CN111598156A

CN111598156A - PM based on multi-source heterogeneous data fusion2.5Prediction model

Info

Publication number: CN111598156A
Application number: CN202010405889.7A
Authority: CN
Inventors: 李晓理; 张博; 王康; 崔桂梅
Original assignee: Beijing University of Technology
Current assignee: Beijing University of Technology
Priority date: 2020-05-14
Filing date: 2020-05-14
Publication date: 2020-08-28

Abstract

The invention provides a PM based on multi-source heterogeneous data fusion_2.5And the prediction model integrates information by fusing multi-source heterogeneous data, and establishes a multi-core support vector regression model on the basis of the fused data to realize PM2.5 concentration prediction. Firstly, collecting two multi-source heterogeneous data of air quality data and images, and preprocessing and extracting characteristics of the collected data. Secondly, multi-source heterogeneous data fusion is completed by using a multi-core expansion method based on a kernel function, the fusion process is mainly completed by constructing and combining a Gram matrix, and the multi-core expansion kernel function is deduced. Then, on the basis of the multi-core extended kernel function and the extended kernel matrix, a multi-core support vector regression model is reconstructed. Finally, the model parameters are optimized by using an improved minimum sequence optimization algorithm. The method is used for predicting PM2.5 based on multi-source heterogeneous data fusion, and can be obtained on the basis of realizing information fusionMore comprehensive and credible judgment is obtained, and the accuracy, stability and credibility of prediction are ensured.

Description

PM based on multi-source heterogeneous data fusion2.5Prediction model

Technical Field

The invention relates to PM (particle matter) based on multi-source heterogeneous data fusion_2.5The prediction model is mainly used for the monitoring and early warning of air quality and other related works, and belongs toThe technical field of environmental monitoring.

Background

The rapid development of economy and the promotion of industrialization process cause unprecedented huge pressure on the ecological environment. For a long time, the hidden troubles accumulated by the economic development mode mainly aiming at economic growth and environmental protection are gradually revealed, and under the condition that the pollution of inhalable particles and total suspended particles cannot be completely solved, the long-term and continuous haze in economically developed areas such as Jingjin Ji and the like seriously threatens the life and production activities of people. In order to effectively prevent and control harm caused by haze and strengthen atmospheric PM_2.5Monitoring and achieving early PM_2.5The concentration prediction has sufficient practical significance.

Conventional PM based on air quality data_2.5The prediction method mainly comprises a numerical method and a statistical method. The numerical method usually combines the interdisciplinary knowledge of atmospheric dynamics, chemistry, mathematics and the like, analyzes the processes of pollutant generation, transportation, conversion, settlement and the like, and utilizes a substance conservation equation to carry out the treatment on the PM_2.5And (5) making quantitative prediction. The statistical method is to establish a model for PM by analyzing the internal change rule of the historical monitoring data_2.5And (6) predicting. At present, most of PM_2.5The prediction is carried out by utilizing a statistical method, and common prediction models comprise ARIMA models, multiple linear regression models, autoregressive moving average models, time series, grey statistics and other prediction methods. PM based on statistical method_2.5The prediction requires high reliability of the original data, or after the data preprocessing process, the obtained training data can accurately reflect PM_2.5Characteristic information of (1). However, PM is based on modeling because of the irreparable data caused by some human factors or accidents_2.5Prediction faces more difficulty. In addition, image-based PM has been developed in recent years_2.5Although the prediction method is free from the constraint of large-scale equipment in data acquisition, the prediction accuracy is poor.

In consideration of the defects of the single-source simple substance data prediction method, the invention provides a PM based on multi-source heterogeneous data fusion_2.5And (4) predicting the model.The model establishes a prediction model for PM by using data from different data sources and different structures_2.5And (6) predicting. The multi-source heterogeneous data can be used for PM from different angles_2.5The change of the data is represented, the information fusion of the multi-source heterogeneous data can be realized through the multi-source heterogeneous data fusion, the confidence coefficient of the data can be increased through the information complementation, the reliability is improved, and the uncertainty is reduced. The regression prediction model established based on the multi-source heterogeneous data can obtain more comprehensive estimation and judgment, and is favorable for improving PM_2.5Accuracy and stability of the prediction.

Disclosure of Invention

The invention provides a PM based on multi-source heterogeneous data fusion_2.5The prediction model collects air quality data and image data as training data, multi-source heterogeneous data fusion is realized by using a kernel method, and a multi-core Support Vector Regression (SVR) model is established on the basis of the fusion data to realize PM_2.5And (6) predicting.

The technical scheme adopted by the invention is PM based on multi-source heterogeneous data fusion_2.5Prediction, comprising the following steps:

step 1: the method comprises the steps of collecting two kinds of heterogeneous data of air quality data and images, and preprocessing the collected data.

Step 2: and (3) performing feature extraction on the preprocessed data, particularly extracting image features of the image data by adopting a digital image processing technology.

And step 3: verifying extracted features with PM_2.5Correlation of concentration values completes feature selection.

And 4, step 4: and mapping by using different kernel functions based on the types of the data characteristics, and realizing multi-source heterogeneous data fusion by using a kernel method.

And 5: and establishing a multi-core SVR regression model based on the multi-source heterogeneous fusion data.

Step 6: and optimizing parameters of the multi-core SVR model by using an improved minimum sequence optimization algorithm (SMO) to determine a model structure.

And 7: and testing the prediction effect of the multi-core SVR model by using the test data.

And 8: and evaluating the multi-core SVR model by using the related evaluation indexes.

PM based on multi-source heterogeneous data fusion of the embodiment of the invention_2.5Forecasting, namely taking two kinds of heterogeneous data of air quality data and images as input, after feature extraction is carried out on the data, utilizing a kernel method to project the features of heterogeneous space to the same kernel space after kernel function mapping is carried out on the features, training a regression model on the kernel space, and completing PM (particle mass model) matching_2.5And (4) predicting. A prediction model obtained based on multi-source heterogeneous data is higher in prediction precision and better in stability.

In addition, the PM based on multi-source heterogeneous data fusion according to the embodiment_2.5The prediction method has the following additional technical features:

in step 1, the air quality data includes PM₁₀、PM_2.5、SO₂、NO₂、CO、O₃Temperature and humidity, and additionally require pre-processing of the image data, including translation and scaling of the image, with 1920 x 1080 pixels of the acquired image and 320 x 240 pixels after scaling.

In the feature extraction process in step 2, a total of 5 features related to spatial contrast, dark channel intensity and HIS color space difference (three dimensions) are extracted.

According to the atmospheric transmission model i (x) ═ j (x) t (x) + a (1-t (x)), extinction of atmospheric light and transmittance are in an inverse relationship, and both satisfy the following formula:

wherein b is_extIs the extinction coefficient, and r (x) is the transmission distance of light. The atmosphere is similar to a low-pass filter and filters out high-frequency information of the image, so that the image information is reduced. Defining local contrast as the first feature:

F_ig＝|▽_xI(x)|。

the image dark channel intensity is defined as:

where Ω (x) is a block centered on pixel x, J is the scene radiation, J^cOne of the color channels is represented. It can be seen that the dark channel intensity value for a given pixel is the minimum of the three color co-channels of the tile. The priori knowledge of a large number of haze-free images shows that the dark channel strength value of the haze-free images is 0, namely: j. the design is a square_dark→ 0, in combination with the atmospheric transmission model, the extinction ratio can be obtained:

in the formula A^cIs atmospheric light, the extinction ratio t (x) is therefore selected as the second feature F_id

Color difference and atmospheric extinction of sky in HIS color space_extThere is an exponential relationship, which can be expressed as: b_ext＝ae^bΔDWhere a and b are model parameters and Δ D is used to describe the difference in HIS space. Due to the difficulty in obtaining b_extThe influence parameters of the three parts of the HIS are defined as follows, and therefore, the difference values of the three parts in the HIS color space are used as features:

d_x(h)＝I_h(x,y)-I_h(x+1,y)

d_y(h)＝I_h(x,y)-I_h(x,y+1)

wherein, I is input image with pixels of m × n, I_h(x, y) is the h value of pixel (x, y). Likewise, F_isAnd F_iiThe definition is as follows:

and 4, performing multi-source heterogeneous data fusion by adopting a multi-core learning method. For a given sample space (x)₁,y₁),(x₂,y₂),…,(x_l,y_l) ∈ X × Y, wherein space is input

Output space

Two different RBF kernel functions are selected in consideration of the learning ability and generalization ability of the regression model. In order not to lose any original information, the multi-core matrix is merged, i.e. the multi-core expansion method is adopted. The multi-core extended core matrix contains all original core matrixes, so that the properties of original core functions are saved. The form of the multi-core extended core matrix is:

wherein the diagonal matrix of the new matrix is the original kernel matrix, (K)_p,p′)_i,j＝K_p,p′(x_i,x_j) Representing the mixture of two different kernel matrices, M being the number of kernel functions used and l being the total number of samples. The mixture of two different kernel matrices can be obtained by the following formula:

it can be seen that when p ═ p', k_p,p′≡k_pWhere σ is_p、σ_p′Is a parameter of the RBF kernel function.

The root mean square error (e) is used in step 8_rmse) Mean absolute percentage error (e)_mape) And correlation coefficient (R)²) The model was evaluated for 3 indices:

in the formula: y is_iRepresents the PM corresponding to the ith sample_2.5The actual value of the concentration of the water,

represents the PM corresponding to the ith sample_2.5The predicted value of the concentration of the active ingredient,

representing the model prediction output average. e.g. of the type_rmseReflecting the stability of the predicted output value of the model, e_mapeReflecting the degree of deviation of the model prediction output value from the actual value, wherein the smaller the two values are, the better the model performance is; r²Reflecting the degree of correlation between the predicted output value and the true value of the model, the closer the value is to 1, the better the performance of the model is.

Drawings

FIG. 1 is a schematic diagram of a multi-source heterogeneous data fusion process according to the present invention

FIG. 2 is a PM process based on multi-source heterogeneous data fusion according to the present invention_2.5Schematic diagram of prediction process

Detailed Description

The following describes embodiments of the present invention in detail, and the embodiments describe technical solutions related to the present invention in detail and explain principles of the present invention in detail. And on the premise of the technical scheme of the invention, detailed implementation modes and specific operation processes are given, but the protection scope of the invention is not limited by the following examples.

The sampling period of the data collected in this embodiment is 1 hour, the air quality data is collected by the micro weather station, the image is collected by the 360-degree camera, the collected sample data is prepared into sample pairs, and 700 groups of sample data are collected in total, wherein 500 groups are used as training data, and 200 groups are used as test data. The method comprises the following steps of carrying out data cleaning, integration, conversion and specification on air quality data, carrying out standardized processing on the data, wherein a conversion function is as follows:

in the formula: x is the original data, X_minIs the minimum value in the sample data, X_maxIs the maximum value, X, in the sample data^*For data after normalization, X^*∈[0,1]. The image is translated and scaled to 1920 x 1080 pixels of the captured image and 320 x 240 pixels after scaling.

The increase of the atmospheric PM2.5 concentration causes the image information amount to be reduced, and the image characteristics are reflected to visually represent that the image contrast is reduced, the dark channel intensity value is increased, and the HIS color space difference is reduced. This example extracts F_ig，F_id，F_ih，F_is，F_iiThere are 5 features in total.

Verification of all features with PM Using Pearson correlation coefficients_2.5And (4) selecting the characteristic with stronger correlation as the prediction characteristic according to the concentration correlation. Through verification, the selected features of the embodiment include: PM (particulate matter)₁₀，PM_2.5，SO₂，NO₂，CO，O₃Temperature, humidity, F_ig，F_id，F_ih，F_is，F_ii。

In this embodiment, in order to retain more local data distribution information, a core matrix is constructed by using a multi-core extended core method. Setting air quality characteristic data u_i＝{PM_2.5，PM₁₀，SO₂，NO₂，CO，O₃Temperature, humidity }, image characteristic data v_i＝{F_ig，F_id，F_ih，F_is，F_ii}. Merging the two characteristic data to obtain heterogeneous characteristic data x_i＝{u_i,v_i1,2, …, l, and the input space X is X ═ X₁,x₂,…,x_l}∈R_13×lOutput space Y ═ Y₁,y₂,…,y_l}∈R_1×l. The subject uses two different RBF kernel functions k_P、k_RComputing a kernel matrix K_P，K_P,R，K_R，K_R,PSynthesizing a multi-core extended core matrix as follows:

According to the multi-core extended core matrix in the step 4, the following optimization problems are constructed:

s.t.f(x_i)-y_i≤+ξ_i,

where ω is a weight vector, tolerable error, ξ_i，

Two relaxation variables, C is a penalty parameter, Lagrange conversion formula is used for omega, b and ξ respectively_i，

The partial derivative is zero:

α therein_i、

For the introduced lagrange multiplier, the dual problem to get the above optimization problem is:

the optimal weight vector omega is obtained by calculating the lagrange multiplier alpha, so that the discriminant function of the multi-core extended kernel is as follows:

α therein_iAnd

cannot be zero at the same time, so b can be calculated as:

step 6: and optimizing parameters of the multi-core SVR model by using an improved minimum sequence optimization algorithm, and determining the model structure.

1) Solving α_jOf optimum value α'_jThe feasible region is L ≦ α'_jH is less than or equal to H. Wherein the content of the first and second substances,

is α_iThe original value of (a) is set,

is α_jOriginal value of (2):

by solving for the knowledge α_jThe optimized values of (A) are:

2) solving α_iOf optimum value α'_i：

3) The offset term b is modified as follows:

wherein, b'_iAnd b'_jTo optimize the value, b^oThe value is the original value.

For both the original problem and the dual problem, if the solution of the original problem is (ω, ξ), the original problem is solved by (ω, ξ)^*) While solving the dual problem as (α)^*) Let R (ω, ξ)^*)＝W(α,α^*) The following can be obtained:

order to

Then there are:

indicating the accuracy of the iterative algorithm when

And when the precision of the algorithm is less than the given precision, the iteration stops.

The stop criterion and the KKT stop criterion are matched for use, and the operation is performed once after a plurality of iterations are set

And detection can greatly improve the algorithm efficiency.

From 700 groups of sample pairs, 500 groups are selected for training the model, and the residual 200 groups are used as a test set to test the prediction effect of the model.

Using root mean square error (e)_rmse) Mean absolute percentage error (e)_mape) And correlation coefficient (R)²) The model was evaluated for 3 indices.

Claims

1. PM based on multi-source heterogeneous data fusion_2.5The concentration prediction model is characterized by comprising the following steps:

step 1: acquiring two kinds of heterogeneous data of air quality data and images, and performing data preprocessing on the two kinds of acquired heterogeneous data;

step 2: performing feature extraction on the preprocessed data; acquiring image characteristics of the image heterogeneous data by adopting a digital image processing technology;

and step 3: verifying the features and PM extracted in the step 2_2.5The relevance of the concentration values completes the feature selection;

and 4, step 4: mapping by using different kernel functions based on the types of the data characteristics to realize multi-source heterogeneous data fusion;

and 5: establishing a multi-core SVR regression model based on multi-source heterogeneous data fusion data;

step 6: optimizing parameters of the multi-core SVR model by using an improved minimum sequence optimization algorithm, and determining a model structure;

and 7: testing the prediction effect of the multi-core SVR model by using the test data;

2. The PM based on multi-source heterogeneous data fusion of claim 1_2.5A concentration prediction model characterized by: collecting two types of data of air quality data and images, and translating and scaling the collected RGB images, wherein the pixels of the collected RGB images are 1920 x 1080, and the pixels of the collected RGB images are scaledThe pixels of the RGB image are 320 × 240.

3. The PM based on multi-source heterogeneous data fusion of claim 1_2.5A concentration prediction model characterized by: extracting image features 5 features are extracted that are related to spatial contrast, dark channel intensity, and HIS color space difference.

4. The PM based on multi-source heterogeneous data fusion of claim 1_2.5A concentration prediction model characterized by: the parameters of the kernel function used are | | | x according to the standard deviation sigma in the RBF kernel function_i-x_j||²Wherein x represents the fused heterogeneous feature samples, and i and j traverse all the samples; the parameters in the polynomial kernel function are usually set to p 0 and q 1.

5. The PM based on multi-source heterogeneous data fusion of claim 1_2.5A concentration prediction model characterized by: performing the following operations on the multi-source heterogeneous data characteristics: setting air quality characteristic data u_i＝{PM_2.5，PM₁₀，SO₂，NO₂，CO，O₃Temperature, humidity }, image characteristic data v_i＝{F_ig，F_id，F_ih，F_is，F_iiIn which F_igFor image contrast, F_idFor dark channel strength characteristics, F_ih，F_is，F_iiIs HIS space difference characteristic; fusing the two characteristic data to obtain a heterogeneous characteristic sample x_i＝{u_i,v_i1,2, …, l, and the input space X is X ═ X₁,x₂,…,x_l}∈R_13×lOutput space Y ═ Y₁,y₂,…,y_l}∈R_1×lWherein y is_iIndicating the corresponding PM2.5 concentration value label.

6. The PM based on multi-source heterogeneous data fusion of claim 1_2.5Concentration prediction model, characterized by using a kernel methodMulti-source heterogeneous data fusion is realized; the multi-core extended core matrix constructed using two different RBF kernel functions is of the form:

it is clear that when p ═ p', there is K_p,p′≡K_p(ii) a Wherein sigma_p、σ_p′Parameters of the RBF kernel function; the corresponding multi-core extended core matrix is:

wherein (K)_p,p′)_i,j＝K_p,p′(x_i,x_j) Representing the mixture of two different kernel matrices, M being the number of kernel functions used and l being the total number of samples.

7. The PM based on multi-source heterogeneous data fusion of claim 1_2.5A concentration prediction model characterized by: and (3) forming sample pairs by the collected multi-source heterogeneous data, wherein the sample pairs are 700 groups, 500 groups are used as model training data, and 200 groups are used as test data.

8. The PM based on multi-source heterogeneous data fusion of claim 1_2.5A concentration prediction model characterized by: when the model parameters are optimized by using the SMO algorithm, the iteration is stopped when the precision of the iterative algorithm is less than 0.01.