CN111563135A - GM(1,3) Model Urban Land Subsidence Prediction Method Combining Terrain Factors and Neural Networks - Google Patents

Abstract

The invention discloses an urban land subsidence prediction method combining terrain factors and GM (1,3) models, which includes collecting historical observation data of a site to be predicted and several nearby sites, and each observation record includes information including observation time, site geography Coordinates and subsidence observations, while collecting local DEM data in adjacent years; using terrain data to extract terrain factors, combining terrain factors to build a geospatial weight matrix, and screening auxiliary variables for subsidence prediction; using the GM(1,3) model for initial prediction, using BP The neural network corrects the prediction error of GM(1.3), and obtains the accurate settlement prediction value. The present invention utilizes subsidence observation historical data and topographic data, and integrates GM(1,3) model and BP neural network to realize the prediction of urban land subsidence, and the accuracy reaches millimeter level, which can be comparable to the results of precise level observation. The prevention and control of subsidence has important practical significance and can improve the efficiency of subsidence disaster prevention and control.

Description

Prediction of Urban Land Subsidence Based on GM(1,3) Model Combining Terrain Factor and Neural Network method

technical field

The invention relates to the technical field of land subsidence prediction, and is a method for accurately predicting urban land subsidence by using historical data of urban deformation observation and using a GM(1,3) model combining geographic space weight matrix and neural network technology.

Background technique

Land subsidence, also known as subsidence or ground subsidence, is a geological phenomenon in which the soil below the surface is compressed under the influence of natural or human factors, resulting in a decrease in local surface elevation. Affected by groundwater, natural gas and oil extraction and urban development, land subsidence has occurred frequently around the world in recent years. In the "Regulations on the Prevention and Control of Geological Hazards" issued in 2003, the phenomenon of land subsidence was defined as a "slowly changing geological hazard". According to statistics, as of 2012, more than 50 large and medium-sized cities in my country have experienced land subsidence disasters. Among them, the developed coastal areas are high-incidence areas of land subsidence, and the area of urban areas with a total land subsidence of more than 200 mm nationwide is as high as 7.9%. million square kilometers. In the past 40 years, my country has suffered losses of up to 300 billion yuan due to land subsidence. Land subsidence not only affects the sustainable development of society, but also threatens the safety of human life. Therefore, the research on model prediction of land subsidence has become a hot issue at present.

SUMMARY OF THE INVENTION

The urban land subsidence prediction method combining the terrain factor and the GM (1,3) model of the present invention is a method using the GM (1,3) prediction model and the geographic space weight matrix fusion technology, using the historical data of subsidence observation, using the GM ( 1,3) The fusion method of the model and the variable screening of the geospatial weight matrix to achieve the effect of accurately predicting the future subsidence trend.

The invention provides an urban land subsidence prediction method combining terrain factors and GM(1,3) model, comprising the following steps:

Step 1), collect the historical observation data of the site to be predicted and a number of nearby sites, each observation record includes information including observation time, site geographic coordinates and subsidence observations; simultaneously collect local adjacent years DEM data;

Step 2), terrain factor extraction, including extracting the elevation value and slope value of each observation site through DEM data and the coordinates of each observation site;

Step 3), the construction of the geospatial weight matrix and the screening of auxiliary variables, including the construction of the geospatial weight matrix according to the terrain factor using the inverse distance weight method, and the auxiliary variable is to filter out the weight corresponding to each site from the geospatial weight matrix The largest two sites get;

Step 4), use the GM(1,3) model for initial prediction, including using the geospatial weight matrix to construct the GM(1,3) model, and carry out rough prediction of settlement;

Step 5), correcting the precise prediction of the neural network fitting error, including fitting the predicted value with the GM(1,3) rough prediction model, and subtracting the true value of the data from the fitted value to obtain the prediction error value of the corresponding period of the original data; Input the error value into the BP neural network to obtain the fitting function of the prediction error, and then superimpose the predicted error and the real prediction value to obtain a more accurate prediction result.

Moreover, in step 1), the historical subsidence observation data of the predicted stations are collected for not less than at least 5 stations.

And, step 3) implementation is as follows,

First, denote the four terrain factors of p stations as (x, y, slope, H) _i , i=1, 2, 3....p, where x represents the x-coordinate of the site, y represents the y-coordinate of the site, slope represents the slope of the site, and H represents the initial elevation value of the site; in the Euclidean space composed of the above four variables, ) calculates the Euclidean distance between any two sites;

Secondly, the inverse distance weight method is used to construct the geospatial weight matrix R, and the geospatial weight value between any two stations is obtained;

Finally, arrange the weight values of the corresponding rows of the spatial weight matrix R for any station, and take the two stations with the largest weight values as auxiliary variables for prediction.

And, step 4) implementation mode is as follows,

First, take the historical elevation observation values of the observation station to be predicted for n consecutive years as the zero-order cumulative sequence to establish the original sequence of the primary variable, and use the consecutive n-year historical observation sequence of the two auxiliary stations after screening by the geospatial weight matrix to establish the auxiliary variable sequence ;

Secondly, by solving the GM(1,3) equation, the settlement prediction curve is obtained;

Finally, the original sequence of the main variable is restored, and the restored predicted value is obtained as the rough prediction result.

Moreover, in step 5), the characteristic index of the neural network preferably has:

①The BP neural network has a total of one hidden layer.

②The number of hidden layer nodes of BP neural network is 12.

③The delay variable of BP neural network is 3.

The method of the invention has the characteristics of good physical explanation for the prediction of urban land subsidence, and the accuracy reaches the millimeter level, which is comparable to the result of precise level observation, and has important practical significance for the prevention and control of urban land subsidence. The invention has the following positive effects: the prediction accuracy of urban area land subsidence is high, thereby improving the accuracy of prevention and control of land subsidence disasters, and at the same time, it has good physical meaning and is beneficial to the exploration of urban land subsidence laws.

Description of drawings

FIG. 1 is a flowchart of an embodiment of the present invention.

Detailed ways

The technical solutions of the present invention will be specifically described below with reference to the accompanying drawings and embodiments.

The present invention utilizes the historical data of settlement observation and adopts the fusion method of GM(1,3) model and the variable screening of geographic space weight matrix, so as to achieve the effect of accurately predicting the future settlement trend. Thereby, the accuracy of subsidence prediction is improved, and the efficiency of subsidence disaster prevention and control is improved. Specifically, it includes: (1) collection of historical subsidence observation data, (2) terrain factor extraction, (3) geospatial weight matrix construction and auxiliary variable screening, (4) GM(1,3) model construction, (5) BP neural network Network correction, (6) settlement value prediction.

Referring to FIG. 1, the embodiment provides a GM(1,3) model urban land subsidence prediction method combining terrain factors and neural networks, including the following steps:

Step 1) Input the subsidence observation data of the site to be predicted

During specific implementation, historical observation data of the site to be predicted and at least 5 nearby sites can be collected in advance, and each observation record contains at least the following information: observation time, site geographic coordinates (latitude and longitude), and settlement observation value. At the same time, it is necessary to collect DEM data of local neighboring years.

In the application example of this embodiment, the annual elevation observation data from 1999 to 2016 of 12 stations in Haiyan County, Zhejiang Province is selected. 1) Attribute information such as the location and name of the subsidence observation point: the location is the CGCS2000 latitude and longitude coordinates, and the name is Chinese name of the observation site; 2) Subsidence observation point of subsidence observation point from 1999 to 2016: the unit of observation value is meter, and the precision is millimeter. The filtered data samples are shown in Table 1:

Table 1 Example of subsidence data in Haiyan County, Zhejiang Province

Step 2) Terrain factor extraction:

In this step, the DEM data and the coordinates of each observation site are used to extract the elevation value and slope value of each observation site.

In the application example of the embodiment, according to the 30m resolution ASTER DEM data in Haiyan area in 1999, ArcGIS was used to extract the slope and initial elevation value of the corresponding site in this year, and the X coordinate defined by the Gauss-Kruger coordinate system of the site, The Y coordinates together constitute the terrain factor data, and the results are shown in Table 2.

Table 2 Topographic factor data of corresponding sites

Step 3) Construction of Geospatial Weight Matrix:

In the embodiment, the geospatial weight matrix is constructed using the inverse distance weighting method according to the terrain factor; the auxiliary variables are obtained by filtering out the two sites with the largest weight corresponding to each site from the geospatial weight matrix.

The four terrain factors of p stations are recorded as (x, y, slope, H) _i , i=1, 2, 3....p, where x represents the X coordinate of the station, y represents the Y coordinate of the station, and slope represents the Site slope, H represents the initial site elevation value. In the Euclidean space composed of the above four variables, the Euclidean distance of any two stations i, j in this space is recorded as D _ij :

Step 3.1, using the inverse distance weight method to construct the geospatial weight matrix R:

The geospatial weight value R _ij between any two stations can be calculated by the following formula:

The spatial weight matrix obtained by using the experimental data is shown in Table 3:

Table 3 Examples of Geospatial Weight Matrix Construction

0 1 2 3 4 5 6 7 8 9 10 11 x0 0.51 0.84 1.04 0.90 1.15 1.05 0.94 0.93 1.39 0.64 1.00 x1 0.51 0.59 0.73 0.50 0.90 0.64 0.64 0.56 0.92 0.32 0.65 x2 0.84 0.59 0.73 0.39 0.33 0.42 0.85 0.68 0.79 0.30 0.47 x3 1.04 0.73 0.73 0.39 0.95 0.47 1.25 1.10 1.02 0.58 0.29 x4 0.90 0.50 0.39 0.39 0.63 0.17 0.91 0.74 0.71 0.29 0.17 x5 1.15 0.90 0.33 0.95 0.63 0.57 1.00 0.82 0.77 0.63 0.67 x6 1.05 0.64 0.42 0.47 0.17 0.57 0.95 0.76 0.60 0.42 0.21 x7 0.94 0.64 0.85 1.25 0.91 1.00 0.95 0.20 0.80 0.78 1.08 x8 0.93 0.56 0.68 1.10 0.74 0.82 0.76 0.20 0.63 0.64 0.91 x9 1.39 0.92 0.79 1.02 0.71 0.77 0.60 0.80 0.63 0.84 0.81 x10 0.64 0.32 0.30 0.58 0.29 0.63 0.42 0.78 0.64 0.84 0.42 x11 1.00 0.65 0.47 0.29 0.17 0.67 0.21 1.08 0.91 0.81 0.42

Step 3.2, Screening of auxiliary variables of any site: for any site i, arrange the weight values of the i-th row (R _i1 , R _i2 , ..., R _ip ) of the spatial weight matrix R, and take the two sites with the largest weight values, namely auxiliary variable for prediction.

The two sites with the largest weights corresponding to each site are screened out from the geospatial weight matrix, which can be used as auxiliary predictors for GM(1,3) prediction. The corresponding tables of the main variables and auxiliary variables obtained from the experimental data are as follows shown in Table 4.

Table 4 Correspondence table of main variables and auxiliary variables

Observatory ID auxiliary variable 1 auxiliary variable 2 x0 x1 x10 x1 x4 x10 x2 x5 x10 x3 x4 x10 x4 x6 x11 x5 x2 x6 x6 x4 x11 x7 x1 x8 x7 x1 x7 x9 x6 x8 x10 x2 x4 x11 x4 x6

Step 4) Use the GM(1,3) model for initial prediction:

This step uses the geospatial weight matrix to construct a GM(1,3) model for rough prediction of settlement.

In order to carry out rough prediction by introducing the GM(1,3) of the geospatial weight matrix, this step firstly establishes the original sequence of main variables based on the historical elevation observations of the observation station to be predicted for n consecutive years (ie, the zero-order cumulative sequence), and then uses the geospatial data to obtain the original sequence of primary variables. After the weight matrix screened out the consecutive n-year historical observation sequences of the two auxiliary stations, an auxiliary variable sequence is established. Secondly, by solving the GM(1,3) equation, the settlement prediction curve can be obtained; The predicted value of is the predicted coarse prediction result.

The specific steps of the prediction in the embodiment are as follows: the original sequence of the main variable is the historical elevation observation value of the observation site to be predicted for consecutive n years (ie, the zero-order cumulative sequence), which is recorded as H ₁ ⁽⁰⁾ = (h ₁ ⁽⁰⁾ (1) ,h ₁ ⁽⁰⁾ (2),h ₁ ⁽⁰⁾ (3),…,h ₁ ⁽⁰⁾ (n)), the auxiliary variable sequence is the consecutive n of two auxiliary stations filtered by the geospatial weight matrix The annual historical observation sequences are recorded as:

H ₂ ⁽⁰⁾ = (h ₂ ⁽⁰⁾ (1),h ₂ ⁽⁰⁾ (2),h ₂ ⁽⁰⁾ (3),…,h ₂ ⁽⁰⁾ (n))

H ₂ ⁽⁰⁾ = (h ₃ ⁽⁰⁾ (1),h ₃ ⁽⁰⁾ (2),h ₃ ⁽⁰⁾ (3),…,h ₃ ⁽⁰⁾ (n))

Step 4.1, follow the steps below to solve the GM(1,3) equation to get the prediction curve:

①Calculation of the first-order accumulation generation sequence: the first-order accumulation generation sequence of the main variable H ₁ and the two auxiliary variables H ₂ and H ₃ is recorded as H ₁ ⁽¹⁾ , H ₂ ⁽¹⁾ , H ₃ ⁽¹⁾ , then :

Hi ⁽¹⁾ = ( _hi ⁽¹⁾ (1), _hi ⁽¹⁾ (2), _hi ⁽¹⁾ (3),…, _hi ⁽¹⁾ (n)) _i =1,2 ,3

Among them, h _i ⁽¹⁾ (1), h _i ⁽¹⁾ (2), h _i ⁽¹⁾ (3),…, h _i ⁽¹⁾ (n) represent the first, second, and third observation sites to be predicted, respectively. 3,…, first-order cumulative value of n years.

② Calculate the adjacent mean sequence Z ₁ ⁽¹⁾ of the first-order cumulative sequence of the main variable sequence, where the k-th value is

③Construct the GM(1,3) whitening differential equation, the equation form is as follows: where t refers to the number of observation periods

其中Solve the middle set:

in

Y=[H ₁ ⁽⁰⁾ (2),H ₁ ⁽⁰⁾ (3),...,H ₁ ⁽⁰⁾ (n)],

Y、B等为求解方程的中间参数。Among them, a, b ₁ , b ₂ ,

Y, B, etc. are the intermediate parameters for solving the equation.

④ Discretize and solve the equation, and obtain the solution of the time response sequence equation as

为真实H₁ ⁽¹⁾(k+1)的估算值，e为自然对数。in,

is an estimate of the true H ₁ ⁽¹⁾ (k+1), and e is the natural logarithm.

该序列即为预测结果。Step 4.2, restore the original sequence to obtain the restored predicted value

This sequence is the prediction result.

表示还原后的预测值，即

in,

represents the predicted value after restoration, that is,

Step 5) Neural network fitting error correction and accurate prediction: The present invention uses BP neural network to correct the prediction error of GM (1.3) to obtain accurate settlement prediction value.

The error between the restoration result and the real value is input into the time series neural network. The embodiment adopts the BP neural network. Preferably, the number of hidden layers is 1, the number of hidden layer nodes is 12, and the delay variable is set to 3, and the neural network algorithm is used for this network structure. After sample training, the fitting function of the prediction error of GM(1,3) can be obtained. The fitting function is superimposed with the prediction sequence to obtain the final high-precision prediction value.

After GM(1,3) rough prediction, the GM(1,3) model fitting prediction value of the original sequence can be obtained, and the prediction error value of the corresponding period of the original data can be obtained by subtracting the true value of the data from the fitting value . Input the error value into the BP neural network to obtain the fitting function of the prediction error, and then superimpose the predicted error with the real prediction value to obtain a more accurate prediction result. The specific steps are as follows:

①Assuming that the total number of periods to be predicted is T, use the GM(1,3) prediction model combined with the terrain factor to fit the T period data, and the error between the T period predicted value and the T period can be obtained,

② After inputting the error statistical sequence into the time series neural network, the time series function of the error can be fitted to obtain the fitting prediction model of the error.

③ Accumulate the prediction function of the neural network error prediction model and the GM(1,3) prediction model combined with the terrain, that is, a more accurate settlement prediction result can be obtained.

During specific implementation, the above process can be automatically run by using computer software technology. The system apparatus for operating the method should also be within the scope of protection of the present invention.

Obviously, the above-mentioned embodiments of the present invention are only examples for clearly illustrating the present invention, rather than limiting the embodiments of the present invention. Various changes and modifications may be made to the embodiments without departing from the spirit and scope of the embodiments of the present invention. Thus, provided that these modifications and variations of the embodiments of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include such modifications and variations.

Claims (5)

1. an urban land subsidence prediction method combining terrain factor and GM (1,3) model, is characterized in that: comprise the following steps, Step 1), collect the historical observation data of the site to be predicted and a number of nearby sites, each observation record includes information including observation time, site geographic coordinates and subsidence observations; simultaneously collect local adjacent years DEM data; Step 2), terrain factor extraction, including extracting the elevation value and slope value of each observation site through DEM data and the coordinates of each observation site; Step 3), the construction of the geospatial weight matrix and the screening of auxiliary variables, including the construction of the geospatial weight matrix according to the terrain factor using the inverse distance weight method, and the auxiliary variable is to filter out the weight corresponding to each site from the geospatial weight matrix The largest two sites get; Step 4), using the GM(1,3) model for initial prediction, including using the geospatial weight matrix to construct the GM(1,3) model for rough prediction of settlement; Step 5), correcting the precise prediction of the neural network fitting error, including fitting the predicted value with the GM(1,3) rough prediction model, and subtracting the true value of the data from the fitted value to obtain the prediction error value of the corresponding period of the original data; Input the error value into the BP neural network to obtain the fitting function of the prediction error, and then superimpose the predicted error and the real prediction value to obtain a more accurate prediction result. 2. a kind of urban land subsidence prediction method combining terrain factor and GM (1,3) model according to claim 1, it is characterized in that: in step 1), collect and predict site historical subsidence observation data not less than at least 5 sites. 3. a kind of urban land subsidence prediction method combining terrain factor and GM (1,3) model according to claim 1, is characterized in that: step 3) realization mode is as follows, First, denote the four terrain factors of p stations as (x, y, slope, H) _i , i=1, 2, 3....p, where x represents the x-coordinate of the site, y represents the y-coordinate of the site, slope represents the slope of the site, and H represents the initial elevation value of the site; In the Euclidean space composed of the above four variables, ) calculates the Euclidean distance between any two stations; Secondly, the inverse distance weight method is used to construct the geospatial weight matrix R, and the geospatial weight value between any two stations is obtained; Finally, arrange the weight values of the corresponding rows of the spatial weight matrix R for any station, and take the two stations with the largest weight values as auxiliary variables for prediction. 4. a kind of urban land subsidence prediction method combining terrain factor and GM (1,3) model according to claim 1, is characterized in that: step 4) realization mode is as follows, First, take the historical elevation observation values of the observation station to be predicted for n consecutive years as the zero-order cumulative sequence to establish the original sequence of the primary variable, and use the consecutive n-year historical observation sequence of the two auxiliary stations after screening by the geospatial weight matrix to establish the auxiliary variable sequence ; Secondly, by solving the GM(1,3) equation, the settlement prediction curve is obtained; Finally, the original sequence of the main variable is restored, and the restored predicted value is obtained as the rough prediction result. 5. a kind of urban land subsidence prediction method combining terrain factor and GM (1,3) model according to claim 1 or 2 or 3 or 4, is characterized in that: in step 5), the characteristic index of neural network is preferably Have: ①The BP neural network has a total of one hidden layer. ②The number of hidden layer nodes of BP neural network is 12. ③The delay variable of BP neural network is 3.

Images