptive statistics are performed, and then data preprocessing is carried out.

[0041] This embodiment takes a certain study area as an example, in which a site network including 12 air quality detection points and a ground meteorological station is deployed. Detailed information of each site is shown in Table 1 and Table 2.

[0042] Table 1 Detailed information of air quality monitoring sites

[0043]

[0044] Table 2 Detailed information of ground meteorological stations

[0045]

[0046] The data used in this example are the air pollutant concentration values ​​of 12 air quality monitoring stations in the study area every three hours and the meteorological factor data of 1 meteorological monitoring station every three hours from January 1, 2021 to September 30, 2023. Among them, air pollutants include PM 2.5 、PM 10 、SO 2 、NO 2 ,CO,O 3 The data comes from the national urban air quality real-time release platform of the China National Environmental Monitoring Center; meteorological factors include air temperature, dew point temperature, air pressure, wind direction, and wind speed. The data comes from the China Meteorological Data Network of the National Meteorological Information Center.

[0047] The data obtained totals 8024 items. The air quality data indicators are shown in Table 3, and the meteorological factor data indicators are shown in Table 4.

[0048] Table 3 Description of air quality data indicators

[0049]

[0050] Table 4 Description of meteorological factor data indicators

[0051]

[0052]

[0053] It should be noted that, for air quality data, this embodiment mainly takes PM 2.5 、PM 10 、SO 2 、NO 2 ,CO,O 3 These six real-time data indicators. The descriptive statistics of air quality data are shown in Table 5, and the descriptive statistics of meteorological factor data are shown in Table 6.

[0054] Table 5 Descriptive statistics of air quality data in Hangzhou

[0055]

[0056] Table 6 Descriptive statistics of meteorological factors in Hangzhou

[0057]

[0058] The above air quality data is preprocessed, such as Figure 1 This is a flowchart of data preprocessing. The data preprocessing stage involves the corresponding integration of time series of air quality records and meteorological factor data, the reasonable interpolation of missing data, and the standardized conversion of numerical values. The overall data set is segmented according to specific rules to serve the subsequent model training, verification and in-depth exploration.

[0059] Due to instrument failures at monitoring points, interruptions in data collection, and other unforeseen circumstances, data for some periods may not be successfully obtained, resulting in missing values ​​in the data set. In order to enhance the prediction performance of the model, it is essential to perform necessary processing steps before actual application.

[0060] The data of air pollutants and meteorological factors show significant spatiotemporal characteristics. If the missing values ​​are simply removed, the validity and reliability of the prediction results will be weakened. Therefore, this paper adopts a missing value filling strategy to construct a continuous and complete time series to improve data completeness. In addition to the missing of some fields in the data set of this paper, there are also missing whole lines with long time intervals of three days or more. At this time, conventional statistical filling methods (such as mean or median interpolation) may not be able to effectively capture the temporal dynamic changes of the data. In view of a large number of related studies, this embodiment finally chooses to use the K nearest neighbor algorithm (KNN) in machine learning as the missing value filling method, which uses the nearest station set of the target location to realize the reasonable speculation of the unknown value of the target based on the measured data of the adjacent stations.

[0061] The K-nearest neighbor algorithm is a non-parametric classification and regression method. In regression problems, it predicts the numerical output of new instances. Among them, k is a custom constant, usually a small positive integer. When k=1, it is the nearest neighbor algorithm (Nearest Neighbor), and when k>1, it is the k-nearest neighbor algorithm. This article defines the k value as 8. The main process of the KNN filling method used in this article is: take the records with missing values ​​in the data set as new instances, calculate its Euclidean distance with all instances in the training set, select the 8 closest instances, and calculate the average based on their corresponding eigenvalues ​​as the filling result of the missing value.

[0062] Eliminating the dimensions of each field in the data set helps the model to more accurately extract features between data, helps to accelerate the convergence of the model, and strengthens the prediction effect. Since the data used in this study are all dimensional, they need to be standardized before analysis and prediction. Commonly used data standardization methods include Min-max standardization and Z-score standardization.

[0063] The Min-max normalization method is a linear transformation of the original data and is calculated using the following formula.

[0064]

[0065] In the formula, represents the new data after standardization, X represents the original data, and X min is the minimum value, X max This method maps the data between [0,1] and is suitable for data with a certain boundary range.

[0066] The Z-Score standardization method is to standardize the mean and standard deviation of the original data. The processed data satisfies the mean of 0 and the standard deviation of 1, which is calculated by the following formula.

[0067]

[0068] Where, X Z is the standardized data, X is the original data, μ is the mean, and σ is the variance. This method is applicable to data with uncertainty boundaries and assumes that the data distribution roughly follows the normal distribution characteristics.

[0069] Considering the effective application range of the two normalization techniques, this example verifies whether the research data satisfies the normal distribution. 2.5 Taking the data as an example, the single-sample Kolmogorov-Smirnov test method was used to rigorously evaluate whether its distribution was normal. The KS test result of this site was statistic 0.070 and P value 1.485e -10 The P value is less than 0.05, rejecting PM 2.5 The null hypothesis that the concentration follows a normal distribution. Similarly, this paper conducted a normal distribution test on the air pollutant concentrations of the remaining 11 air quality monitoring stations and the meteorological factor data of one ground meteorological monitoring station, and none of them followed a normal distribution. Therefore, in this embodiment, the Min-max normalization method was selected to perform a normalization processing step on the studied data.

[0070] This embodiment provides a PM based on time series 2.5 Concentration prediction method, based on PM 2.5Starting from the dimension of the concentration series itself, a VMD-LSTM-GWO combined prediction model based on time series is built, with data from January 2021 to August 2023 as the training set and data from September 2023 as the test set. At the same time, this embodiment designs ablation experiments and comparative experiments to verify the feasibility, effectiveness and superiority of the proposed model.

[0071] The implementation of the VMD-LSTM-GWO combined prediction model mainly involves the use of Pycharm and Anaconda3 integrated development platforms, with the help of the deep learning architecture TensorFlow to build and implement the model, and the use of libraries such as Pandas and Numpy to perform preprocessing operations on the data, and the use of Matplotlib to complete the corresponding visualization tasks. The specific experimental environment is shown in Table 7 below.

[0072] Table 7 Experimental environment

[0073]

[0074] The PM based on time series 2.5 The concentration prediction method is based on the PM 2.5 The concentration value prediction model is implemented, namely the VMD-LSTM-GWO model. The model framework diagram is as follows Figure 2 As shown in the figure, the model prediction process is mainly divided into the following five parts:

[0075] Part I: Data preprocessing. Obtain the PM2.5 data from each air quality monitoring station in Hangzhou every three hours from January 1, 2021 to September 30, 2023. 2.5 The concentration value time series data is preprocessed and the KNN method is used to fill the missing values ​​to eliminate the impact of missing values ​​on the prediction results.

[0076] The second part: sequence decomposition. This part uses VMD decomposition (variational mode decomposition) to decompose the PM 2.5 The concentration value sequence is split and converted into a series of stable time series components, aiming to reveal the intrinsic characteristic information of the original sequence in multiple frequency dimensions, thereby overcoming the PM 2.5The prediction difficulties caused by the inherent non-stationary characteristics of the concentration time series. When implementing the decomposition step, the number of modes selected is defined by the parameter k, and the key regulatory factor for weighing the bandwidth of each mode is the penalty parameter α, which directly affects the spectrum coverage of each decomposed subsequence. In order for VMD to obtain more accurate decomposition results and for the overall prediction model to have higher prediction accuracy, this embodiment uses the envelope entropy in VMD as the objective function, and the minimum envelope entropy as the target value. The particle swarm optimization algorithm is used to optimize the two parameters k and α in VMD, and then the center frequency of each IMF output when different k values ​​are selected is combined to determine the final k and α values. The input of this part is the PM after preprocessing. 2.5 A sequence of concentration values, output as several eigenmode functions.

[0077] Empirical Mode Decomposition (EMD) is an intuitive and easy-to-use signal decomposition processing technology. Since its introduction, it has penetrated into many fields and has been widely used. EMD decomposes the original signal into a limited number of intrinsic mode functions (IMFs) and a residual component. These IMFs each represent an independent intrinsic fluctuation mode in the original sequence, and the residual component summarizes the overall trend of the original sequence over time. The decomposed signal can be expressed by the following formula.

[0078]

[0079] In the formula, u i (t) is the IMF component after decomposition; r(t) is the residual term after decomposition. The IMF generated by EMD processing must comply with two criteria: (1) the difference in the number of extreme points and zero-crossing points of the IMF component is ≤ 1; (2) at any given moment, the envelope constructed by averaging the adjacent maximum and minimum points should be 0.

[0080] However, the IMFs obtained by EMD decomposition are prone to the problem of modal aliasing. Modal aliasing can be intuitively described as the confusion and superposition of different modal features in the same analytical result. In this regard, the variational mode decomposition (VMD) technology has been proven to be able to effectively suppress the occurrence of such phenomena.

[0081] The VMD method avoids the repeated selection process of each component in the EMD and defines the intrinsic mode function as the amplitude-frequency modulation signal u k (t), which is expressed by the following formula.

[0082] u k (t) = A k (t)cos(φ k (t)).

[0083] In the formula, u k(t) is the instantaneous amplitude; u k (t) is the instantaneous frequency ω k (t) is φ k (t) first-order derivative. The constraint model of VDM is:

[0084]

[0085] Where f(t) is the original signal; u k = {u 1 ,...u K} are the K IMFs obtained after decomposition of f(t); {ω k}={ω 1 ,...ω K} represents the center frequency corresponding to each IMF obtained by decomposition; It is the partial derivative of t, and * is the convolution operation.

[0086] Introducing the Lagrange multiplier λ t and the penalty factor α turns this problem into an unconstrained problem:

[0087]

[0088] In the formula, δ(t) is the impulse function, and α can ensure the reconstruction accuracy of the signal in the presence of Gaussian noise. The saddle point of the Lagrange function, i.e. the optimal solution of formula (2-5), can be obtained by using the Alternating Direction Method of Multipliers (ADMM). Solution to the optimization problem, center frequency update, and operator λ n+1 Use the following formulas to update respectively.

[0089]

[0090]

[0091]

[0092] In the formula, Can be seen as Wiener filtering; is the center of gravity of the corresponding IMF power spectrum; τ is the noise tolerance.

[0093] The specific steps of the VMD method are as follows:

[0094] (1) Initialization λ 1 ,n=0;

[0095] (2) Update based on the solution to the optimization problem, the center frequency, and the operator update formula λ 1 ;

[0096] (3) Repeat the above steps, with n=n+1 in the iteration. When the following formula is satisfied, the loop ends.

[0097]

[0098] Where ε is the given judgment accuracy (ε>0).

[0099] In the specific application of VMD method to PM 2.5 When predicting concentration, VMD is applied to the pre-processed PM 2.5 Before decomposing the concentration data, it is necessary to determine the number of decomposed modal components k. The size of k affects the decomposition effect of the time series. If the k value is too small, mode aliasing will occur in the modal components. If the k value is too large, mode repetition or additional noise will occur. At the same time, in the process of using the VMD method, the penalty parameter α also has a significant effect on the final decomposition effect. When the value of this parameter is smaller, the frequency span of each IMF component will increase accordingly. On the contrary, as it increases, the frequency concentration of each component signal will increase, which is manifested as a decrease in bandwidth. Given the PM in the real world, 2.5 Concentration sequences often show highly complex and dynamically changing characteristics. It is not easy to accurately set the key parameters in VMD. Therefore, how to systematically find a suitable parameter combination for decomposing such sequences has become a major challenge for the effective use of VMD decomposition technology. The particle swarm algorithm in swarm intelligence optimization technology is favored for its excellent overall search performance. In this embodiment, the envelope entropy in the VMD decomposition process is selected as the optimization objective function, and the minimum value of the envelope entropy is used as the expected optimal solution. In order to achieve this goal, this embodiment uses the advantages of the particle swarm optimization algorithm to carry out concurrent adjustment and optimization of the two core parameters k and α involved in VMD, aiming to eliminate the subjective influence brought by the artificial setting of thresholds, and then automatically identify the parameter matching scheme that optimizes the VMD performance.

[0100] In a D-dimensional solution space scenario, suppose there is a particle swarm X=(X 1 ,X 2 ,...,X M ), where the position of any single particle i in its corresponding D-dimensional search region is defined as X i =(x i1 ,x i2 ,...,x iD ), which represents a possible answer to the optimization problem to be solved, and the current velocity of particle i is recorded as V i =(v i1 ,vi2 ,...,v iD ), and the individual local extreme value it reaches is recorded as P i =(p i1 ,p i2 ,...,p iD ), and the global optimal solution shared by the entire group is G i =(g 1 ,g 2 ,...,g D ), each particle cyclically updates its velocity vector and position based on its own local optimal information and the global optimal information shared by all particles, so as to continuously approach the optimal solution. This process is calculated by the following particle velocity and position update formula:

[0101]

[0102] Where w is the inertia weight, d = 1, 2, ..., D, i = 1, 2, ..., M, k is the current iteration number, c 1 and c 2 is the acceleration factor, and η is a random number between [0,1].

[0103] When using the particle swarm optimization algorithm to explore the key influencing parameters of the VMD algorithm, one of the key steps is to establish an evaluation index, namely the fitness function. Whenever a particle adjusts its position in the search space, it will re-evaluate its quality based on the function and make a position update decision based on the newly calculated fitness value. This paper selects envelope entropy as the function to measure fitness. After performing demodulation on the signal and obtaining the envelope components, the envelope information is converted into a probability distribution sequence p j , the entropy value derived from this can effectively reflect the sparse characteristics of the original signal. The envelope entropy E of the zero-mean signal x(j), j=1,2,...N p Use formula (4-2) to calculate.

[0104]

[0105] In the formula, p j is the normalized form of a(j); a(j) is the envelope signal obtained after the signal x(j) is decomposed. i The position here represents a set of potential influencing parameter combinations α and k configurations. The corresponding envelope entropy of all IMF components obtained by VMD decomposition under this position condition is calculated, and the minimum envelope entropy value is selected and defined as the local minimum entropy value, denoted as min L Based on this, the IMF component corresponding to the local minimum entropy value is determined. This component is considered to capture the richest PM under the current parameter combination. 2.5 The ideal component of the characteristic information of the concentration time series, but it should be noted that this is only the optimal component in the local range. In order to decompose the IMF component containing the richest characteristic information from the sequence, that is, to determine the global optimal component, this paper uses the local minimum entropy value as an adaptive indicator to measure the optimization potential of each component, and sets the goal of the entire optimization process to minimize the local minimum entropy value. The optimization steps of α and k in the VMD method are as follows:

[0106] (1) When applying the particle swarm optimization algorithm, all the parameters involved are initialized and reasonably set, and the fitness function used to evaluate the quality of candidate solutions during the optimization iteration process is clearly defined.

[0107] (2) Initialize the particle population to influence the parameter combination [α, k] as the position of the particle, and randomly generate a series of such combinations as the initial position of the particle. At the same time, the initial moving speed of each particle is randomly set to ensure the diversity of the population in the search space.

[0108] (3) Perform VMD operation on the signal under different particle position conditions and calculate the corresponding fitness value min of each particle position L

[0109] (4) By comparing the fitness values ​​of each individual, the local extreme value state maintained by each is updated according to the order of their merits; at the same time, the global extreme value shared by the entire particle swarm is synchronously refreshed to reflect the current optimal performance level.

[0110] (5) Use the particle velocity and position update formula to update the particle velocity and position.

[0111] (6) Repeat the iteration cycle and go to step (3) until the number of iterations reaches the maximum set value, then output the optimal fitness value and the position of the particle.

[0112] This embodiment sets the maximum k value of the particle swarm optimization parameter space to a constant value of 16. This value reflects that in actual application scenarios, the upper limit of the intrinsic frequency components of most signals usually does not exceed this threshold. At the same time, when selecting the bandwidth constraint strength, this embodiment selects a generally applicable range of [100,3100] to ensure the effectiveness and rationality of the optimization process. After optimization, VMD will study the PM 2.5The concentration value sequence is decomposed into 10 IMFs with a center frequency bandwidth constraint of 1000. In order to further verify the rationality of the VMD decomposition results, this embodiment verifies the optimal value of k by observing the center frequency. First, the center frequency bandwidth constraint is kept unchanged at 1000, and the original data is decomposed with different k values. Then, the center frequencies of different layers are calculated. 8 lists the center frequency values ​​of each IMF when k is 2 to 12, and compares the center frequencies of the last layer of modal components under different k values. When it is observed that the center frequencies of the tail modal layers are basically consistent, k can be regarded as a potential optimal value point. It can be seen from Table 8 that when k is greater than or equal to 10, the center frequency of the last layer is relatively stable around 0.46, so k = 10 is the optimal value, which is consistent with the optimal parameter value obtained by particle swarm optimization. Therefore, this paper sets the parameter values ​​of the VMD algorithm to α = 1000, k = 10, that is, the PM value of the study area is calculated every three hours from January 1, 2021 to September 30, 2023. 2.5 The concentration values ​​are decomposed into 10 IMFs with a center frequency bandwidth constraint of 1000.

[0113] In order to achieve a more intuitive display of the modal features obtained after VMD decomposition, this embodiment adopts visualization technology to present the original PM2.5 sequence and its decomposed sub-modes together. 2.5 Time series graph Figure 3 As shown, the original PM 2.5 The time series contains a large number of large-scale fluctuations and subtle peak changes, and the structure is complex and changeable, which increases the difficulty of prediction. The VMD decomposition results are as follows: Figure 4 As shown, the VMD algorithm converts highly complex PM 2.5 The concentration data is decomposed into multiple periodic and regular modal components, each of which contains different frequency information, which can better extract and learn the information contained in the sequence. The sub-modes after VMD decomposition are arranged in order from low frequency to high frequency, showing distinct frequency domain characteristics, which proves the effectiveness of VMD decomposition and greatly reduces PM 2.5 The difficulty of extracting internal features improves prediction accuracy.

[0114] Table 8 Central frequencies of PM2.5 concentration series at station 1223A at different k values

[0115]

[0116] Part 3: Prediction model construction. First, the data of several IMF sequences output by VMD are normalized, that is, the original data is mapped to numbers in the interval [0,1]. This operation can speed up the gradient descent to find the optimal solution and accelerate the convergence speed. Further, the normalized IMF sequences are input into the LSTM network model respectively, and the predicted values ​​of each IMF are output after training and fitting. Among them, the network model structure of LSTM is shown in Table 9. It can be seen that the LSTM network structure used in this prediction framework is not complicated, and a satisfactory prediction accuracy can be achieved without many network layers. This is because VMD converts PM 2.5 The sequences are several intrinsic mode functions, all of which are stationary sequences. They can be understood as extracting features at different frequencies. Therefore, there is no need for a complex LSTM model to predict each IMF, which can save resources and improve prediction efficiency.

[0117] Table 9 LSTM network structure in VMD-LSTM-GWO model

[0118]

[0119] Different from the standard RNN, the LSTM network structure uses a unique LSTM layer to replace the conventional hidden layer, and introduces a mechanism that can transmit state information across multiple time intervals, save the information of data points, and effectively retain and integrate the long-term dependency information of data sequences, thus overcoming the limitations of traditional RNN in this regard. The internal structure of LSTM neurons is as follows: Figure 5 shown.

[0120] The uniqueness of LSTM neural network lies in its ability to effectively integrate short-term transient information and long-term dependent information in time series data, thanks to its use of a "gating" mechanism, which implements selective retention and update operations of information through a precisely designed gate structure. In fact, it is these internal gating units that are responsible for the selection and accumulation of input information streams, thereby ensuring that the network can properly process and permanently maintain key information in the time series.

[0121] (1) Forget Gate

[0122] The forget gate plays a key role in regulating the output information in LSTM. Its core function is to determine which information components should be removed from the neuron state. The input of this gating mechanism comes from the output results of the previous moment and the information currently input to the model. After the Sigmoid activation function is operated, a weight coefficient between 0 and 1 is generated. This coefficient essentially reflects the degree of information retention. Then, by performing an element-by-element multiplication operation with the stored memory cell information, effective screening and filtering of past information is achieved. This process is calculated using the following formula.

[0123] ft =σ(w f ·[h t-1 ,x t ]+b f ).

[0124] In the formula, h t-1 is the output data of the previous moment; x t is the input data at the current moment; [h t-1 ,x t ] means concatenating two vectors into a new vector; b f is the bias term; w f is the connection weight of the forget gate; σ is the activation function.

[0125] (2) Input Gate

[0126] The input gate plays a key regulatory role in LSTM, which is to decide what new information should be integrated into the state of the neuron. To achieve this function, the input gate receives two inputs: on the one hand, it receives the output i after being processed by the Sigmoid activation function. t On the other hand, it also receives the output after the tanh activation function These two information flows work together to update the contents of the state unit, ensuring that the network can efficiently process and remember key information. This process is calculated using the following formula.

[0127] i t =σ(w f ·[h t-1 ,x t ]+b i ),

[0128]

[0129] (3) Status Unit

[0130] The state unit plays a key role in refreshing memory in LSTM. The internal architecture of the unit can be divided into two major operating blocks: the first block is responsible for fusing the output information of the previous moment with the new data stream currently input to the model; and the other block is focused on refreshing the content of the memory cell. The final state c t It is the comprehensive result obtained through the collaborative operation of the forget gate and the input gate, and is calculated using the following formula.

[0131]

[0132] In the formula, f t *c t-1 Indicates the previous state information c t-1Eliminate some of the information that needs to be forgotten, and add the output activated by tanh The final state c is obtained t .

[0133] (4) Output gate

[0134] The output gate plays a key role in determining the next hidden state value in LSTM. Specifically, the value in the hidden state is not determined by a single factor, but by the hidden state h activated by Sigmoid in the previous process. t-1 The input information x at this moment t The result of the joint impact t , these two are combined with the hidden state c processed by the tanh activation function t Multiply them together to get the final hidden state value. This process is calculated using the following formula.

[0135] o t =σ(w o ·[h t-1 ,x t ]+b o ).

[0136] The output information is calculated using the following formula.

[0137] h t =o t *tanh(c t ).

[0138] LSTM uses a unique gating mechanism to carefully screen and regulate the accumulation and discard of historical information in time series, thereby avoiding the gradient vanishing problem while retaining key historical clues, and thus providing a powerful solution for revealing the intrinsic relationship between long-term series data. Compared with traditional recurrent neural networks, the LSTM architecture constructed by chain loops is better at dealing with time series analysis tasks, especially in dealing with long-term dependencies. Therefore, within this theoretical framework, using LSTM to predict PMs rich in deep historical information is 2.5 Concentration values ​​are expected to achieve more accurate and reliable prediction results.

[0139] Part 4: Model weight optimization. If we directly add up the forecast results of each IMF to get PM 2.5 The prediction results of the concentration value sequence are given by 1. In fact, not every IMF has an ideal prediction accuracy. If the same weight is given to each IMF, the best prediction result of the overall model may not be obtained. Therefore, this paper uses the Gray Wolf Optimization Algorithm to compare the model prediction value with PM2.5. 2.5The mean square error between the actual concentration values ​​is used as the objective function, and the minimum mean square error is used as the target value. The weight values ​​of each IMF are reconstructed and optimized to improve the prediction accuracy of the model. The final PM value of the model is obtained by multiplying the predicted sequence values ​​of each IMF with the weight values ​​of each IMF sequence obtained by the gray wolf optimization algorithm. 2.5 Predicted concentration.

[0140] Part V: Model Evaluation. Comparison of model prediction results with PM 2.5 The true value of concentration and calculation error are combined with the evaluation indicators of each model to comprehensively evaluate the prediction performance of the model.

[0141] Designed to facilitate comparison with other PM prediction methods 2.5 To compare the prediction effects of concentration values, this embodiment experiment uses three performance evaluation indicators, namely Mean Absolute Percentage Error (MAPE), Mean Absolute Error (MAE) and Root Mean Square Error (RMSE), to evaluate the prediction effect of the model. The smaller the three errors, the higher the prediction accuracy of the model, which are calculated by the following formulas respectively.

[0142]

[0143]

[0144]

[0145] In the formula, is the i-th predicted value; y i is the i-th true value; n is the length of the time series

[0146] PM 2.5 After VMD decomposition of the concentration time series, the IMFs obtained were input into the LSTM network for fitting and prediction. It is noted that there is no unified standard parameter configuration scheme for the LSTM model architecture, so in practice, the research results of several authoritative scholars were referred to. [85~86]The proposed model parameter setting strategy, and the LSTM network hyperparameters were adjusted and set in a targeted manner based on multiple rounds of experimental measurements. The final model network has 3 layers. The first layer is the LSTM layer, and the number of hidden neurons is set to 64; the second layer is the Dropout layer, and the proportion of randomly discarded neurons is set to 0.1; the third layer is a fully connected layer, and the output dimension is 1. The learning rate is set to 0.001; the activation function is set to ReLU; the optimization algorithm uses the Adam algorithm to optimize the parameters; the number of iterative training is set to 150 times; the time window is set to 60, that is, the first 60 historical data are used to predict the 61st value; Loss is used to quantify the degree of deviation between the model prediction result and the actual target value, and then the model parameters are continuously guided to adjust in the direction of reducing Loss to achieve effective training of the model. The Loss function used in the experiment in this paper is the mean square error function, and its expression is:

[0147]

[0148] In the formula, y i is the i-th true value, is the i-th predicted value, and n is the length of the time series.

[0149] After setting the LSTM model parameters, each IMF sequence is input into the LSTM model for fitting and prediction. The last 240 data are used as the test set and the previous data are used as the training set. The prediction results are visualized as follows: Figure 6 As shown in Table 10, the yellow solid line is the true value of the IMF component and the red solid line is the model predicted value of the IMF component. The forecast errors of each IMF are shown in Table 10.

[0150] From Table 10 and Figure 6 It can be seen that the LSTM model has a good overall prediction effect on each IMF, among which the prediction accuracy of IMF1 is as high as 99.45%, the prediction accuracy of IMF2 and IMF5 are both above 90%, and the prediction accuracy of IMF3, IMF4, IMF6 and IMF7 are all above 80%. However, the prediction effect of IMF8, IMF9 and IMF10 is not good, with prediction accuracy of 33.24%, 1.02% and 30.80% respectively. This is because these three components are high-frequency components, fluctuate violently, have many peaks and troughs, contain more information, and bring difficulties to fitting; while the IMF1-IMF7 components are relatively low in frequency, smaller in fluctuation, and contain less fluctuation information. Therefore, the prediction results of IMF1-IMF7 components are more accurate than those of IMF8-IMF10.

[0151] Table 10 Forecast errors of various IMF series

[0152]

[0153] According to the forecast results of each IMF, the LSTM model does not have an ideal forecast effect on IMF8, IMF9 and IMF10. If we directly add up the forecast results of each IMF, we can get PM 2.5 The final predicted value of the concentration time series cannot obtain the best prediction result of the model. At this time, the weight assigned to each IMF prediction result is 1. In order to improve the prediction accuracy, this embodiment uses the gray wolf optimization algorithm to reconstruct and optimize the weight value of each IMF, and takes the mean square error between the model prediction value and the true value of the sequence obtained by multiplying each subsequence with the weight as the objective function, and takes the minimum value of the mean square error as the target value to recalculate the weight of each component. For the parameter setting of the gray wolf optimization algorithm, referring to the relevant research results and conducting multiple rounds of experimental measurements, this paper sets the population number to 50; the search space dimension is 10, because there are 10 IMF components; the initialization weight upper limit is 1.5, the initialization weight lower limit is 0.5; the number of iterations is 100 times. The optimization algorithm outputs the weight value of each IMF component, as shown in Table 11, the predicted value of each IMF sequence is multiplied and added with its corresponding weight calculated by the gray wolf optimization algorithm, that is, the final predicted value of the VMD-LSTM-GWO model is obtained.

[0154] Table 11 The weights of each IMF component output by the Grey Wolf optimization algorithm

[0155]

[0156] The VMD-LSTM-GWO model calculates the PM every three hours at station 1223A in the study area in September 2023. 2.5 The comparison between the predicted concentration value and the actual value is shown in the figure below. Figure 7 As shown, the red solid line is PM 2.5 The model predicted value of the concentration value, the blue solid line is PM 2.5 The actual value of the concentration. From the comparison chart, it can be seen that the overall prediction effect of the model is good. After calculation, the MAE is 0.57, the RMSE is 0.75, and the MAPE is 3.75%. It can be seen that the VMD-LSTM-GWO model proposed in this embodiment has good prediction performance and the prediction accuracy reaches 96.25%.

[0157] To further verify the VMD-LSTM-GWO model for PM in the study area 2.5 The universality of concentration value prediction, this example applies the model to the PM concentration of the other 11 air quality monitoring stations in the study area. 2.5 The concentration values ​​are predicted and the PM values ​​of each air quality monitoring station in the study area are 2.5The concentration value prediction evaluation indicators are shown in Table 12. As can be seen from Table 12, the MAPE of the VMD-LSTM-GWO model for the 12 air quality monitoring stations is within 10%, that is, the prediction accuracy of the model is above 90%, and the prediction performance is good. 2.5 The prediction of concentration value has certain universality.

[0158] Table 12 Evaluation indicators of PM2.5 concentration values ​​at each air quality monitoring station

[0159]

[0160] In order to further verify the feasibility and effectiveness of the VMD-LSTM-GWO prediction model proposed in this embodiment, this embodiment designs and conducts an ablation experiment. In deep learning, an ablation experiment refers to systematically cutting or adjusting certain parts or parameters of a neural network to evaluate their impact on model performance. This experiment can help understand the contribution of different components in the model to the overall performance and reveal the key characteristics of the network. This embodiment designs two different models to compare with the proposed VMD-LSTM-GWO model. The comparison results are shown in Table 13, showing the prediction errors of the prediction models lacking different parts.

[0161] Table 13 Comparison of model prediction errors

[0162]

[0163] To ensure the comparability of the prediction results, the LSTM model uses the model framework parameters introduced in Section 4.4.1 to analyze the PM 2.5 The concentration data is fitted and predicted. It can be seen from Table 13 that the prediction effect is not ideal at this time, with a MAPE of 29.64%, that is, the prediction accuracy of the model is 70.36%, indicating that there is room for further optimization of the model.

[0164] The VMD-LSTM model adds a decomposition operation to the original time series based on LSTM, and applies the VMD method and steps introduced above to PM 2.5The concentration value sequence is decomposed to obtain 10 IMF components, which are input into the LSTM network for fitting and prediction. The results of each component are directly added to obtain the prediction value of the VMD-LSTM model. Table 13 shows that the prediction result at this time is significantly improved compared with LSTM. The MAPE at this time is 4.91%, which is 24.73 percentage points lower than the LSTM average absolute percentage error, which means that the prediction accuracy of the VMD-LSTM model is 24.73 percentage points higher than that of LSTM; at the same time, MAE drops from 4.91 to 0.81, and RMSE drops from 7.01 to 1.09. This shows that the original PM 2.5 It is necessary to perform VMD decomposition on the concentration value time series, extract the features of different frequencies, output a set of stable sub-modes and predict them separately, which greatly reduces the fitting difficulty of LSTM and can greatly improve the prediction accuracy.

[0165] The VMD-LSTM-GWO model is the prediction model proposed in this paper. On the basis of VMD-LSTM, the weights of the prediction values ​​of each IMF component are further reconstructed and optimized. The weights of the prediction values ​​of each IMF component in VMD-LSTM are all 1, and the VMD-LSTM-GWO model applies the WGO optimization algorithm. The mean square error between the model prediction value and the true value of the sequence is used as the optimization function. The weights of the prediction values ​​of each IMF component are recalculated, and the prediction values ​​of each IMF component are multiplied by the weight and then added to obtain the final prediction value of the VMD-LSTM-GWO model. Table 13 shows that the prediction results at this time are improved compared with the other two models. The MAPE at this time is 3.75%, which is 1.16 percentage points lower than the average absolute percentage error of VMD-LSTM and 25.89 percentage points lower than LSTM. At the same time, MAE is reduced from 0.81 to 0.57, a decrease of 0.24, and a decrease of 4.34 compared with LSTM. RMSE is reduced from 1.09 to 0.75, a decrease of 0.34, and a decrease of 6.26 compared with VMD-LSTM. This shows that the application of the GWO optimization algorithm to reconstruct the weights of the prediction results of each IMF component has contributed to the improvement of the prediction accuracy. By continuously reducing the mean square error between the model prediction value and the true value of the sequence, the weight ratio of each subsequence is continuously adjusted and optimized, so as to achieve the effect of improving the prediction accuracy. At this time, the prediction accuracy of the model reaches 96.25%.

[0166] In order to more intuitively demonstrate the prediction performance of each model in the ablation experiment, this paper draws the tracking curves of the test set at different times, that is, the prediction performance of each model for PM at station 1223A in the study area in September 2023. 2.5 Concentration predictions, such as Figure 8 As shown, the black solid line is PM 2.5The real value of the concentration sequence, the red solid line is the prediction result of the LSTM model, the purple solid line is the prediction result of the VMD-LSTM model, and the cyan solid line is the prediction result of the VMD-LSTM-GWO model. By observing the comparison chart, it can be clearly found that the prediction result of the LSTM model for the sequence is not ideal. Although it can roughly predict the basic volatility and trend of the sequence as a whole, there is a certain lag, and the prediction is not good for data with large fluctuations. For example, in the data interval from No. 140 to No. 150 in the sequence, the model has no good prediction for concentrations exceeding 40μg / m 3 The data prediction error is large. Similarly, the same problem exists for the data near No. 200, indicating that LSTM has not fully learned the data fluctuations and trends, and the prediction accuracy is not high. Both the VMD-LSTM and VMD-LSTM-GWO models show excellent prediction accuracy. The prediction result curves of the two models basically coincide with the true value curve. By zooming in on the local area, that is, the part marked by the black box in the figure, it can be observed that the prediction result of VMD-LSTM-GWO is closer to the true value, indicating that the weights of each IMF component optimized by GWO can effectively improve the prediction performance of the model.

[0167] In order to further verify the superiority of the prediction model proposed in this embodiment, the VMD-LSTM-GWO prediction model is compared with three currently commonly used prediction models, namely ARIMA, GRU and EMD-LSTM. In the experiment, the same training set and test set are used to train each model, and the prediction error is shown in Table 14. It can be seen from Table 14 that the prediction effect of the ARIMA model based on traditional statistics is the least ideal, with a prediction accuracy of only 65.88%. This may be because the model can only learn the linear relationship in the sequence, but cannot capture the nonlinear relationship in the data well; the prediction effect of the GRU model based on deep learning is better than ARIMA, because the neural network can better learn the nonlinear features in the sequence, but its prediction accuracy is 71.43%, and there is still room for further optimization; the EMD-LSTM model is a relatively popular model in the field of time series prediction recently. The model first uses EMD to decompose the sequence to obtain Several IMF components are then input into the LSTM model for fitting training. Finally, the IMF prediction components are added together to get the final output of the model. The prediction accuracy of the model is 93.02%, which has obvious advantages over ARIMA and GRU. It can be seen that decomposing the original sequence into several components at different frequencies has a great contribution to improving the prediction accuracy. However, compared with the VMD-LSTM model designed in the ablation experiment, the prediction accuracy of the EMD-LSTM model is still 2.07 percentage points lower. This is because EMD is prone to modal aliasing, and VMD has improved this disadvantage well, further improving the prediction accuracy.

[0168] Table 14 Comparison of model prediction errors

[0169]

[0170] Combining Table 13 and Table 14, it can be seen that the VMD-LSTM-GWO prediction model proposed in this paper has better prediction performance, higher prediction accuracy, and can fit PM well. 2.5 The changing characteristics of the concentration series are analyzed and short-term predictions are made.

[0171] The embodiment of the present invention also provides a PM based on multi-feature fusion 2.5 Concentration value prediction method, from PM 2.5 From the perspective of multi-dimensional influencing factors of concentration, a CNN-LSTM combined prediction model based on multi-feature fusion is built, with data from January 2021 to August 2023 as the training set and data from September 2023 as the test set. At the same time, this embodiment designs ablation experiments and comparative experiments to verify the feasibility, effectiveness and superiority of the proposed model.

[0172] The implementation of the CNN-LSTM combined prediction model mainly involves the use of Pycharm and Anaconda3 integrated development platforms, the deep learning architecture Pytorch to build and implement the model, and the use of libraries such as Pandas and Numpy to perform preprocessing operations on the data. At the same time, Matplotlib is used to complete the corresponding visualization tasks. The specific experimental environment is shown in Table 15.

[0173] Table 15 Experimental environment

[0174]

[0175] In order to accurately describe the degree of interdependence of individuals in the spatial dimension, a spatial weight matrix is ​​introduced as a measurement tool to provide key spatial information for subsequent numerical prediction and machine learning models. The spatial weight matrix is ​​often defined as an n×n binary symmetric matrix. Its core function is to characterize the intrinsic spatial connection between n different entity geographic locations, which is calculated by the following formula:

[0176]

[0177] In the formula, w ij Represents the distance between position i and position j in space, which is used to represent the spatial correlation between position i and position j.

[0178] A basic spatial weight matrix expression is a binary spatial weight matrix, which is expressed as follows.

[0179]

[0180] Where, d ij represents the distance between position i and position j, δ represents the distance threshold, when d ij When <δ, position i and position j have spatial correlation, w ij =1; when d ij When ≥δ, there is no correlation between position i and position j, w ij =0.

[0181] Although the binary spatial weight matrix can be used to determine whether there is an association between the spatial locations of each entity, it does not yet have the ability to reveal the degree of association between the locations. In order to more accurately measure the relative importance of spatial distances, the weight definition method shown in the following formula is generally used.

[0182]

[0183] Where, d ij represents the distance between position i and position j, and δ represents the distance threshold.

[0184] Existing experience in spatial weight research has revealed that the size of the weight will show a rapid downward trend as the spatial distance increases, that is, there is a close exponential relationship between the weight and the inverse of the spatial distance between two entities. In practical applications, the diagonal elements of the spatial weight matrix are generally set to 0 using the following formula.

[0185]

[0186] As an excellent correlation quantification method, the maximum information coefficient (MIC) is widely favored in the field of correlation analysis. Its characteristic is that it can map values ​​to a unified [0,1] closed interval. If the MIC value is high, it means that there is a close correlation between the data variables under investigation, that is, there is a strong correlation. On the contrary, when the MIC value is equal to 0, it strongly implies that the data variables under investigation are independent or lack substantial correlation. Given two variables X = {x i ,i=1,...,n},Y={y i ,i=1,...,n}, and use the following formula to define MIC.

[0187]

[0188]

[0189] Where P(x,y),P(x),P(y) are the joint and marginal probability densities of x and y. I(X,Y) is the mutual information of X and Y. x ·n yare the number of divisions on the x-axis and y-axis respectively; B(n) is the upper limit of the number of divided grids.

[0190] PM at each air quality monitoring station 2.5 The historical concentration data were paired and considered as two groups of variables, and then the PM concentration of each station was calculated according to the formula. 2.5 The MIC between the concentration data is then directly integrated into the constructed spatial weight matrix to act as each element in the matrix. This process is expressed by the following formula.

[0191]

[0192] In the formula, r ij It represents the spatial correlation weight between air quality monitoring site i and site j, which is the MIC value in this embodiment.

[0193] PM collected at each air quality monitoring point 2.5 Concentration time series data are first collected to form a complete PM 2.5 Concentration data set, denoted as X(t)=(X 1 (t),X 2 (t),...,X s (t)) T Next, the MIC between different sites was calculated and used as the basic element to construct the PM 2.5 The weight matrix W for the spatial dependence of concentration s×s Then, use the weight matrix W s×s With PM 2.5 The concentration dataset X(t) is subjected to matrix multiplication operation, aiming to transform the abstract spatial features into quantifiable indicators and generate PM 2.5 The spatial characteristics of the concentration series are calculated using the following formula.

[0194]

[0195] The spatial feature X to be constructed Space (t) Integrate it into the original data set to facilitate the construction of a comprehensive set of input variables for subsequent prediction tasks, thereby enriching and improving the data input level of the prediction model.

[0196] For the existing PM 2.5 In the concentration prediction research, the regional effect of pollutants was not fully considered, and more influencing factors were ignored to be incorporated into the prediction model. This embodiment proposes a PM 2.5 The concentration prediction model uses PM 2.5The spatial characteristics of concentration changes, air pollutant factors and meteorological factors are taken into account at the same time. The CNN-LSTM model framework based on multi-feature fusion is as follows: Fig. 9 shown.

[0197] The prediction process of the CNN-LSTM model based on multi-feature fusion is mainly divided into the following four parts:

[0198] Part I: Data preprocessing. Obtain air pollutant concentration data from each air quality monitoring station every three hours and meteorological factor data from each meteorological data monitoring station every three hours from January 1, 2021 to September 30, 2023. Air pollutants include: PM 2.5 、PM 10 、SO 2 、NO 2 , CO and O 3 ,Meteorological factors include: air temperature, dew point temperature, air pressure, wind direction and wind speed.,The data is preprocessed and the KNN method is used to fill the missing values ​​to eliminate the,effect of missing values ​​on the prediction results.

[0199] Part 2: Constructing spatial characteristic factors. According to the PM introduced above 2.5 The spatial characteristics of the concentration series are generated to construct the PM 2.5 The spatial characteristic factors of concentration are first calculated. 2.5 The MIC coefficients between the concentrations are used as elements to construct a spatial weight matrix, which is multiplied by the matrix with the original data set to extract the PM 2.5 Spatial characteristic factor of concentration.

[0200] Part 3: Prediction model construction. First, the spatial feature factors, air pollutant data and meteorological factor data obtained in the second part are normalized, that is, the original data is mapped to numbers in the interval [0,1]. This operation can speed up the gradient descent to find the optimal solution. Further, the normalized data is input into the CNN-LSTM network model, and the PM2.5 concentration prediction value is output after training and fitting. Among them, the CNN-LSTM network structure is shown in Table 16. First, there are two 2D convolution layers, which are used to deeply extract features from the data; because the convolution operation of the pooling process will cause the loss of the location information of the features, the convolution process in this paper does not use the pooling layer; a Dropout layer is added after the 2D convolution layer to prevent overfitting; followed by two LSTM layers to learn the historical information of the data and improve the fitting accuracy; finally, two fully connected layers are connected to output the prediction results of the final model.

[0201] Table 16 CNN-LSTM network model structure

[0202]

[0203]

[0204] Part 4: Model evaluation. Comparison of model prediction results with PM 2.5 The true value of concentration and calculation error are combined with the evaluation indicators of each model to comprehensively evaluate the prediction performance of the model.

[0205] The experiment in this embodiment continues to take the data of the air quality monitoring station and meteorological monitoring station in the study area 1223A as an example, and uses three performance evaluation indicators, namely, mean absolute percentage error, mean absolute error and root mean square error, to evaluate the prediction effect of the model.

[0206] According to the steps of constructing spatial characteristic factors described above, the PM 2.5 The MIC value of the concentration is plotted and a heat map is drawn as Fig.10 shown.

[0207] The spatial weight matrix constructed based on this is as follows:

[0208]

[0209] The matrix is ​​compared with the PM 2.5 By multiplying the concentration value sequence, we can get the spatial characteristic factor of each site.

[0210] The constructed spatial characteristic factors are used as PM 2.5 Spatial characteristics of concentration, and PM 2.5 Other air pollutants and meteorological factors are used as PM 2.5 The characteristic variables include PM 2.5 、PM 10 、SO 2 、NO 2 ,CO,O 3 , air temperature, dew point temperature, air pressure, wind direction and wind speed as inputs to the model CNN-LSTM.

[0211] At present, there is no fixed parameter setting method for CNN, LSTM and their combined CNN-LSTM network. The network hyperparameters are set according to multiple rounds of experimental measurements. The final model network has 7 layers. The 1st and 2nd layers are two-dimensional convolution layers with a convolution kernel size of 3×3; the 3rd layer is a Dropout layer, and the ratio of randomly discarded neurons is set to 0.1; the 4th layer is an LSTM layer, and the number of hidden neurons is set to 128; the 5th layer is an LSTM layer, and the number of hidden neurons is set to 64; the 6th and 7th layers are fully connected layers, and the final output dimension is 1. The learning rate is set to 0.001; the activation function is set to ReLU; the optimization algorithm uses the Adam algorithm to optimize the parameters; the number of iterative training is set to 300 times; the time window is set to 10, that is, the first 10 historical feature variable data are used to predict the label value of the 11th item, that is, PM 2.5 Concentration value; The Loss function used in the experiment of this embodiment is the mean square difference function.

[0212] After setting the parameters of the CNN-LSTM model, the feature variables are input into the model for fitting and prediction. The data of September 2023 in the study area is used as the test set, and the data from January 2021 to August 2023 is used as the training set. The prediction results are visualized as follows: Fig.11 As shown in the figure, the overall prediction effect of the model is good. After calculation, the MAE is 2.19, the RMSE is 2.84, and the MAPE is 12.79%. Fig.11 It can be observed that the model has good 2.5 The fitting of concentration fluctuation is still insufficient. For example, there is a large gap between the predicted value and the true value for data No. 50 to 100 in the figure. At the same time, for the prediction of data No. 200 to 230, it can be found that the prediction error of the model for the local maximum value is large. But overall, the model can basically predict the trend and most of the fluctuations of the sequence. It can be seen that the CNN-LSTM model based on multi-feature fusion proposed in this embodiment has good prediction performance, with a prediction accuracy of 87.21%.

[0213] To further verify the CNN-LSTM model based on multi-feature fusion for PM 2.5 The universality of concentration value prediction, this example applies the model to the PM 2.5 The concentration value is predicted, and the PM 2.5 The concentration value prediction evaluation indicators are shown in Table 17. It can be seen from Table 17 that the average absolute percentage error of the CNN-LSTM model based on multi-feature fusion for the 12 air quality monitoring stations in Hangzhou is within 15%, that is, the prediction accuracy of the model is above 85%, and the prediction performance is good. 2.5 The prediction of concentration value has certain universality.

[0214] Table 17 PM at each air quality monitoring station 2.5 Concentration value evaluation index

[0215]

[0216] In order to further verify the feasibility and effectiveness of the CNN-LSTM prediction model based on multi-feature fusion proposed in this embodiment, this embodiment designs and conducts ablation experiments. This embodiment designs multiple prediction models based on three different basic models of CNN, LSTM and CNN-LSTM and changes the input variables to compare with the proposed model. To ensure the comparability of the prediction results, all model parameters are applied to the PM of the 1223A air quality monitoring station in the study area using the model framework and parameters introduced above. 2.5 The concentration data were fitted and predicted.

[0217] First, CNN is used as the basic model. This is equivalent to removing the LSTM network part and observing the fitting prediction of CNN for different input variables. The comparison results are shown in Table 18, which shows the average prediction error of the CNN model. As can be seen from Table 18, the prediction accuracy of CNN based on spatial feature factors is 83.53%, the prediction accuracy of CNN based on air pollutants is 73.21%, the prediction accuracy of CNN based on meteorological factors is 52.87%, and the prediction accuracy of CNN based on multi-feature fusion is 85.30%. It can be found that when air pollutants and meteorological factors are used as input features respectively, the prediction accuracy of CNN is not ideal, especially the prediction accuracy of CNN based on meteorological factors is only 53.87%. This may be because PM2.5 cannot be extracted well through these features. 2.5 When the spatial characteristic factor is used as the input feature, the spatial characteristic factor is connected with the PM 2.5 The concentration value is calculated, which can be indirectly understood as based on PM 2.5 The spatiotemporal characteristics of concentrations. At this time, the accuracy of the model is improved compared with that based on air pollutants or meteorological factors, indicating that the CNN network can better extract PM 2.5 The spatiotemporal characteristics of PM show better prediction performance; when all the characteristics are input into the model at the same time, the prediction accuracy of the model reaches the maximum value. 2.5 When considering the spatiotemporal characteristics of concentration, the influence of other air pollutants and meteorological factors, the CNN network can learn more useful information between the data, thereby improving the prediction accuracy of the model. 2.5 The predicted results of the concentration and the true value are plotted as a line comparison chart as shown in Fig.12 As shown. Fig.12It can be seen that the CNN based on meteorological factors only learns the weak trends and fluctuations in the data, and the error with the true value is the largest; and at the 203rd data, it can be observed that the CNN prediction value based on air pollutants is about twice the true value, and the error is very large. At the same time, a careful observation of other predicted values ​​of the model is mostly greater than the true value, indicating that the prediction results of the model are too large and have a large error; other models can basically predict PM 2.5 The variation pattern of concentration.

[0218] Table 18 Comparison of CNN prediction errors for different input features

[0219]

[0220] Secondly, taking LSTM as the basic model, it is equivalent to removing the CNN network part and observing the fitting prediction of LSTM for different input variables separately. The prediction error comparison results are shown in Table 19. The prediction accuracy of LSTM based on spatial feature factors is 83.61%, the prediction accuracy of LSTM based on air pollutants is 60.97%, the prediction accuracy of LSTM based on meteorological factors is 47.32%, and the prediction accuracy of LSTM based on multi-feature fusion is 80.17%. It can be found that similar to the prediction results of the CNN model, when air pollutants and meteorological factors are used as input features respectively, the prediction accuracy of the LSTM model is not ideal, and except for the LSTM based on spatial feature factors, the prediction accuracy of other models is lower than that of the CNN model. This is because CNN has a stronger ability to extract data features than LSTM and can learn more information in the data; LSTM mainly learns the historical information of the data, but its ability to extract features is not as good as CNN, and the input of LSTM based on spatial feature factors is PM 2.5 LSTM can fit the historical spatiotemporal characteristics of the concentration series very well, so only this model has a higher prediction accuracy than CNN, and it has the highest accuracy among the four LSTM models. The prediction accuracy of other models is lower than that of CNN network. 2.5 The predicted results of the concentration and the true value are plotted as a line comparison chart as shown in Fig.13 ,Through observation, it can be seen that the LSTM based on meteorological factors has basically not learned the trend of the data, with only some weak fluctuations, and the prediction error is the largest. The second largest error is the LSTM based on air pollutants. The other models can basically predict PM 2.5 The variation pattern of concentration.

[0221] Table 19 Comparison of LSTM prediction errors for different input features

[0222]

[0223] Finally, taking CNN-LSTM as the basic model, the fitting prediction of the model based on different input variables is observed, and the prediction error comparison results are shown in Table 20. The prediction accuracy of CNN-LSTM based on spatial feature factors is 83.91%, the prediction accuracy of CNN-LSTM based on air pollutants is 74.22%, the prediction accuracy of CNN-LSTM based on meteorological factors is 53.14%, and the prediction accuracy of CNN-LSTM based on multi-feature fusion is 87.21%. 2.5 The predicted results of the concentration and the true value are plotted as a line comparison chart as shown in Fig.14 As shown, the black solid line is PM 2.5 The actual value of concentration, the red solid line is the CNN-LSTM prediction value based on spatial feature factors, the purple solid line is the CNN-LSTM prediction value based on air pollutants, the green solid line is the CNN-LSTM prediction value based on meteorological factors, and the cyan solid line is the CNN-LSTM prediction value based on multi-feature fusion. It can be seen from the observation that the characteristics of the CNN-LSTM prediction results based on different input variables are basically similar to those of CNN, among which the CNN-LSTM prediction error based on meteorological factors is the largest, and the predicted value line is very different from the actual value line; the CNN-LSTM prediction value based on multi-feature fusion is basically consistent with the actual value line, and the prediction accuracy is the highest.

[0224] Table 20 Comparison of CNN-LSTM prediction errors for different input features

[0225]

[0226] The comparison of all the above ablation experiment errors is summarized in Table 21. Through comparison, it can be found that the CNN-LSTM model based on multi-feature fusion proposed in this embodiment has the smallest error, with a MAPE of 12.79%, a MAE of 2.19, and a RMSE of 2.84, that is, the prediction accuracy of the model is 87.21%, which is 1.91 percentage points higher than the prediction accuracy of the CNN based on multi-features, and 7.04 percentage points higher than the prediction accuracy of the LSTM based on multi-feature fusion. A detailed analysis of Table 21 shows that the prediction effects of the three models of CNN, LSTM, and CNN-LSTM based on meteorological factors are not ideal, with a prediction accuracy of about 50%, while the prediction accuracy of the three models based on air pollutants is about 70%, and the prediction accuracy of the three models based on spatial feature factors is about 80%. This result shows that although meteorological factors are related to PM, the prediction accuracy of the three models based on air pollutants is about 70%, and the prediction accuracy of the three models based on spatial feature factors is about 80%. 2.5 There is a certain correlation between the concentrations, but the model cannot learn more PM 2.5 Changes in information, that is, the effect of meteorological factors on PM 2.5The contribution of the concentration value prediction result is not significant; 2.5 The variable that contributes most to the concentration prediction result is the spatial characteristic factor, which shows that the model is more effective in predicting the concentration based on PM2. 2.5 The dimension of spatiotemporal changes in concentration is used to mine information in the learning sequence. When spatial characteristic factors, air pollutants, and meteorological factors are all used as input features, the prediction accuracy of the CNN-LSTM model is the highest, indicating that the factors of the three dimensions are indispensable and should all be taken into account. The ablation experiment also confirms the superiority of the CNN-LSTM model based on multi-feature fusion proposed in this paper.

[0227] Table 21 Comparison of prediction errors of ablation experiment models

[0228]

[0229]

[0230] To further verify the superiority of the prediction model proposed in this embodiment, the CNN-LSTM prediction model based on multi-feature fusion is compared with three commonly used machine learning models, namely BP, SVR and LightGBM. In the experiment, all features are used as input variables, namely PM 10 、SO 2 、NO 2 ,CO,O 3 , air temperature, dew point temperature, air pressure, wind direction, wind speed, spatial characteristic factors, all use the same training set and test set to train each model, and the prediction error is shown in Table 22. It can be seen that the prediction effect of the BP neural network is the least ideal, with a prediction accuracy of 81.16%, which may be because the model is prone to local fitting, and the learning speed of the model is slow, and the running time is too long; the prediction accuracy of the SVR model is 82.37%, which is 1.21 percentage points higher than the BP prediction accuracy. The model has strong linear and nonlinear feature fitting capabilities, but there are still problems with more complex calculations, especially when the data dimension is very high, SVR consumes more computing resources; the LightGBM algorithm is improved based on the GBDT algorithm, and supports distributed training, which can process large-scale data sets, improve training efficiency, and effectively solve the problem of processing massive data. The prediction accuracy of this model is 85.92%, which has obvious advantages over the BP and SVR prediction performance.

[0231] Combining Table 21 and Table 22, it can be seen that the CNN-LSTM prediction model based on multi-feature fusion proposed in this embodiment has better prediction performance, higher prediction accuracy, and can fit PM well. 2.5 The changing characteristics of the concentration series are analyzed and short-term predictions are made.

[0232] Table 22 Comparison of model prediction errors

[0233]

[0234] Therefore, this embodiment innovatively constructs a spatial feature factor based on the maximum information coefficient, proposes a CNN-LSTM combined prediction model based on multi-feature fusion, and explains the model construction steps and experimental results analysis. First, the experimental platform and environment are introduced, and then the framework and construction process of the model are analyzed in depth; secondly, the training set is input into the model for fitting training, and the test set is predicted and the prediction results are output, and the prediction accuracy of the model is discussed in combination with the evaluation indicators; finally, ablation experiments are carried out to design multiple models to verify the necessity of each component of the model proposed in this article, and the superiority of the model proposed in this article is verified by comparing it with three currently commonly used time series prediction models.

[0235] The above implementation modes are only used to illustrate the present invention, but not to limit the present invention. Ordinary technicians in the relevant technical field can make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, all equivalent technical solutions also belong to the scope of the present invention. The patent protection scope of the present invention should be defined by the claims.

Claims

1. A PM 2.5 The concentration prediction method is characterized in that The method comprises: About PM 2.5 The concentration value time series data is subjected to variational mode decomposition to obtain multiple decomposition sequences, which are eigenmode functions. When performing variational mode decomposition, the envelope entropy in the variational mode decomposition is used as the first objective function, and the minimum envelope entropy is used as the target value of the first objective function. The particle swarm optimization algorithm is used to determine the two parameters k and α in the variational mode decomposition, where the parameter k is used to define the PM 2.5 The number of modes selected when performing variational mode decomposition of concentration time series data, with reference to α, is used to determine the spectrum coverage of each decomposition sequence; Feeding the multiple decomposition sequences into the LSTM network model to obtain the predicted value of each decomposition sequence; The predicted values ​​of each decomposition sequence and PM 2.5 The mean square error between the true values ​​of concentration is used as the objective function, the minimum mean square error is used as the target value, and the weight value of each decomposition sequence is determined by the gray wolf optimization algorithm; The weight value of each decomposition sequence is multiplied by the predicted value and then added up to get PM 2.5 Predicted concentration.

2. The PM according to claim 1 2.5 The concentration prediction method is characterized in that The method of using the particle swarm optimization algorithm to determine the two parameters k and α in the variational modulus decomposition specifically includes: Taking the envelope entropy in variational modulus decomposition as the fitness function, Initialize the particle population, use the influencing parameter combination [α, k] as the particle position, randomly generate multiple influencing parameter combinations as the initial positions of the particles, and randomly set the initial moving speed for each particle; In one iteration, PM is analyzed under different particle position conditions. 2.5 The concentration value time series data is decomposed by variational mode to calculate the corresponding fitness value of each particle position Compare the fitness values ​​of each particle, and update the speed and position of the particle according to the order of merit. At the same time, synchronously refresh the global extreme value shared by the entire particle group; The loop is iterated until the number of iterations reaches the maximum set value, and then the optimal fitness value and particle position are output.

3. The PM according to claim 2 2.5 The concentration prediction method is characterized in that The velocity and position of the particle are determined by the following formula: Where w is the inertia weight, c1 and c2 are acceleration factors, η is a random number between [0,1], k is the number of iterations, represents the velocity of the i-th particle in the d dimension in the k+1-th iteration, represents the velocity of the i-th particle in the d dimension in the k-th iteration, represents the individual local extremum of the ith particle at degree d in the kth iteration, represents the position of the i-th particle in the d-dimension in the k-th iteration, represents the global optimal solution shared by the entire particle swarm in dimension d, represents the position of the i-th particle in the d dimension in the k+1-th iteration.

4. The PM according to claim 3 2.5 The concentration prediction method is characterized in that The envelope entropy in variational modulus decomposition is calculated by the following formula: In the formula, E p represents the envelope entropy, p j is the normalized form of a(j), representing a probability distribution sequence; a(j) is the envelope signal obtained by decomposing the zero-mean signal, j represents the sequence number of the signal, and N represents the total number of signals.

5. The PM according to claim 1 2.5 The concentration prediction method is characterized in that The method further comprises: Calculate PM at different locations 2.5 The MIC coefficient between the concentrations is calculated, and the spatial weight matrix is ​​constructed with the MIC coefficient as an element. The spatial weight matrix is ​​multiplied with the original matrix to obtain the PM at different locations. 2.5 The spatial characteristic factor of concentration, the original matrix consists of PM at all locations 2.5 Concentration composition; Construct a prediction model and a data set for training the prediction model, wherein the data set includes the spatial characteristic factors, air pollutant data and meteorological factor data, the spatial characteristic factors, air pollutant data and meteorological factor data are used as inputs of the prediction model, and the output of the prediction model is PM 2.5 Concentration values, using the trained prediction model to achieve PM 2.5 Concentration prediction.

6. The PM according to claim 5 2.5 The concentration prediction method is characterized in that The PM at different locations is calculated using the following formula 2.5 MIC coefficients between concentrations: Where X is the PM of the first position point 2.5 Concentration set, x is the PM at the first location 2.5 A PM in the concentration set X 2.5 concentration, Y is the PM at the second location 2.5 Concentration set, y is the PM at the second location 2.5 A PM in the concentration set X 2.5 concentration, P(x,y), P(x), P(y) are the joint and marginal probability densities of x and y, I(X,Y) is the mutual information of X and Y, n x and n y are the number of divisions on the x-axis and y-axis respectively; B(n) is the upper limit of the number of grids divided, and MIC is the PM of the first and second position points. 2.5 MIC coefficients between concentrations.

7. The PM according to claim 6 2.5 The concentration prediction method is characterized in that The spatial weight matrix is ​​multiplied with the original matrix by the following formula to obtain PM at different locations: 2.5 Spatial characteristic factor of concentration: Where, X Space (t) represents PM 2.5 Spatial characteristics of concentration series, W s×s represents the spatial weight matrix, X(t) represents the PM collected at each station 2.5 Concentration time series data, X s (t) represents the PM collected by each station s 2.5 Concentration time series data, r 12 、r 1s 、r 21 、r 2S 、r s1 、r s2 They represent the MIC coefficients between site 1 and site 2, site 1 and site s, site 2 and site 1, site s and site 1, and site s and site 2 respectively.

8. A PM 2.5 The concentration prediction device is characterized in that The apparatus comprises a processor configured to: About PM 2.5 The concentration value time series data is subjected to variational mode decomposition to obtain multiple decomposition sequences, which are eigenmode functions. When performing variational mode decomposition, the envelope entropy in the variational mode decomposition is used as the first objective function, and the minimum envelope entropy is used as the target value of the first objective function. The particle swarm optimization algorithm is used to determine the two parameters k and α in the variational mode decomposition, where the parameter k is used to define the PM 2.5 The number of modes selected when performing variational mode decomposition of concentration time series data, with reference to α, is used to determine the spectrum coverage of each decomposition sequence; Feeding the multiple decomposition sequences into the LSTM network model to obtain the predicted value of each decomposition sequence; The predicted values ​​of each decomposition sequence and PM 2.5 The mean square error between the true values ​​of concentration is used as the objective function, the minimum mean square error is used as the target value, and the weight value of each decomposition sequence is determined by the gray wolf optimization algorithm; The weight value of each decomposition sequence is multiplied by the predicted value and then added up to get PM 2.5 Predicted concentration.

9. A PM 2.5 The concentration prediction system is characterized by: The system comprises: Memory for storing computer programs; A processor, configured to execute the computer program to implement the method according to any one of claims 1 to 7. 10 . A non-transitory computer-readable storage medium storing instructions, which, when executed by a processor, perform the method according to claim 1 .

Citation Information

Patent Citations

  • PM2.5 concentration prediction method and system

    CN116050270A

  • Air quality time series prediction model research based on optimized variational mode decomposition

    CN116468170A

Cited By

  • Sub-industry load prediction method and system based on type analysis and feature reconstruction

    CN119209529A

  • Industry-specific load forecasting method and system based on type analysis and feature reconstruction

    CN119209529B

  • Air pollutant concentration prediction method and related equipment thereof

    CN119474762A

  • A method for predicting air pollutant concentration and related equipment

    CN119474762B

  • Urban fine-scale day-by-day atmospheric pollutant concentration estimation method

    CN119740205A