Air quality index prediction method based on improved grey wolf algorithm and DBN-ELM evolutionary deep learning model

Through the improved gray wolf optimization algorithm and DBN-ELM model, combined with TVFEMD and SE technology, the nonlinear processing and model parameter optimization problems in air quality index prediction are solved, the prediction accuracy and efficiency are improved, and more effective air pollution is provided Warning information.

CN118918984APending Publication Date: 2024-11-08NANJING TECH UNIV
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202410718226.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-06-04
Publication Date
2024-11-08

AI Technical Summary

Technical Problem

The prior art has problems such as difficulty in nonlinear data processing, large amount of calculation, long time to optimize model parameters, and slow convergence speed of deep learning models in air quality index prediction, resulting in unsatisfactory prediction accuracy and efficiency.

Method used

The improved Gray Wolf Optimization Algorithm (IGWO) combined with the deep learning model DBN-ELM is used to process data through TVFEMD decomposition technology and sample entropy (SE) technology, and introduce reverse search strategies and dimensional learning search strategies in the initialization and search stages. Optimize model parameters to improve prediction accuracy and efficiency.

Benefits of technology

It achieves higher air quality index prediction accuracy and efficiency, providing more effective air pollution warning information, which is better than traditional models and other benchmark models.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118918984A_ABST
    Figure CN118918984A_ABST
Patent Text Reader

Abstract

The invention belongs to the field of weather prediction, and particularly relates to an air quality index prediction method based on an improved grey wolf algorithm and a DBN-ELM evolutionary deep learning model, which comprises the following steps of: 1) acquiring an air quality index (AQI) data set of a certain station in four seasons, and processing an AQI sequence by using an empirical mode decomposition technology; 2) constructing a deep belief network-extreme learning machine (DBN-ELM) model, performing feature extraction and data analysis on the input data through the DBN, and then inputting the extracted effective features into the ELM model; 3) setting operation parameters of the improved grey wolf optimization algorithm, and obtaining parameters of a DBN-ELM model; 4) utilizing the DBN-ELM model to predict each component of an internal model function (IMF); and 5) reconstructing the predicted value of each component of the IMF to obtain an overall AQI predicted value. The model is compared with other reference models, and the result shows that the model is superior to other models in the aspects of precision and generalization, which indicates that the model can effectively predict the air quality.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of air quality prediction and proposes an air quality index prediction method based on an improved grey wolf algorithm and a DBN-ELM evolutionary deep learning model. Background Art

[0002] With the rapid development of the world economy and the advancement of industrialization and urbanization, air pollution has become increasingly serious. Air pollutants mainly include carbon monoxide (CO), carbon dioxide (CO 2 ), sulfur dioxide (SO 2 ,SO 2 ), nitrogen oxides (nitrogen oxide, NO x ) and other particles, inhalable particles (PM 10 ), fine particulate matter (PM 2.5 ) etc. The presence of air pollutants causes the air quality to deteriorate continuously, which not only seriously affects people's daily life and health, but also brings great obstacles to the sustainable development of society. The Air Quality Index (AQI) is a dimensionless indicator used to evaluate air quality. It mainly considers six pollutants: carbon monoxide (CO), sulfur dioxide (SO 2 ), nitrogen dioxide (NO 2 ), ozone (O 3 ), fine particulate matter (PM 2.5 ) and inhalable particulate matter (PM 10 ). And combined with the new environmental air quality, the calculation method was evaluated. The larger the AQI value, the worse the air quality. Therefore, accurate prediction of the air quality index is of great significance to the prevention and control of air pollution. However, the atmospheric environment is a dynamic, nonlinear, and non-stationary system, which makes accurate analysis and prediction of AQI very difficult. In recent years, in order to obtain scientific and accurate AQI prediction results, many scholars have conducted in-depth research on AQI and proposed many effective prediction methods. These methods can be divided into statistical methods, physical methods, intelligent methods, and hybrid methods.

[0003] Statistical methods were used to analyze the changing patterns of pollutants in the air. Statistical models mainly include gray model (GM) and multivariate linear regression (MLR). However, the performance of these models is not enough to predict nonlinear data. Physical methods can predict the concentration of air pollutants by simulating the transport and diffusion of pollutants in the atmosphere. Commonly used models for physical methods are Weather Research and Forecasting (WRF) model, chemical coupling meteorological chemistry model, etc. However, the implementation of such models has problems such as large amount of calculation, complex process, and difficult data collection. Intelligent methods have good learning ability for nonlinear features in data, which makes intelligent methods widely used in the prediction field. Commonly used intelligent methods include support vector regression (SVR), artificial neural network (ANN) and other models. However, these machine learning models still have limitations for large data sets. With the rapid development of deep learning technology, deep learning has been widely favored by scholars and has been applied to air quality prediction, such as long short-term memory (LSTM), deep belief network (DBN), etc. Deep learning models have strong learning ability and high prediction accuracy, and have good learning ability for a large number of experimental samples. But at the same time, deep learning models also have problems such as generally long training time and non-optimal model parameters. Therefore, it is of great significance to improve the deep learning model and optimize the parameters. The DBN model based on deep learning theory is an efficient deep learning prediction model with powerful data analysis and feature extraction capabilities. However, in a large number of iterative calculations, its convergence speed is prone to be too slow, which makes the nonlinear approximation performance of DBN in the prediction field unsatisfactory. In order to solve the above problems, the superior performance of DBN in data analysis and feature extraction is combined with the fast learning speed and good generalization performance of the extreme learning machine (ELM). In the DBN-ELM model, DBN is used for feature extraction and dimensionality reduction. The hidden layer of the last restricted Boltzmann machine (RBM) in DBN is used as the input layer of ELM.

[0004] The parameters of the deep learning neural network model are usually set empirically, which makes it take a lot of time to test the best prediction performance of the model. The swarm intelligence optimization algorithm can find the optimal parameters of the model according to the characteristics of the data set. Commonly used intelligent optimization algorithms include particle swarm optimization (PSO), whale optimization algorithm (WOA), sine cosine algorithm (SCA), Harris Hawk optimization algorithm (HHO) and gray wolf optimization algorithm (GWO). The GWO algorithm has a convergence factor and information feedback mechanism that can be adjusted adaptively, which can achieve a balance between local optimization and global search, and has good performance in problem solving accuracy and convergence speed. Therefore, the present invention uses an improved GWO algorithm to obtain the parameters of the deep learning model. Summary of the invention

[0005] In order to solve the problems existing in the prior art, the present invention combines data processing, intelligent optimization and time series prediction, and proposes an air quality index prediction method based on an improved gray wolf optimization algorithm and an evolutionary deep learning model of DBN-ELM. The air quality index prediction method based on an improved gray wolf optimization algorithm and an evolutionary deep learning model of DBN-ELM described in the present invention comprises the following steps:

[0006] 1) Combining the advantages of TVFEMD decomposition technology, improved GWO algorithm and deep learning model DBN-ELM, a new hybrid model (TVFEMD-SE-IGWO-DBN-ELM) is proposed for air quality index prediction to provide air pollution warning and promote ecosystem recovery. The superior performance of the hybrid model in air quality prediction is verified by applying it to the air quality index (AQI) dataset of a station throughout the year;

[0007] 2) The TVFEMD decomposition technique is used to process the AQI sequence, and the SE technique is used to reconstruct the decomposed components. In view of the nonlinear and non-statistical characteristics of AQI data, the TVFEMD decomposition technique is used to eliminate the interference and noise in the original sequence and extract the effective information of the original sequence. Since the TVFEMD technique obtains too many components, the SE technique is used to reduce the calculation time and the cumulative error of repeated modeling;

[0008] 3) The DBN-ELM model is built for subcomponent modeling. Based on the traditional DBN model, the DBN-ELM model introduces ELM as a regression layer. In the DBN-ELM model, the hidden layer of the last RBM of DBN serves as the input layer of ELM. The DBN-ELM model combines the advantages of DBN in feature extraction and ELM in nonlinear fitting, improving the accuracy of AQI prediction;

[0009] 4) In order to solve the problem of lack of population diversity and easy to fall into local optimum of GWO algorithm, the original GWO algorithm is improved. In the initialization stage, the anti-search strategy is introduced to enrich the population information, and the DLH strategy is introduced in the search stage to solve the problem of falling into local optimum and accelerate the convergence speed. The experimental results show that the IGWO algorithm proposed in this invention is better than the unimproved GWO algorithm in overall performance;

[0010] 5) The parameters of the DBN-ELM model are optimized using the IGWO algorithm, and the DBN-ELM model with the optimal parameters is selected to predict AQI. The excellent performance of the TVFEMD-SE-IGWO-DBN-ELM model is verified by comparing it with benchmark models such as BP, DBN-ELM, PSO-DBN-ELM, GWO-DBN-ELM, and TVFEMD-SE-GWO-DBN-ELM. This model can be used as an effective tool for predicting the air quality index (AQI) and provide reliable information for early warning of air pollution.

[0011] The steps of the improved gray wolf optimization algorithm in the air quality index prediction method based on the improved gray wolf optimization algorithm and the DBN-ELM evolutionary deep learning model are as follows:

[0012] As a social animal, gray wolves have a strict social hierarchy. The Gray Wolf Optimization Algorithm (GWO) divides the gray wolf population into four levels from top to bottom: α, β, δ, and ω according to its hierarchy. When searching for prey, gray wolves will approach and surround the prey. The mathematical model of this search is:

[0013]

[0014] Among them, A and C are coefficient vectors; x prey (t) represents the position of the prey; x(t) represents the current position of the gray wolf; a is the convergence factor, which decreases linearly from 2 to 0 during the entire iteration process; r 1 and r 2 is a random vector in [0, 1] and t is the current iteration number.

[0015] When the gray wolf hunts prey, the ω wolf updates its own position based on the position information of the other three wolves. The mathematical expressions are as follows:

[0016]

[0017]

[0018]

[0019] where x α (t), x β (t) and x δ (t) represents the position of α, β, and δ wolves in the current population, respectively, and x(t) represents the position of the gray wolf; D α , D β and D δ Respectively represent the distance between the current candidate gray wolf and the best three wolves; A 1 , C 1 , A2 , C 2 , A 3 and C 3 The coefficient vectors represent α, β, and δ respectively. When |A|>1, the gray wolves try to disperse in various areas to find prey; when |A|<1, the gray wolves will concentrate on finding prey in one or some areas.

[0020] In the GWO algorithm, the ω wolf searches for the optimal solution in the search space under the leadership of the a, β, and δ wolves. Due to the disadvantages of being easily trapped in the local optimal solution and reducing the diversity of the population during the search process, the initialization and search strategies of the gray wolf optimization algorithm are improved to address the above problems.

[0021] In the initialization phase, the reverse search strategy is introduced to randomly generate the initial population and its reverse individuals, and the fitness of the reverse individuals is compared with the original individuals. If the fitness of the other individual is better than that of the original individual, the fitness of the other individual is adopted, otherwise the fitness of the original individual is adopted. The position vector expression of the relative individual is as follows:

[0022]

[0023] Where x'(t) represents the position of the opposite individual; represents the position of the original population; l bound and u bound are the upper and lower bounds of the search space.

[0024] The improvement of the search strategy is to introduce the dimensional learning based optimization (DLH) search strategy. In the DLH strategy, the mathematical model of the updated position of the gray wolf individual in each dimension is obtained by formula (1). Individual wolves learn from different neighbors, and the wolves are randomly selected from ordinary wolves. In addition to x ig In addition to (t+1), the DLH search strategy also generates a new position x of another candidate wolf. id (t+1). To this end, first, by calculating the current position x of the gray wolf individual i (t) and the position x of the candidate gray wolf ig The Euclidean distance between (t+1) and the radius r is calculated according to formula (6) i (t).

[0025] r i (t) = x i (t)-x ig (t+1) (6)

[0026] Then, x i The neighborhood n of (t) i (t) is constructed according to equation (7), where d i is xi (t) and x j (t) is the Euclidean distance between them.

[0027] n i (t) = {x j (t)|d i (x i (t), x j (t))≤r i (t), x j (t)∈pop} (7)

[0028] When x i When the neighborhood of (t) is constructed, multi-neighborhood learning can be constructed by equation (8). id The m-dimensional m(m+1) of (t+1) is obtained by i (t) and the random number x of the ordinary wolf r,m A random neighborhood x selected from (t) n,m (t) calculated.

[0029] x id,m (t+1)=x i,m (t)+rand×(x n,m (t)-x r,m (t)) (8)

[0030] In the update phase of selection, first, the fitness values ​​x of the two candidates are compared by equation (9): ig (t+1) and x id (t+1) to select a more suitable candidate solution.

[0031]

[0032] Then update x i (t+1) If the fitness value of the candidate solution is less than x i (t), then update x by the current candidate solution i (t), otherwise, x i The position of (t) remains unchanged. Finally, after performing this process for all individuals, the number of iterations is increased by 1 until the maximum number of iterations is reached.

[0033] Compared with other benchmark models, the model designed in this scheme has the following two advantages:

[0034] 1. The model designed by this scheme is superior to other models in terms of accuracy and generalization;

[0035] 2. Compared with other models, this scheme can predict the air quality index more effectively. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 A site model performance comparison chart;

[0037] Figure 2 TVFEMD-SE-IGWO-DBN-ELM overall flow chart; DETAILED DESCRIPTION

[0038] The present invention is further described in detail below in conjunction with the accompanying drawings and specific implementation methods. The specific implementation scheme of the improved gray wolf algorithm and the evolutionary deep learning model of DBN-ELM in the air quality index prediction method: First, the original air quality index data is decomposed into multiple intrinsic mode function components by using the time-varying filtering-based empirical mode decomposition (TVFEMD) method; in order to reduce the amount of calculation, sample entropy (SE) is introduced to estimate multiple IMF components. Secondly, the gray wolf optimization algorithm is improved, and an optimization strategy based on dimensionality learning is introduced to avoid falling into local optimality; at the same time, a reverse search strategy is introduced in the initialization of the DLH strategy to enrich population information. Third, the IGWO algorithm is used to obtain parameters of the deep belief network-extreme learning machine (DBN-ELM) model. Then, the optimal DBN-ELM parameter model obtained is used to predict each IMF component respectively. Finally, the predicted values ​​of each IMF component are reconstructed to obtain the overall AQI prediction value.

[0039] The improvement percentage can compare the performance between different models, and Table 2 lists the model improvement percentages for site A. The abbreviations of different models are given in Table 1. The eight models from BP model to TVF-SE-IGWO-DBN-ELM are set as M1, M2, …, M8.

[0040] From Table 2 and Figure 1 It can be seen that the overall performance of M4 is better than that of M3. Taking PRMSE as an example, compared with M4, M3 increases by 3.6% in spring, 1.2% in summer, 2.07% in autumn, and 3.48% in winter. From the comparison between M4 and M6, it can be seen that the performance can be effectively improved by optimizing the model parameters using this algorithm. By using the TVFEMDSE technology, that is, the comparison between M6 and M7, it can be seen that the TVFEMD-SE technology can greatly improve the forecast accuracy of AQI. Compared with M7 and M8, the use of the IGWO algorithm can further improve the performance of the model.

[0041] Table 1 Model abbreviations

[0042]

[0043] Table 2A Site improvement percentage results

[0044]

Claims

1. Air quality index prediction method based on improved grey wolf algorithm and DBN-ELM evolutionary deep learning model, its features include Follow these steps: 1) Collect the air quality index data set of a station in four seasons. In case of missing values ​​in the collected AQI data, preprocess the data by supplementing the missing values. The method of supplementing the missing values ​​is cubic spline interpolation; 2) The TVFEMD decomposition technology is used to process the AQI sequence, and the SE technology is used to reconstruct the decomposed components. In view of the nonlinear and non-statistical characteristics of AQI data, the TVFEMD decomposition technology is used to eliminate the interference and noise in the original sequence and extract the effective information of the original sequence; 3) Construct a DBN-ELM model, perform feature extraction and data analysis on the input data through DBN, and then input the extracted valid features into the ELM model; 4) In order to solve the problem of lack of population diversity and easy to fall into local optimum of GWO algorithm, the original GWO algorithm was improved. In the initialization stage, the anti-search strategy was introduced to enrich the population information. In the search stage, the search strategy based on dimensionality learning was introduced to solve the problem of falling into local optimum and accelerate the convergence speed. The parameters of DBN-ELM model were optimized by using IGWO algorithm. 5) The optimized DBN-ELM model is used to predict the AQI sequence and the residual sequence to obtain the predicted value of the subsequence, and then the predicted value of the subsequence is reconstructed to obtain the predicted value of AQI.

2. According to the air quality index prediction method based on the improved gray wolf algorithm and the evolutionary deep learning model of DBN-ELM described in claim 1, the initialization and search strategy of the gray wolf optimization algorithm are improved as follows: In the initialization stage, the reverse search strategy is introduced to randomly generate the initial population and its reverse individuals. The fitness of the reverse individuals is compared with that of the original individuals. If the fitness of the other individual is better than that of the original individual, the fitness of the other individual is adopted. Otherwise, the fitness of the original individual is adopted. The position vector expression of the relative individual is as follows: where x'(t) represents the position of the opposite individual, represents the position of the original population, l bound and u bound are the upper and lower limits of the search space. The improvement of the search strategy is to introduce the dimension learning based optimization (DLH) search strategy. In the DLH strategy, the mathematical model of the updated position of the gray wolf individual in each dimension is obtained by formula (1). The individual wolf learns from different neighbors, and the wolf is randomly selected from the ordinary wolves, except for x ig In addition to (t+1), the DLH search strategy also generates a new position x of another candidate wolf. id (t+1), to this end, first, by calculating the current position x of the gray wolf individual i (t) and the position x of the candidate gray wolf ig The Euclidean distance between (t+1) and the radius r is calculated according to formula (2) i (t): r i (t) = x i (t)-x ig (t+1) (2) Then, x i The neighborhood n of (t) i (t) is constructed according to equation (3), where d i is x i (t) and x j (t) The Euclidean distance between ni (t) = {x j (t)|d i (x i (t), x j (t))≤r i (t), x j (t)∈pop} (3) When x i When the neighborhood of (t) is constructed, multi-neighborhood learning can be constructed by equation (4); id The m-dimensional m(m+1) of (t+1) is obtained by i (t) and the random number x of the ordinary wolf r,m A random neighborhood x selected from (t) n,m (t) Calculated: x id,m (t+1)=x i,m (t)+rand×(x n,m (t)-x r,m (t)) (4) In the update phase of selection, first, the fitness values ​​x of the two candidates are compared by equation (5) ig (t+1) and x id (t+1) to select a more suitable candidate solution, Then update x i (t+1) If the fitness value of the candidate solution is less than x i (t), then update x by the current candidate solution i (t), otherwise, x i The position of (t) remains unchanged. Finally, after performing this process for all individuals, the number of iterations is increased by 1 until the maximum number of iterations is reached.

3. According to the air quality index prediction method based on the improved gray wolf algorithm and the evolutionary deep learning model of DBN-ELM according to claim 1, the steps of the improved gray wolf optimization algorithm (IGWO) are as follows: a. Set the population size, number of iterations, and upper and lower boundaries of the search space of the gray wolf optimization algorithm; b. Randomly generate an initial population, generate individuals corresponding to the initial population, and calculate the fitness value of each gray wolf; c. Compare the fitness of the relative individual with the fitness of the original individual and save the individual with the best fitness; d. Introduce the DLH search strategy, search for the optimal solution, and update the position of the gray wolf according to formula (5); e. Determine whether the maximum number of iterations has been reached. If so, output the optimal parameters of DBN-ELM, otherwise jump to step b.

Citation Information

Patent Citations

  • Short-term air quality prediction method based on sparrow search algorithm and decomposition error correction

    CN115660167A

  • Short-term traffic flow prediction model based on variational mode decomposition multi-stage optimization

    CN116721542A