AU2021104794A4 - Stage-wise stochastic inversion identification method of aquifer structure based on deep learning - Google Patents
Stage-wise stochastic inversion identification method of aquifer structure based on deep learning Download PDFInfo
- Publication number
- AU2021104794A4 AU2021104794A4 AU2021104794A AU2021104794A AU2021104794A4 AU 2021104794 A4 AU2021104794 A4 AU 2021104794A4 AU 2021104794 A AU2021104794 A AU 2021104794A AU 2021104794 A AU2021104794 A AU 2021104794A AU 2021104794 A4 AU2021104794 A4 AU 2021104794A4
- Authority
- AU
- Australia
- Prior art keywords
- stage
- aquifer
- aquifer structure
- model
- solute transport
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Ceased
Links
- 238000000034 method Methods 0.000 title claims abstract description 49
- 238000013135 deep learning Methods 0.000 title claims abstract description 23
- 238000004088 simulation Methods 0.000 claims abstract description 79
- 238000012549 training Methods 0.000 claims abstract description 71
- 238000002910 structure generation Methods 0.000 claims abstract description 38
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 claims description 35
- VWDWKYIASSYTQR-UHFFFAOYSA-N sodium nitrate Chemical compound [Na+].[O-][N+]([O-])=O VWDWKYIASSYTQR-UHFFFAOYSA-N 0.000 claims description 9
- 230000006870 function Effects 0.000 claims description 7
- 238000005516 engineering process Methods 0.000 abstract description 8
- 238000006467 substitution reaction Methods 0.000 abstract description 8
- 230000007704 transition Effects 0.000 abstract description 4
- 230000008569 process Effects 0.000 description 11
- 238000010586 diagram Methods 0.000 description 4
- 239000003673 groundwater Substances 0.000 description 4
- 238000010606 normalization Methods 0.000 description 4
- 238000005070 sampling Methods 0.000 description 4
- 238000013528 artificial neural network Methods 0.000 description 3
- 239000004927 clay Substances 0.000 description 3
- 238000004364 calculation method Methods 0.000 description 2
- 238000005553 drilling Methods 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000012545 processing Methods 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 238000012795 verification Methods 0.000 description 2
- 208000037170 Delayed Emergence from Anesthesia Diseases 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 238000013136 deep learning model Methods 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 239000003344 environmental pollutant Substances 0.000 description 1
- 238000013508 migration Methods 0.000 description 1
- 230000005012 migration Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000001537 neural effect Effects 0.000 description 1
- 238000007500 overflow downdraw method Methods 0.000 description 1
- 231100000719 pollutant Toxicity 0.000 description 1
- 239000000700 radioactive tracer Substances 0.000 description 1
- 239000002901 radioactive waste Substances 0.000 description 1
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01V—GEOPHYSICS; GRAVITATIONAL MEASUREMENTS; DETECTING MASSES OR OBJECTS; TAGS
- G01V9/00—Prospecting or detecting by methods not provided for in groups G01V1/00 - G01V8/00
- G01V9/02—Determining existence or flow of underground water
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01V—GEOPHYSICS; GRAVITATIONAL MEASUREMENTS; DETECTING MASSES OR OBJECTS; TAGS
- G01V99/00—Subject matter not provided for in other groups of this subclass
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- E—FIXED CONSTRUCTIONS
- E21—EARTH OR ROCK DRILLING; MINING
- E21B—EARTH OR ROCK DRILLING; OBTAINING OIL, GAS, WATER, SOLUBLE OR MELTABLE MATERIALS OR A SLURRY OF MINERALS FROM WELLS
- E21B2200/00—Special features related to earth drilling for obtaining oil, gas or water
- E21B2200/20—Computer models or simulations, e.g. for reservoirs under production, drill bits
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Physics & Mathematics (AREA)
- Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Physics & Mathematics (AREA)
- General Life Sciences & Earth Sciences (AREA)
- Geophysics (AREA)
- Theoretical Computer Science (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Artificial Intelligence (AREA)
- Medical Informatics (AREA)
- Software Systems (AREA)
- Computer Hardware Design (AREA)
- Geometry (AREA)
- General Engineering & Computer Science (AREA)
- Hydrology & Water Resources (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a stage-wise stochastic inversion identification method of aquifer
structure based on deep learning, which comprises the following steps: collect site prior data
and set prior distribution of aquifer structure parameters; randomly generate aquifer structural
parameter samples to form training samples of the first-stage solute transport simulation
substitute model, and train the first-stage solute transport simulation substitute model; acquire
prior interval of aquifer data, invert aquifer structural parameter and acquiring posterior aquifer
structural parameters; construct and train aquifer structure generation model; construct and
train the second stage solute transport simulation alternative model; collect observation data
and invert the input parameters of aquifer structure generation model; input the input
parameters into the aquifer structure generation model to obtain the posterior aquifer structure.
According to the invention, the aquifer structure identification method is combined with the
deep learning technology, and the inversion identification of the aquifer structure is carried out
by utilizing the field and observation data, so that the uncertainty of the random simulation of
the aquifer structure is effectively reduced.
1/3
FIGURES
I First stage: Second stage:
ITraining sample Potential variableI
Update Aquifer structure
parameter I
toIcatorkriging simulationprogram I
based on transition probability model I OCAAE aquifer structure
I generation program
I I
Aquifer structure Aquifer structureI
ISubstitution simulation program T
Ifor solute transport Substitution simulation programI F for solute transportI
Aquifer structural parametersPoeta
Parameter inverse derivation program IPtnilvariableI
inversion program
No NoI
It~i. bes II - e
YYes
1 77 ye
Posterior aquifer structure UPosterior aquifer structure
Figure1I
Aquifer structure Latent Vector -iReconstructed structure
1 True:Z-p(Z
False:Z-q(Z)
Figure 2
Description
1/3
I First stage: Second stage: ITraining sample Potential variableI Update Aquifer structure parameter I
toIcatorkriging simulationprogram I based on transition probability model I OCAAE aquifer structure I generation program I I Aquifer structure Aquifer structureI
ISubstitution simulation program T Ifor solute transport Substitution simulation programI F for solute transportI Aquifer structural parametersPoeta Parameter inverse derivation program IPtnilvariableI inversion program
No NoI
It~i. besII- e YYes 1 77 ye
Posterior aquifer structure UPosterior aquifer structure
Figure1I
Aquifer structure Latent Vector -iReconstructed structure
1 True:Z-p(Z
False:Z-q(Z)
Figure 2
Stage-wise stochastic inversion identification method of aquifer structure based on
deep learning
The invention relates to the technical field of aquifer structure identification, in particular
to an aquifer structure phased random inversion identification method based on deep
learning.
Groundwater flow and solute simulation are widely used in various hydrogeology fields,
such as site pollutant migration prediction, C02 geological storage, groundwater resource
management and radioactive waste geological storage.
Accurate simulation of groundwater flow and solute depends on accurate description of
aquifer structure. Aquifer structure identification method based on geostatistical theory
can obtain aquifer structure by borehole data. However, because borehole distribution is
sparse, aquifer structure parameters and corresponding aquifer structure are uncertain. In
order to improve the accuracy of aquifer structure description, it is a very effective
method to use actual observation data and multi-source data fusion method to invert
aquifer structure parameters and their posterior structure.
In recent years, with the continuous development of deep learning technology, some
image processing technologies can be effectively applied to aquifer structure recognition,
such as generating countermeasure network, self-encoder and convolutional neural
network, etc., to improve the accuracy and speed of aquifer structure recognition. At
present, the inversion method of aquifer structure based on deep learning technology usually needs to provide a large number of aquifer structure samples with similar geostatistical characteristics to the target aquifer structure, but in practice, this sample is difficult to provide. How to make full use of the drilling data and observation data of the site to obtain effective aquifer structure samples for the training of deep learning model, so as to realize the inversion and identification of aquifer structure is a key technology to be solved urgently at present.
The purpose of the present invention is to provide a stage-wise stochastic inversion
identification method of aquifer structure based on deep learning, which solves the
problems in the prior art, and can effectively combine the aquifer structure identification
method based on geological statistics with deep learning technology, which does not need
to provide additional training samples of aquifer structure, it can directly use site drilling
data and observation data to carry out inversion identification on aquifer structure, and
can effectively reduce the uncertainty of random simulation of aquifer structure, and
provides technical support for accurate simulation of subsequent solute transport process,
thus having great practical value.
In order to achieve the above purpose, the present invention provides the following
scheme: the present invention provides a stage-wise stochastic inversion identification
method of aquifer structure based on deep learning, which comprises the following steps:
Si: Collecting site prior data and setting prior distribution of aquifer structural
parameters;
S2: Randomly generating aquifer structural parameter samples based on the prior
distribution of the aquifer structural parameters to form training samples of the first-stage solute transport simulation substitute model, and training the first-stage solute transport simulation substitute model;
S3. Collecting an initial sample of aquifer structural parameters, and inverting the aquifer
structural parameters based on the first-stage solute transport simulation substitute model
to obtain posterior aquifer structural parameters;
S4. Building and training an aquifer structure generation model based on the posterior
aquifer structure parameters;
S5. Based on the site prior data, constructing and training a second-stage solute transport
simulation substitute model;
S6. Collecting observation data, and inverting input parameters of the aquifer structure
generation model based on the trained aquifer structure generation model and the second
stage solute transport simulation substitute model;
S7. Inputting the input parameters after inversion into an aquifer structure generation
model to obtain a posterior aquifer structure.
Preferably, S2 comprises:
S2.1. Randomly generating aquifer structure parameter samples based on the prior
distribution of aquifer structure parameters, generating a first-stage aquifer structure by
using an indicator Kriging model, and obtaining a first-stage state field corresponding to
the first-stage aquifer structure to form a first-stage training sample of a first-stage solute
transport simulation substitute model;
S2.2. Normalizing the first-stage training samples;
S2.3. Constructing a first-stage solute transport simulation substitute model, and training
the first-stage solute transport simulation substitute model by using the normalized first
stage training sample.
Preferably, the S3 comprises:
S3.1. Based on the prior distribution of the aquifer structure parameters, extracting the
initial samples of the first-stage aquifer structure parameters, randomly generating the
first-stage aquifer structure, and inputting the first-stage solute transport simulation
substitute model to obtain the corresponding first-stage water head and concentration
prediction values;
S3.2. Collecting the first-stage waterhead and concentration measured values, comparing
the first-stage water head and concentration predicted values with the first-stage water
head and concentration measured values, updating the first-stage aquifer structure based
on the updated first-stage aquifer structure parameters, updating the first-stage water head
and concentration predicted values, repeating step S3.2 until preset iteration times, and
extracting the finally updated first-stage aquifer structure
Preferably, the S4 comprises:
S4.1. Generating a plurality of first-stage aquifer structure samples based on the posterior
aquifer structure parameters, wherein the first-stage aquifer structure samples are used as
training samples of an aquifer structure generation model;
S4.2. Constructing the aquifer structure generation model, and training the aquifer
structure generation model by using a plurality of first-stage aquifer structure samples.
Preferably, the S5 comprises:
S5.1. Randomly generating a plurality of potential variables based on the site prior data,
inputting the potential variables into the trained aquifer structure generation model,
obtaining a second-stage aquifer structure, generating corresponding water head and
concentration distribution fields, and establishing a second-stage training sample set of a
second-stage solute transport simulation substitute model;
S5.2. Normalizing the second-stage training samples;
S5.3, Constructing a second-stage solute transport simulation substitute model, and
training the second-stage solute transport simulation substitute model by using the
second-stage training sample.
Preferably, S6 comprises:
S6.1. Randomly generating a plurality of potential variable samples of the aquifer
structure generation model, training the aquifer structure generation model based on the
potential variable samples, obtaining a second-stage aquifer structure, and generating
corresponding second-stage water head and concentration prediction values;
S6.2. Comparing the predicted values of water head and concentration at the second stage
with the measured values of water head and concentration, updating the potential variable
samples, updating the aquifer structure at the second stage, updating the predicted values
of water head and concentration at the second stage, repeating step S6.2 until a preset
iteration number, extracting the finally updated potential variable samples, and obtaining
the input parameters of the aquifer structure generation model.
Preferably, the training of the first-stage solute transport simulation substitute model and
the training of the second-stage solute transport simulation substitute model both adopt
the Li loss function based on regularization: f1 K I~ £1I=k- kI- l WT y 9x1+2-67 K Yk=1 2
In the formula, y represents the simulation result of solute transport simulation at grid
k, yA represents the prediction result of alternative model, 0 is a trainable parameter in
alternative model and w is a regularization coefficient.
Preferably, the training of the first stage solute transport simulation substitute model and
the training of the second stage solute transport simulation substitute model adopt an
early stopping method.
Preferably, the parameter inversion in S3 and S6 adopts iteration-local update-ensemble
smoother, as shown in the following formula:
pn+1 = p + C+ ier CD)-1[d + Pk kp Cpnm(C~mn +Nite Niter* terkC - f(pn)]
In the formula, pn is the kth parameter sample after the nth iteration, CRM is the cross
covariance between the parameter sample after the n* iteration and the predicted value of
the alternative model, CmnMis the autocovariance between the predicted values of the
alternative model, Niter is the iteration times, CD represents the covariance of the
observation error, e' represents the kth observation error randomly selected, and f
represents the alternative model.
Preferably, the stage-wise stochastic inversion identification method of aquifer structure
based on deep learning according to claim 1, characterized in that the aquifer structure
generation model is constructed based on a discrete convolution-confrontation-self
encoder network, wherein the discrete convolution-confrontation-self-encoder network
adopts a discrete convolution layer and a dense residual connection structure.
The invention discloses the following technical effects:
The method can combine multi-source data, such as borehole data, water head,
concentration data, etc., can effectively combine the aquifer structure identification
method based on geological statistics with deep learning technology without providing
additional aquifer structure training samples, and can directly use site borehole data and
observation data to carry out inversion identification on aquifer heterogeneous structure,
thus reducing the uncertainty of aquifer heterogeneous structure, and significantly
improving the simulation of groundwater flow and solute simulation in the site.
In order to explain the embodiments of the present invention or the technical scheme in
the prior art more clearly, the drawings needed in the embodiments will be briefly
introduced below. Obviously, the drawings in the following description are only some
embodiments of the present invention, and for ordinary technicians in the field, other
drawings can be obtained according to these drawings without paying creative labor.
Fig. 1 is a schematic flow chart of a phased random inversion identification method for
aquifer structure proposed in an embodiment of the present invention;
Fig. 2 is a schematic diagram of the aquifer structure generation model network structure
proposed in the embodiment of the present invention;
Fig. 3 is a schematic diagram of a discrete convolution dense residual block structure
used in an embodiment of the present invention;
Fig. 4 is a schematic structural diagram of a deep discrete convolution residual network
used in the present invention in an embodiment of the present invention;
Fig. 5 is a comparison diagram between the posterior aquifer structure and the real
structure in the embodiment of the present invention.
The following will clearly and completely describe the technical solutions in the
embodiments of the present invention with reference to the drawings in the embodiments
of the present invention. Obviously, the described embodiments are only part of the
embodiments of the present invention, not all of them. Based on the embodiments of the
present invention, all other embodiments obtained by ordinary technicians in the field
without creative labor belong to the scope of protection of the present invention.
In order to make the above objects, features and advantages of the present invention more
obvious and easy to understand, the present invention will be further explained in detail
with reference to the drawings and specific embodiments.
The invention provides a stage-wise stochastic inversion identification method of aquifer
structure based on deep learning, which is shown in Figures 1-5. This embodiments are
hypothetical embodiments. Based on five virtual borehole data, a target aquifer structure
is generated by presetting a group of aquifer structure parameters. Based on this aquifer
structure, a tracer GEOST scene is simulated by solute transport simulation software, and
corresponding water head and concentration distribution fields are generated. A total of
168 water head and concentration data from 28 different positions and different depths in
time periods are extracted as observation data, and the aquifer structure is inverted
based on these observation data.
Si. Collect the site prior data of the test site and set the prior distribution of aquifer
structural parameters. The obtained prior distribution and true values of aquifer structural
parameters are shown in Table 1:
Table 1
Clay volume Elongation of clay in x Extension length of clay in ratio direction Z direction Prior interval [0.35,0.45] [15,80] [4,5] Actual value 0.4 20 4.5 Inversion 0.4079 19.84 4.56 value S2. Randomly generate aquifer structural parameter samples according to the prior
distribution of aquifer structural parameters. in this embodiment, 10,000 groups of
aquifer structural parameters are extracted by Latin hypercube sampling method, and the
extracted 10,000 groups of aquifer structural parameters are input into the indicator
Kriging model. The indicator kriging simulation is based on transition probability
analysis, and its analytical solution can be obtained by lithofacies volume ratio and
average extension length, which enables inversion of aquifer structure based on
geostatistics. The analytical solution of transition probability is as shown in formula (1)
tij(hp) = p; + (S; - p1)e -M o(Li(I-Pi)) (for i,j = 1,N) (1)
In which, tij(hp) is the probability that lithofacies is j at the interval hp from
lithofacies i in q direction, pi is the volume ratio of lithofacies i, N is the number of
lithofacies, 6og is Kronecker function, and Lig is the average extension length of
lithofacies i in # direction.
Generating a corresponding aquifer structure, namely a first-stage aquifer structure, by
analyzing the indicator kriging model, inputting the obtained first-stage aquifer structure
into a solute transport simulation model, obtaining a corresponding water head and
concentration distribution field, saving in the form of a numpy array, and establishing a
training sample of a first-stage solute transport simulation substitute model composed of
"aquifer structure-state field" samples. The aquifer structure and state field samples are regarded as picture data files, then different aquifer structures are continuously input into the neural network, and the weights of the neural network are gradually updated through reverse error propagation, so that the state field output by the neural network can be as close as possible to the state field corresponding to the input aquifer structure. After a certain number of iterations, the training of the first stage solute transport simulation substitute model is completed, so that it can achieve similar effects to the solute transport simulation model.
Normalize the obtained training samples of the substitute model for solute transport
simulation in the first stage. In this embodiment, the method of "0-1 normalization" is
adopted for processing, and the normalization calculation is shown in Formula (2):
Ck (2) Cmax-cmin
In the formula, Ck represents the normalized waterhead or concentration value at grid k,
ek represents the normalized waterhead or concentration value at grid k, cmin represents
the minimum head or concentration at grid k in training samples and cmax represents the
maximum head or concentration at grid k in training samples are expressed.
The first-stage solute transport simulation substitute model is constructed by using the
deep discrete convolution residual network as the first-stage inversion solute transport
simulation substitute model, and the first-stage solute transport simulation substitute
model is trained by the normalized training samples of the first-stage solute transport
simulation substitute model. In the training process, the L1 loss function based on
regularization is adopted, and its expression is shown in formula (3):
£1= illy -fk1+ OTO (3)
In the formula, yA represents the simulation result of solute transport simulation at grid
k, yk represents the prediction result of forward substitution model, 0 is a trainable
parameter in forward substitution model and w is the regularization coefficient;
In the training process, the exponential learning rate attenuation strategy is adopted, so
that the learning rate can be reduced smoothly with the increase of iteration times, which
can effectively prevent over-fitting phenomenon, and its attenuation expression is shown
in formula (4):
RN = R0 * e" (4)
Where, RO represents the initial learning rate is expressed, RN represents the learning
rate after the nth iteration, and e is the decay coefficient.
The early stopping method is adopted in the training process, that is, the optimal network
model parameters are saved in the training process, and the advantages and disadvantages
of the network model parameters are measured by the root mean square error (RMSE)
between the predicted results of the alternative model and the real results, and its
expression is shown in Formula (5):
eRMSE - =IYk Vk||' (5)
Wherein, y represents the simulation result of solute transport simulation at grid k,
Ykrepresents the prediction result of forward substitution model, k is the total grid
number.
In this embodiment, 80% of the training samples of the first-stage solute transport
simulation substitute model are randomly selected as the training set and the remaining
% as the verification set. The initial learning rate and decay coefficient of the training samples of the first-stage solute transport simulation substitute model are set to 0.0005 and 0.99, respectively, and the regularization coefficient of LI loss function is set to
1 x 10s, the training time is 500.
The alternative model of solute transport simulation in the first stage in this embodiment
is established based on the idea of "image conversion", that is, the aquifer structure, water
head and concentration distribution field are regarded as image pixel data, and the
prediction process of the alternative model is to convert the aquifer structure image pixel
data into water head and concentration image pixel data. In addition, the alternative
model of solute transport simulation in the first stage adopts discrete convolution layer
and dense residual connection structure, which can reduce the GPU memory occupied in
network training and improve the prediction accuracy and training speed of the
alternative model.
S3. Collect aquifer data, extract 2,000 groups of aquifer structure parameters as initial
samples through Latin hypercube sampling, and input the initial samples into an indicator
Kriging model to generate a corresponding second-stage aquifer structure, and input the
second-stage aquifer structure into the first-stage solute transport simulation substitute
model to obtain first waterhead and concentration prediction values;
Collect the measured values of water head and concentration, compare the first predicted
values of water head and concentration with the measured values of water head and
concentration, and update the aquifer structure parameters in the first stage by using
iterative-local update-ensemble smoother, which can assimilate all the observed data at
one time and update the parameter samples by iteration. Compared with ensemble
Kalman filter method, it has faster inversion speed, and its parameter update equation is
shown in Formula (6):
pk = pM +C +(C +Niter*CD) [d + Nter * e f (p )](6)
In which: p' is the kt parameter sample after the nth iteration, CRM represents the cross
covariance between the parameter sample after the nth iteration and the predicted value of
the alternative model, CMM represents the autocovariance between the predicted values
of the alternative model, Niter represents the iteration times, CD is the covariance of the
observationerror,en is the kth observation error randomly selected, and f represents the
alternative model.
The updated first-stage aquifer structure parameters are input into the indicator Kriging
model to obtain the updated first-stage aquifer structure. The updated first-stage aquifer
structure is input into the first-stage solute transport simulation substitute model to obtain
the updated predicted values of water head and concentration, and iterative updating is
carried out continuously until the set number of iterations is reached. After iteration, the
final first-stage aquifer structure parameters are obtained, that is, the posterior aquifer
structure parameters of the first stage are obtained, and the inversion of the first-stage
aquifer structure parameters is completed. In this embodiment, the number of iterations is
set to 40.
S4. The final posterior aquifer structure parameters are input into the indicator kriging
model, and several first-stage aquifer structure samples are generated. In this
embodiment, the number of first-stage aquifer structure samples is 20,000, and the first
stage aquifer structure samples are used as training sets of aquifer structure generation
model, of which 80% are used as training sets and the remaining 20% as test sets. The aquifer structure generation model is established by using discrete convolution countermeasure self-encoder, and the aquifer structure generation model is trained by using training set, and the training number is set to 50 times.
The aquifer structure generation model in this embodiment is established based on
discrete convolution - countermeasure - self-encoder network, which combines the
advantages of generating countermeasure network and self-encoder. In addition, the
network adopts discrete convolution layer and dense residual connection structure, which
makes the aquifer structure generation model adopted by this method have faster training
speed and stronger learning ability, and can generate more accurate aquifer structure. The
establishment and training of this model are realized under the framework of pytorch.
S5. Randomly generate a plurality of potential variables according to the site information.
In this embodiment, 10,000 groups of potential variables are extracted by using Latin
hypercube sampling method, and the potential variables are input into the trained aquifer
structure generation model, and the solute transport simulation program is called to obtain
the second-stage aquifer structure and generate corresponding water head and
concentration distribution fields, which are stored in the form of numpy array. A training
sample set of the second-stage solute transport simulation substitute model is established.
During the training process, 80% of the training sample set of the second-stage solute
transport simulation substitute model is randomly selected as the training set and the
remaining 20% as the verification set.
The training samples in the second stage are normalized by the "0-1 normalization"
method, and the normalization calculation is shown in Formula (2):
Ck (2)
In the formula, Ck represents the normalized head or concentration value at grid k, ek
represents the normalized head or concentration value at grid k, cmn represents the
minimum head or concentration at grid k in training samples and the maximum head or
concentration at grid k, cmax indicates the maximum value of water head or
concentration at grid k in the training sample.
The second-stage solute transport simulation substitute model is constructed by using the
deep discrete convolution residual network, and the second-stage solute transport
simulation substitute model is trained by using the normalized second-stage training
samples. In the second stage, the initial learning rate and attenuation coefficient of solute
transport simulation substitute model were set to 0.0005 and 0.99 respectively, and the
regularization coefficient of L loss function is set to 1 x 10-s, the training time is 500.
In the training process, the L1 loss function based on regularization is adopted, and its
expression is shown in formula (3):
£1I = 1||yk K k= -- fYkj j+wi9" 2 (3)
In the formula, y represents the simulation result of solute transport simulation at grid
k, yArepresents the prediction result of forward substitution model, 0 is a trainable
parameter in forward substitution model and w represents regularization coefficient;
In the training process, the exponential learning rate attenuation strategy is adopted, so
that the learning rate can be reduced smoothly with the increase of iteration times, which
can effectively prevent over-fitting phenomenon, and its attenuation expression is shown
in formula (4):
RN = Ro * en (4)
Wherein, RO represents the initial learning rate, RN represents the learning rate after the
nth iteration, and e is the decay coefficient.
The early stopping method is adopted in the training process, that is, the optimal network
model parameters are saved in the training process, and the advantages and disadvantages
of the network model parameters are measured by the root mean square error (RMSE)
between the predicted results of the alternative model and the real results, and its
expression is shown in Formula (5):
eRMSE 1 jk Zk- YkI|2 (5)
Where, y represents the simulation result of solute transport simulation at grid k and k
represents the prediction result of forward substitution model, and k is the total grid
number.
S6. Randomly generate a plurality of potential variable samples of aquifer structure
generation models. In this embodiment, 2,000 groups of potential variable samples of
aquifer structure generation models are extracted by Latin hypercube sampling as initial
samples, which are input into the trained aquifer structure generation model to obtain the
second-stage aquifer structure, and the second-stage aquifer structure is input into the
second-stage solute transport simulation substitute model to generate corresponding
second-stage water head and concentration prediction values;
Compare the predicted values of water head and concentration with the measured values
of water head and concentration, and update the potential variable samples by using
iterative-local update-ensemble smoother. Iterative-local update-ensemble smoother can assimilate all the observed data at one time and update the parameter samples by iteration. Compared with ensemble Kalman filter, it has faster inversion speed, and its parameter update equation is shown in Equation (6):
P+ p +C (C +Nter * CD)-1[d + NIter * e- ) (6)
In which: p' is the kth parameter sample after the nth iteration, CM represents the cross
covariance between the parameter sample after the nth iteration and the predicted value of
the alternative model, CM represents the autocovariance between the predicted values
of the alternative model, Niter represents the iteration times, CD represents the
covariance of the observation error, e' is the kthobservation error randomly selected,
and f represents the alternative model.
Input the updated potential variable samples into aquifer structure generation model,
update the second-stage aquifer structure, input the updated second-stage aquifer
structure into the second-stage solute transport simulation substitute model, update the
corresponding second-stage water head and concentration prediction values, repeat this
step until the preset iteration times, and extract the finally updated potential variable
samples to obtain the input parameters of aquifer structure generation model. The number
of iterations is set to 40.
S7. Input the input parameters of the aquifer structure generation model into the aquifer
structure generation model to obtain the final posterior aquifer structure. Randomly select
one of the structures, and compare it with the target aquifer structure as shown in Figure
5. The similarity between the inverted aquifer structure and the real structure can reach
more than 95%. It can be seen from Table 1 that more accurate aquifer structure
parameters can be obtained through the first-stage inversion.
The aquifer structure phased random inversion identification method provided by the
invention effectively combines the aquifer structure identification method based on
geostatistics with the deep learning technology, and decomposes the inversion
identification problem of the aquifer structure by stages, including the first-stage
inversion and the second-stage inversion as shown in fig.1. The first-stage inversion
mainly carries out inversion identification on aquifer structure parameters, which can
reduce certain aquifer structure uncertainty after the completion of this stage. In the
second stage inversion, the aquifer structure is inverted on the basis of the aquifer
structure parameters obtained in the first stage inversion, which further reduces the
uncertainty of aquifer structure.
The above embodiments only describe the preferred mode of the invention, but do not
limit the scope of the invention. On the premise of not departing from the design spirit of
the invention, various modifications and improvements made by ordinary technicians in
the field to the technical scheme of the invention shall fall within the protection scope
determined by the claims of the invention.
Claims (10)
1. A stage-wise stochastic inversion identification method of aquifer structure based on
deep learning, characterized by comprising the following steps:
Si: collect site prior data and setting prior distribution of aquifer structural parameters;
S2: randomly generate aquifer structural parameter samples based on the prior
distribution of the aquifer structural parameters to form training samples of the first-stage
solute transport simulation substitute model, and train the first-stage solute transport
simulation substitute model;
S3: collect an initial sample of aquifer structural parameters, and invert the aquifer
structural parameters based on the first-stage solute transport simulation substitute model
to obtain posterior aquifer structural parameters;
S4: build and train an aquifer structure generation model based on the posterior aquifer
structure parameters;
S5: based on the site prior data, construct and train a second-stage solute transport
simulation substitute model;
S6: collect observation data, and invert input parameters of the aquifer structure
generation model based on the trained aquifer structure generation model and the second
stage solute transport simulation substitute model;
S7: input the input parameters after inversion into an aquifer structure generation model
to obtain a posterior aquifer structure.
2. The stage-wise stochastic inversion identification method of aquifer structure based on
deep learning according to claim 1, wherein S2 comprises:
S2.1: randomly generate aquifer structure parameter samples based on the prior
distribution of aquifer structure parameters, generate a first-stage aquifer structure by
using an indicator Kriging model, and obtain a first-stage state field corresponding to the
first-stage aquifer structure to form a first-stage training sample of a first-stage solute
transport simulation substitute model;
S2.2: normalize the first-stage training samples;
S 2.3: construct a first-stage solute transport simulation substitute model, and train the
first-stage solute transport simulation substitute model by using the normalized first-stage
training sample.
3. The stage-wise stochastic inversion identification method of aquifer structure based on
deep learning according to claim 1, wherein the step S3 comprises:
S3.1: based on the prior distribution of the aquifer structure parameters, extract the initial
samples of the first-stage aquifer structure parameters, randomly generate thefirst-stage
aquifer structure, and input the first-stage solute transport simulation substitute model to
obtain the corresponding first-stage water head and concentration prediction values;
S3.2: collect the first-stage water head and concentration measured values, compare the
first-stage water head and concentration predicted values with the first-stage water head
and concentration measured values, update the first-stage aquifer structure based on the
updated first-stage aquifer structure parameters, updating the first-stage water head and
concentration predicted values, repeating step S3.2 until preset iteration times, and
extracting the finally updated first-stage aquifer structure
4. The stage-wise stochastic inversion identification method of aquifer structure based on
deep learning according to claim 3, wherein S4 comprises:
S4.1: generate a plurality of first-stage aquifer structure samples based on the posterior
aquifer structure parameters, wherein the first-stage aquifer structure samples are used as
training samples of an aquifer structure generation model;
S4.2: construct the aquifer structure generation model, and train the aquifer structure
generation model by using a plurality of first-stage aquifer structure samples.
5. The stage-wise stochastic inversion identification method of aquifer structure based on
deep learning according to claim 4, wherein the step S5 comprises:
S5.1: randomly generate a plurality of potential variables based on the site prior data,
input the potential variables into the trained aquifer structure generation model, obtain a
second-stage aquifer structure, generate corresponding water head and concentration
distribution fields, and establish a second-stage training sample set of a second-stage
solute transport simulation substitute model;
S5.2: normalize the second-stage training samples;
S5.3: construct a second-stage solute transport simulation substitute model, and training
the second-stage solute transport simulation substitute model by using the second-stage
training sample.
6. The stage-wise stochastic inversion identification method of aquifer structure based on
deep learning according to claim 5, wherein S6 comprises:
S6.1: randomly generate a plurality of potential variable samples of the aquifer structure
generation model, train the aquifer structure generation model based on the potential
variable samples, obtaining a second-stage aquifer structure, and generate corresponding
second-stage water head and concentration prediction values;
S6.2: compare the predicted values of water head and concentration at the second stage
with the measured values of water head and concentration, update the potential variable
samples, update the aquifer structure at the second stage based on the updated potential
variable samples, update the predicted values of water head and concentration at the
second stage, repeat step S6.2 until a preset iteration number, extracting the finally
updated potential variable samples, and obtain the input parameters of the aquifer
structure generation model.
7. The stage-wise stochastic inversion identification method of aquifer structure based on
deep learning according to claim 1, characterized in that the training of the first stage
solute transport simulation substitute model and the training of the second stage solute
transport simulation substitute model both adopt Li loss function based on regularization:
1 ,k=2
In the formula, yA represents the simulation result of solute transport simulation at grid
k, pk represents the prediction result of alternative model, 0 is a trainable parameter in
alternative model and w represents the regularization coefficient.
8. The stage-wise stochastic inversion identification method of aquifer structure based on
deep learning according to claim 1, characterized in that the training of the first stage
solute transport simulation substitute model and the training of the second stage solute
transport simulation substitute model adopt an early stopping method.
9. The stage-wise stochastic inversion identification method of aquifer structure based on
deep learning according to claim 1, characterized in that the parameter inversion in S3
and S6 adopts iteration-local update-ensemble smoother, as shown in the following
formula: pk+1 =k+ CAM(CM + Niter * CD)[d + Niter * e - f(p)]
In which, p' is the kth parameter sample after the nh iteration, CRM represents the cross
covariance between the parameter sample after the nth iteration and the predicted value of
the alternative model,CM the autocovariance between the predicted values of the
alternative model, Niter represents the iteration times, CD represents the covariance of
the observation error, en represents the k' observation error randomly selected, and f
represents the alternative model.
10. The stage-wise stochastic inversion identification method of aquifer structure based
on deep learning according to claim 1, wherein the aquifer structure generation model is
constructed based on a discrete convolution - confrontation - self-encoder network,
wherein the discrete convolution - confrontation - self-encoder network adopts a discrete
convolution layer and a dense residual connection structure.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
AU2021104794A AU2021104794A4 (en) | 2021-08-02 | 2021-08-02 | Stage-wise stochastic inversion identification method of aquifer structure based on deep learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
AU2021104794A AU2021104794A4 (en) | 2021-08-02 | 2021-08-02 | Stage-wise stochastic inversion identification method of aquifer structure based on deep learning |
Publications (1)
Publication Number | Publication Date |
---|---|
AU2021104794A4 true AU2021104794A4 (en) | 2021-09-30 |
Family
ID=77857738
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
AU2021104794A Ceased AU2021104794A4 (en) | 2021-08-02 | 2021-08-02 | Stage-wise stochastic inversion identification method of aquifer structure based on deep learning |
Country Status (1)
Country | Link |
---|---|
AU (1) | AU2021104794A4 (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN117828955A (en) * | 2024-03-05 | 2024-04-05 | 山东科技大学 | Aquifer solute transport numerical simulation method and system based on scale lifting |
-
2021
- 2021-08-02 AU AU2021104794A patent/AU2021104794A4/en not_active Ceased
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN117828955A (en) * | 2024-03-05 | 2024-04-05 | 山东科技大学 | Aquifer solute transport numerical simulation method and system based on scale lifting |
CN117828955B (en) * | 2024-03-05 | 2024-05-28 | 山东科技大学 | Aquifer solute transport numerical simulation method and system based on scale lifting |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN113610945B (en) | Ground stress curve prediction method based on hybrid neural network | |
CN113687433B (en) | Bi-LSTM-based magnetotelluric signal denoising method and system | |
CN110031895B (en) | Multipoint geostatistical stochastic inversion method and device based on image stitching | |
CN111105097B (en) | Dam deformation prediction system and method based on convolutional neural network | |
CN111242377A (en) | Short-term wind speed prediction method integrating deep learning and data denoising | |
CN111967079A (en) | Foundation pit deformation prediction method based on improved artificial bee colony algorithm and BP neural network | |
AU2021104794A4 (en) | Stage-wise stochastic inversion identification method of aquifer structure based on deep learning | |
CN113470316B (en) | Debris flow monitoring and early warning method based on self-coding single classification model | |
CN115327616B (en) | Automatic positioning method for mine microseism focus driven by massive data | |
CN115293197A (en) | Borehole strain data anomaly detection method based on long-term and short-term memory network | |
CN115169543A (en) | Short-term photovoltaic power prediction method and system based on transfer learning | |
CN112949089A (en) | Aquifer structure inversion identification method based on discrete convolution residual error network | |
CN113537354B (en) | Aquifer structure staged stochastic inversion identification method based on deep learning | |
CN117633512A (en) | Reservoir porosity prediction method based on one-dimensional convolution gating circulation network | |
CN117557100A (en) | Deep foundation pit excavation induced adjacent building risk assessment method based on deep learning | |
CN113920124B (en) | Brain neuron iterative segmentation method based on segmentation and error guidance | |
CN115860165A (en) | Neural network basin rainfall runoff forecasting method and system considering initial loss | |
CN115032682A (en) | Multi-station seismic source parameter estimation method based on graph theory | |
AU2021102281A4 (en) | Inverse identification method of aquifer structure based on octave convolution residual network | |
Li et al. | A gated recurrent unit network model for predicting open channel flow in coal mines based on attention mechanisms | |
CN117454762B (en) | Coal-penetrating tunnel face gas concentration prediction method of Markov-neural network | |
CN117148433B (en) | Microseism P wave S wave classification method based on GPT-L model | |
CN113627761B (en) | Parallel evaluation method for geotechnical engineering water inrush probability prediction | |
CN116796622A (en) | Wind power output scene generation method based on conditional importance weighted self-encoder | |
CN117151486A (en) | Park carbon emission prediction method and system based on graph neural network |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
FGI | Letters patent sealed or granted (innovation patent) | ||
MK22 | Patent ceased section 143a(d), or expired - non payment of renewal fee or expiry |