CN107728231A - One kind prediction nuclear magnetic resonance log T2 T2The method of distribution - Google Patents
One kind prediction nuclear magnetic resonance log T2 T2The method of distribution Download PDFInfo
- Publication number
- CN107728231A CN107728231A CN201710969507.1A CN201710969507A CN107728231A CN 107728231 A CN107728231 A CN 107728231A CN 201710969507 A CN201710969507 A CN 201710969507A CN 107728231 A CN107728231 A CN 107728231A
- Authority
- CN
- China
- Prior art keywords
- mrow
- distribution
- data
- msub
- model
- 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.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01V—GEOPHYSICS; GRAVITATIONAL MEASUREMENTS; DETECTING MASSES OR OBJECTS; TAGS
- G01V11/00—Prospecting or detecting by methods combining techniques covered by two or more of main groups G01V1/00 - G01V9/00
Landscapes
- Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Life Sciences & Earth Sciences (AREA)
- General Physics & Mathematics (AREA)
- Geophysics (AREA)
- Geophysics And Detection Of Objects (AREA)
Abstract
The invention discloses one kind prediction nuclear magnetic resonance log T2 T2The method of distribution, including screening determine that related conventional logging data predict T as step 52The input data of distribution;Set classification parameter, the existing T of quantitative classification2Distribution curve is as model training in step 5 and the input data of test;Existing T is fitted using multiple normal distribution curves2Distribution, obtains the characterization parameter of normal distribution, and model training in step 5 and the output data of test are used as after assessing fitting accuracy by the use of coefficient correlation;Processing early stage is carried out to step 1 conventional logging data, step 2 classification parameter, step 3 characterization parameter;By all data sorting and groupings of step 4, for the training and test of artificial nerve network model, forecast model is obtained;With coefficient correlation and standard root mean square deviation assessment prediction model exactness, T is predicted2Distribution.The present invention obtains conventional logging data and T using artificial nerve network model2Distribution relation model, to T2The accurate prediction of distribution.
Description
Technical field
The present invention relates to reservoir physics and log data electric powder prediction, particularly a kind of prediction nuclear magnetic resonance log is horizontal
To relaxation time T2The method of distribution.
Background technology
Magnetic resonance well logging is highly important a kind of logging mode in petroleum industry well logging.Surveyed using nuclear magnetic resonance
Well, the significant datas such as the distribution of pores of subsurface reservoir, effecive porosity, permeability, irreducible water saturation can be accurately obtained.Its
In, the T2 T of nuclear magnetic resonance log2Distribution profile is one of primary log data of nuclear magnetic resonance log, and it can essence
The reservoir pore space distribution of underground is really characterized, without being influenceed by Rock Species and property.But in practical logging, people seldom use
To nuclear magnetic resonance log, it is primarily due to NMR logging instrument and monopolized for a long time by the oilfield service companies of foreign countries, and price is very
Costliness, and inconvenient maintenance are too high so as to carry out the practical application cost of nuclear magnetic resonance log.
At present, people can be by establishing artificial nerve network model, to predict the conventional porosity in stratum, infiltration
The single parameter such as rate, water saturation, it can also utilize what artificial nerve network model prediction nuclear magnetic resonance log can obtain
Some related datas, including effecive porosity, effective permeability, irreducible water saturation, T2The single parameter such as logarithmic mean.It is but single
The calculating advantage of artificial neural network is underused in the prediction of individual parameter, and the parameter of prediction is also all relatively easy to obtain,
Actual application value is smaller.
The content of the invention
The purpose of the present invention is to overcome above-mentioned the deficiencies in the prior art, there is provided one kind prediction nuclear magnetic resonance log laterally relaxes
Henan time T2The method of distribution, passes through artificial nerve network model, it is only necessary to conventional logging data, you can Accurate Prediction laterally relaxes
Henan time T2The method of distribution, so as to the expensive NMR Logging Technology of fictitious hosts, realize using lower-cost normal
Advise log data, the expensive NMR Logging Technology of alternative cost.
To achieve the above object, the present invention uses following technical proposals:During one kind prediction nuclear magnetic resonance log transverse relaxation
Between T2The method of distribution, comprises the following steps:
(1) screening determines related conventional logging data and mineralogical composition and fluid volume inverting data, as step
(5) nuclear magnetic resonance log T is predicted2The input data of distribution.
(2) classification parameter, the existing T of quantitative classification are set2Distribution curve, as model training in step (5) and
The input data of test.In order to accurately portray the different types of T of identification2Distribution characteristics, different classification parameters need to be set to T2Point
Cloth curve carries out quantitative classification.
(3) the existing nuclear magnetic resonance log T of multiple normal distribution curves fitting is utilized2Distribution, obtains the sign of normal distribution
Parameter, after assessing fitting accuracy using coefficient correlation, as model training in step 5 and the output data of test.People
The output par, c of artificial neural networks model is the T of nuclear magnetic resonance log2Distribution profile.Based on actual measurement T2It is distributed modal data.
(4) is carried out to step (1) conventional logging data, step (2) classification parameter, step (3) characterization parameter early stage
Processing.Before application input data carries out model prediction, data processing early stage is most important, it is ensured that is preferably inputted
The Relationship Prediction model of value and output valve.
(5) by all data sorting and groupings of step (4), the training and survey of artificial nerve network model are respectively used to
Examination, obtains forecast model.
(6) coefficient correlation and standard root mean square deviation NRMSE assessment prediction model exactness are used, predicts nuclear magnetic resonance log
T2Distribution.
Preferably, the conventional logging data in the step (1) are the well logging number related to distribution of pores and permeability
According to.The screening conventional logging data related to measure porosity and permeability physical property are collected, these conventional logging data will
As the input data of artificial nerve network model, prediction nuclear magnetic resonance log T2Distribution, with the expensive direct survey of fictitious hosts
Amount technology.
Preferably, the mineralogical composition in the step (1) and fluid volume inverting data include illite, chlorite, beam
Tie up water, quartz, potassium feldspar, calcite, dolomite, gypsum, can not expelling water, the inverting data of not drainable oil.
Preferably, the log data in the step (1) includes gamma ray well logging, density log, neutron well logging, electricity
The well logging of resistance rate, acoustic logging.
Preferably, the setting classification parameter in the step (2), the existing T of quantitative classification2Distribution curve, it is by nearest
Adjacent disaggregated model k-nearest neighbor classification model, using conventional logging data, predict T2Distribution
Required classification parameter, selection determine, into optimal k values, to meet prediction distribution curve and actual curve best fit degree.
Preferably, classification parameter is set in the step (2) to be included setting the different reservoir lithology of different depth, distribution
Peak number, peak correspond to porosity value size, abnormal short peak presence or absence and the flat degree of distribution curve.
Preferably, multiple normal distribution curves in the step (3), its fitting formula are as follows:
Wherein T '2=log (T2), ɡiIt is the probability density function of normal distribution:Corresponding μiIt is average, σiIt is standard deviation, αi
It is the coefficient of probability density function, therefore characterizes every layer of T2Distribution needs 6 parameters.If certain layer of T2Distribution only has a peak, then
α2=μ2=σ2=0.
Preferably, the method that checking is fitted accuracy in the step (3), is by coefficient correlation, abbreviation R2, to comment
Estimate and utilize normal distribution curve fitting actual measurement T2The accuracy of distribution profile, wherein, R2Formula is as follows:
Wherein, ffit(T’2) it is match value or predicted value, f (T '2) it is initial value,It is initial mean value, RSS is error
Quadratic sum, TSS are totals sum of squares;Through being successively fitted, every layer of T is obtained2The R of the matched curve of distribution2Value, R2Closer to 1, show
As a result it is more accurate.
Preferably, data pre-processing includes suppressing exception value and standard normalized in the step (4):First, need
The exceptional value in initial data is deleted, initial data is subjected to standard normalized, ensures that all data have identical point
Cloth scope so that for the scope of data all between -1 to 1, its standard normalizing equation is as follows after all processing:
Wherein x is initial data, and y is standardized data.After the completion of model training, all data are inversely converted back into each
Original scope.
Preferably, the data of step (5) sorting and grouping, 85% data are therefrom randomly selected as training number
According to residue 15% is used as test data;The artificial nerve network model is divided into input layer, output layer and hidden layer, input layer
As input data, output layer are output data, and hidden layer is used for the relation for finding input layer and output layer.In the present invention, people
It is two layers that artificial neural networks model, which sets the hidden layer number of plies,.Work as input layer, the neuron number of hidden layer and output layer close to etc.
During difference series, the neuron number of model, which is set, to be more adapted to.The formula for calculating weight and deviation factor number is as follows:
Nt=(N1+1)*N2+(N2+1)*N3+(N3+1)*N4
Wherein NtIt is the sum of weight and deviation factor, NiIt is each layer of neuron number.
The algorithm that the artificial nerve network model is used to restrain object function includes, Levenberg-Ma Kuite
(Levenberg-Marquardt, hereinafter referred to as LM) algorithm and convolution gradient (Conjugate Gradient, hereinafter referred to as CG)
Algorithm.
LM algorithms are applied to weight and the sum of deviation factor and the model less than 300, and CG algorithms are applied to total and more
In 500 model.When the sum of weight and deviation factor is between 300-500, two model computational efficiencies are close, and different is pre-
Survey example and have different comparative results, therefore in the range of 300-500, generally two algorithms are separately in model, lead to
The time for crossing comparison model training determines the computational efficiency height of algorithm, from the required model training time is less, computational efficiency
Higher algorithm.
In order to avoid over-fitting, add regularization process, and punishment parameter is set in model.Introduce square of punishment parameter
Difference and (Sum of Squared Errors, abbreviation SSE) formula are as follows:
Wherein n is total number of samples, and P is the neuron number of output layer, that is, calculates the formula of weight and deviation factor number
In N4;λ is punishment parameter, yiIt is original output,It is to predict output valve, σj 2It is output valve variance, it is people that SSE formula, which are,
The object function of artificial neural networks model.
Preferably, coefficient correlation and standard root mean square deviation NRMSE assessment prediction model exactness are used in the step (6)
Specific method be, by using NRMSE and R2Measured data fitting result in model prediction result and step (3) is carried out
Compare, assess gained forecast model accuracy, wherein, NRMSE is defined as follows:
Wherein RMSE is root mean square deviation, and NRMSE span is 0 to 1, shows that prediction result is more accurate closer to 0.
The invention has the advantages that
(1) realize and utilize artificial nerve network model prediction T2Distribution, the expensive NMR Logging Technology of alternative cost
The application of artificial nerve network model in the oil industry is in the starting stage, and is mainly used in predicting at present
The single parameter such as porosity, permeability, water saturation, it is relatively simple.The present invention can be direct prediction T2Distribution, than prediction
Single parameter is more complicated.T2The distribution of pores that can reflect in stratum is distributed, therefore predicts T2Distribution can be obtained than predicting single ginseng
The more useful formation informations of number
The present invention utilizes neural network prediction T2Distribution, by using artificial intelligence technology, instead of expensive
NMR Logging Technology direct measurement, improve the operating efficiency of petroleum industry, also save economy in well logging link and
Time cost.
(2) for more set distribution curve systems, classification parameter concept is firstly introduced, improves prediction accuracy
For the one-parameter prediction such as prediction porosity, permeability, prediction such as T2The distribution curve of distributional class is more
It is complicated.The present invention improves prediction T by introducing classification parameter2The accuracy of distribution.Classification parameter can take into full account more set ground
The features such as layer different lithology, distribution curve morphological feature, peak Distribution feature, sample data is quantified to classify, quantization parameter point
Cloth curve law, realize and predicted using artificial nerve network model precise quantification.
(3) it is fitted T2Distribution, greatly reduces model amount of calculation and ensures prediction accuracy
In the present invention, original T264 discrete point compositions are distributed with, the present invention utilizes Gauss Distribution Fitting T2Distribution, by original
This prediction to 64 discrete points is reduced to the prediction to 6 Gaussian Distribution Parameters so that the amount of calculation of model reduces 90%
More than, it not only can guarantee that prediction T2The accuracy of distribution, also substantially increase the efficiency of training and test model.
(4) multi objective R2With NRMSE assessment models accuracy
R2Degree of correlation between key reaction parameter, R2Show that the accuracy of model prediction is higher closer to 1.NRMSE master
The size of response prediction error is wanted, NRMSE shows that the accuracy of model prediction is higher closer to 0.This model applies R simultaneously2With
NRMSE carrys out assessment models accuracy so that assessment result is more accurately and reliably.
The present invention utilizes artificial nerve network model, based on accurately predicting T using conventional logging data2One kind side of distribution
Method, training and test have obtained conventional logging data and nuclear magnetic resonance log T2The relational model of distribution, avoid direct use
The high NMR Logging Technology of cost consumer, economy and time cost are saved, while prediction obtains high-precision T2Distribution profile,
Help understands subsurface deposit feature, such as distribution of pores, permeability.
Brief description of the drawings
Fig. 1 is flow chart of steps of the present invention;
Fig. 2 is U.S. Ba Ken (Bakken) oil reservoir well series log and nuclear magnetic resonance log gained T in embodiment2
Distribution map;
Fig. 3 is according to T in embodiment2It is distributed the classification parameter and its setting value summary view of definition;
Fig. 4 is Gauss normal distribution curve matching T in embodiment2Distribution is with surveying original T2Figure is compared in distribution;
Fig. 5 is that normal distribution curve is fitted T in embodiment2The levels of precision R of distribution2Distribution map;
Fig. 6 is artificial nerve network model structural representation;
Fig. 7 is to predict gained T using artificial nerve network model2Distribution is with surveying original T2Figure is compared in distribution;
Fig. 8 is all layer model prediction results and actual T2Distribution relatively gained R2And NRMSE statistical Butut.
Embodiment
The present invention is further described with reference to the accompanying drawings and examples.
U.S. Ba Ken (Bakken) oil reservoir is one of maximum shale oil gas field in the U.S..The present invention is with U.S.'s Ba Ken petroleums
Method displaying is carried out exemplified by certain well in system.The well target layer depth is XX675 feet to XX985 feet, is total to through observing nuclear-magnetism
Shake well logging T2Distribution characteristics, find the T more than 95% in target interval2Distribution curve only has a peak or two peaks, therefore
When being fitted in the later stage, it is only necessary to most two normal distribution curves fittings, you can accurate Characterization T2Distribution.
A kind of as shown in figure 1, prediction nuclear magnetic resonance log T2 T2The method of distribution, comprises the following steps:
(1):Screening determines related conventional logging data and mineralogical composition and fluid volume inverting data, as step
Suddenly (five) predict nuclear magnetic resonance log T2The input data of distribution.
Collect the screening conventional logging data related to measure porosity and permeability physical property, these conventional logging numbers
According to using as the input data of artificial nerve network model, prediction nuclear magnetic resonance log T2Distribution, with the straight of fictitious hosts costliness
Connect e measurement technology.
First, prediction accuracy can be improved as the input data of model for as often as possible selection conventional logging data,
The conventional loggings such as gamma ray well logging, density log, neutron well logging, resistivity logging, acoustic logging are determined by screening not
Input data with log for artificial nerve network model, as shown in Figure 2.In addition, be to improve model prediction accuracy, 10
Kind of mineralogical composition and fluid volume (illite, chlorite, irreducible water, quartz, potassium feldspar, calcite, dolomite, gypsum, can not
Expelling water, not drainable oil) inverting data also by as input data.
(2):Set classification parameter, the existing T of quantitative classification2Distribution curve, as model training in step (5) and
The input data of test.
In order to accurately portray the different types of T of identification2Distribution characteristics, different classification parameters need to be set to T2Distribution curve
Quantitative classification is carried out, by arest neighbors disaggregated model (k-nearest neighbor classification model), is utilized
Conventional logging data, predict T2Classification parameter needed for distribution, selection determine, into optimal k values, to meet prediction distribution curve and reality
Curve best fit degree.
It is described to set different classification parameters to include setting the different reservoir lithology of different depth, distribution of peaks number, peak pair
Answer porosity value size, abnormal short peak presence or absence and the flat degree of distribution curve.
First group of classification parameter in embodiment:Seven kinds of different lithologies according to existing for the depth segment, totally 7 class, is designated as respectively
1-7.Second group of classification parameter:According to T2The peak number of distribution is different to be set, and 0 corresponding 1 peak, 1 corresponding 2 peaks, shares 2 classes;The
Three groups of classification parameters:The pore size according to corresponding to distribution of peaks is different to be set, and setting peak corresponds to T2Reference value situation is 0, peak pair
Answer T2It is worth larger, then plus 1, peak corresponds to T2Value is smaller, then subtracts 1;4th group of classification parameter, according to whether being set with the presence of abnormal short peak
Put, be set to 1 if it abnormal short peak be present, be set to 0 if nothing, totally 2 class;5th group of classification parameter:According to T2Distribution curve is flat
Degree (peak span) is set, and is designated as 1 if all peak spans of distribution curve are all very big, is otherwise designated as 0, totally 2 class.
After having set classification parameter, by arest neighbors disaggregated model, after 12 kinds of conventional logging data and 10 predictions four are utilized
Group classification parameter, the scope selection 1 to 10 determine, into optimal k values, to meet prediction distribution curve and actual curve best fit
Degree.As a result show, in this example, as k=3, prediction result is most accurate.
(3):Existing nuclear magnetic resonance log T is fitted using multiple normal distribution curves2Distribution, obtains the master of normal distribution
Characterization parameter is wanted, after assessing fitting accuracy using coefficient correlation, as model training in step 5 and the output of test
Data.
The output par, c of artificial nerve network model is the T of nuclear magnetic resonance log2Distribution profile.Based on actual measurement T2Distribution profile number
According to the present invention characterizes the T of measured data using the fitting of multiple normal distribution curves2Distribution profile.T in this example2Distribution profile
Every layer has 64 data to form, all log datas and T2Distribution profile as shown in figure 3, be respectively from left to right in figure:In 1 row
Depth is depth measurement, and GR be gamma ray in 2 row, and DCAL is hole diameter, 3 arrange in VPVS be compressional wave transverse wave speed ratio, DTCO/DTSM
For acoustic logging, AT be resistivity logging in 4 row, and DPHZ is density log in 5 row, and NPOR is neutron well logging, 6 arrange in PEFZ be
Photoelectric effect, VCL are clay content, and RHOZ be rock density in 7 row, and QUARTZ be quartz content in 8 row, and CALCITE is just
Solve stone content, DOLOMITE is that BOUND_WATER is bound water content during dolomite content 9 arranges, UWATER for can not expelling water contain
Amount, UOIL are not drainable oil content, and T2 is T in 10 row2Distribution.It is fitted using multiple normal distribution curves as follows:
Wherein T '2=log (T2), ɡiIt is the probability density function of normal distribution:Corresponding μiIt is average, σiIt is standard deviation, αi
It is the coefficient of probability density function, therefore characterizes every layer of T2Distribution needs 6 parameters.If certain layer of T2Distribution only has a peak, then
α2=μ2=σ2=0.
The present invention utilizes coefficient correlation (hereinafter referred to as R2) come assess utilize normal distribution curve fitting actual measurement T2Distribution profile
Accuracy, wherein, R2Formula is as follows:
Wherein ffit(T’2) it is match value or predicted value, f (T '2) it is initial value,It is initial mean value, RSS is error
Quadratic sum, TSS are totals sum of squares.Through being successively fitted, every layer of T is obtained2The R of the matched curve of distribution2Value, R2Closer to 1, show
As a result it is more accurate.
Example of the present invention shares 416 layers, the R of all matched curves2Median 0.9834.Fig. 4 is part T2Distribution
Fitting result, wherein, dotted line is actual T2Distribution, solid line is matched curve.Fig. 5 is whole 416 layers of matched curve R2Statistics
Distribution map.
Step 4:Before being carried out to step (1) conventional logging data, step (2) classification parameter, step (3) characterization parameter
Phase is handled.
Before application input data carries out model prediction, data processing early stage is most important, it is ensured that obtains preferably
The Relationship Prediction model of input value and output valve.Data pre-processing mainly includes:First, some are different existing for deletion initial data
Constant value.For example, some layer of position in this example, the numerical value of gamma ray is more than 1000API, it is clear that it is unreasonable, because, generally
The API values of gamma ray are less than 200.Secondly, in view of every group of data area difference, has same distribution model for all data of guarantee
Enclose, initial data is subjected to standard normalized.The data area of every layer of group is also different in example, in order that must own
Data have identical scope, it is necessary to which initial data is standardized, and the data area after guarantee processing is all -1 to 1
Between, normalized equation is as follows:
Wherein x is initial data, and y is the data after standardization.After the completion of training, all data are inversely converted into respective
Original scope.
Step 5:By all data sorting and groupings of step (4), be respectively used to artificial nerve network model training and
Test, obtains forecast model.
Data need to be divided into two parts, training data and test data, the data of step (5) sorting and grouping, from
In randomly select 85% data and be used as test data as training data, residue 15%.
Artificial nerve network model is divided into input layer, output layer and hidden layer, and input layer is input data, and output layer is
Output data, hidden layer are used for the relation for finding input layer and output layer, as shown in Figure 6.This artificial nerve network model is set
The hidden layer number of plies is two layers.Work as input layer, when the neuron number of hidden layer (two layers) and output layer is close to arithmetic progression, model
Neuron number set more be adapted to.For example, we have 27 groups of input values, 6 groups of output valves, then, input layer and output layer
There are 27 and 6 neurons respectively, hidden layer can be respectively set to 20 and 13 neurons, and so, this model just has 917
Individual weight and deviation factor need to calculate.The formula for calculating weight and deviation factor number is as follows:
Nt=(N1+1)*N2+(N2+1)*N3+(N3+1)*N4
Wherein NtIt is the sum of weight and deviation factor, NiIt is each layer of neuron number.
Artificial nerve network model has many algorithms, for restraining object function, mainly includes:Levenberg-Ma Kuite
(Levenberg-Marquardt, hereinafter referred to as LM) algorithm and convolution gradient (Conjugate Gradient, hereinafter referred to as CG)
Algorithm.LM algorithms are applied to total and less, the model less than 300 of weight and deviation factor, CG algorithms be applied to sum with
It is more, the model more than 500.When the sum of weight and deviation factor is between 300-500, due to two model computational efficiency phases
Closely, different prediction examples has different comparative results, therefore, it is necessary to which two algorithms are applied respectively in the range of 300-500
Into model, by comparison model train time determine algorithm computational efficiency height, from the required model training time compared with
Less, the higher algorithm of computational efficiency.Because this model only has 917 weights and deviation factor, CG algorithms are more suitable for.
In addition, in order to avoid over-fitting, need to add regularization process in model, and set punishment parameter.Introduce punishment
The SSE formula of parameter are as follows:
Wherein n is total number of samples, and P is the neuron number of output layer, that is, calculates the formula of weight and deviation factor number
In N4, λ is punishment parameter, yiIt is original output,It is to predict output valve, σj 2It is output valve variance, it is people that SSE formula, which are,
The object function of artificial neural networks model.λ spans are [0,1], and by comparing, λ=0.5 is chosen as the punishment ginseng of model
Number.
Step 6:With coefficient correlation and standard root mean square deviation NRMSE assessment prediction model exactness, prediction nuclear magnetic resonance is surveyed
Well T2Distribution.
Obtaining neural network prediction T2After the model of distribution, the present invention uses NRMSE and R simultaneously2To model prediction
As a result compared with measured data fitting result in step 3, gained forecast model accuracy is assessed, wherein, NRMSE definition
It is as follows:
Wherein, RMSE is root mean square deviation, and NRMSE span is 0 to 1, shows that prediction result is more accurate closer to 0.
Fig. 7 is illustrated in the layer position of part and is predicted T2Distribution and actual T2The comparative result of distribution, illustrate to be established pre-
It is more accurate to survey mould.In addition, such as Fig. 8, model prediction result and the actual T of all layers of position2The determination coefficients R of distribution relatively gained2
And standard root mean square deviation NRMSE statistical distribution, wherein, R2Median is that 0.7553, NRMSE medians are 0.1575, as a result
Indicate the accuracy of prediction result.
Finally it should be noted that the above embodiments are merely illustrative of the technical solutions of the present invention, rather than the present invention is protected
The limitation of scope is protected, although being explained with reference to preferred embodiment to the present invention, one of ordinary skill in the art should
Work as understanding, technical scheme can be modified or equivalent substitution, without departing from the reality of technical solution of the present invention
Matter and scope.
Claims (10)
1. one kind prediction nuclear magnetic resonance log T2 T2The method of distribution, it is characterized in that, comprise the following steps:
(1) screening determines related conventional logging data and mineralogical composition and fluid volume inverting data, pre- as step 5
Survey nuclear magnetic resonance log T2The input data of distribution;
(2) classification parameter, the existing T of quantitative classification are set2Distribution curve, as model training in step 5 and test defeated
Enter data;
(3) the existing nuclear magnetic resonance log T of multiple normal distribution curves fitting is utilized2Distribution, obtains the characterization parameter of normal distribution,
After assessing fitting accuracy using coefficient correlation, as model training in step 5 and the output data of test;
(4) processing early stage is carried out to step 1 conventional logging data, step 2 classification parameter, step 3 characterization parameter;
(5) by all data sorting and groupings of step 4, the training and test of artificial nerve network model is respectively used to, is obtained
Forecast model;
(6) coefficient correlation and standard root mean square deviation NRMSE assessment prediction model exactness, prediction nuclear magnetic resonance log T are used2Point
Cloth.
A kind of 2. prediction nuclear magnetic resonance log T2 T as claimed in claim 12The method of distribution, it is characterized in that,
Conventional logging data in the step 1 are the log data related to distribution of pores and permeability, mineralogical composition and fluid body
Product inverting data include illite, chlorite, irreducible water, quartz, potassium feldspar, calcite, dolomite, gypsum, can not expelling water, no
The inverting data of drainable oil.
A kind of 3. prediction nuclear magnetic resonance log T2 T as described in any one of claim 1 or 22The method of distribution,
It is characterized in that log data in the step 1 include gamma ray well logging, density log, neutron well logging, resistivity logging,
Acoustic logging.
A kind of 4. prediction nuclear magnetic resonance log T2 T as claimed in claim 12The method of distribution, it is characterized in that,
Setting classification parameter in the step 2, the existing T of quantitative classification2Distribution curve, it is by arest neighbors disaggregated model k-
Nearest neighbor classification model, using conventional logging data, predict T2Classification needed for distribution is joined
Number, selection determine, into optimal k values, to meet prediction distribution curve and actual curve best fit degree.
A kind of 5. prediction nuclear magnetic resonance log T2 T as described in any one of claim 1 or 42The method of distribution,
It is characterized in that the setting classification parameter includes setting the different reservoir lithology of different depth, distribution of peaks number, peak correspond to hole
It is worth size, abnormal short peak presence or absence and the flat degree of distribution curve.
A kind of 6. prediction nuclear magnetic resonance log T2 T as claimed in claim 12The method of distribution, it is characterized in that,
Multiple normal distribution curves in the step 3, its fitting formula are as follows:
<mrow>
<mi>f</mi>
<mrow>
<mo>(</mo>
<msubsup>
<mi>T</mi>
<mn>2</mn>
<mo>&prime;</mo>
</msubsup>
<mo>)</mo>
</mrow>
<mo>=</mo>
<munderover>
<mo>&Sigma;</mo>
<mrow>
<mi>i</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
<mn>2</mn>
</munderover>
<mrow>
<mo>(</mo>
<msub>
<mi>&alpha;</mi>
<mi>i</mi>
</msub>
<mo>)</mo>
</mrow>
<msub>
<mi>g</mi>
<mi>i</mi>
</msub>
<mrow>
<mo>(</mo>
<msub>
<mi>&mu;</mi>
<mi>i</mi>
</msub>
<mo>,</mo>
<msub>
<mi>&sigma;</mi>
<mi>i</mi>
</msub>
<mo>,</mo>
<msubsup>
<mi>T</mi>
<mn>2</mn>
<mo>&prime;</mo>
</msubsup>
<mo>)</mo>
</mrow>
</mrow>
Wherein T '2=log (T2), ɡiIt is the probability density function of normal distribution:Corresponding μiIt is average, σiIt is standard deviation, αiIt is general
The coefficient of rate density function, therefore characterize every layer of T2Distribution needs 6 parameters;If certain layer of T2Distribution only has a peak, then α2=
μ2=σ2=0.
A kind of 7. prediction nuclear magnetic resonance log T2 T as claimed in claim 12The method of distribution, it is characterized in that,
The method of checking fitting accuracy, is by coefficient correlation, abbreviation R in the step 32, normal distribution curve is utilized to assess
Fitting actual measurement T2The accuracy of distribution profile, wherein, R2Formula is as follows:
<mrow>
<msup>
<mi>R</mi>
<mn>2</mn>
</msup>
<mo>=</mo>
<mn>1</mn>
<mo>-</mo>
<mfrac>
<mrow>
<mi>R</mi>
<mi>S</mi>
<mi>S</mi>
</mrow>
<mrow>
<mi>T</mi>
<mi>S</mi>
<mi>S</mi>
</mrow>
</mfrac>
</mrow>
<mrow>
<mi>R</mi>
<mi>S</mi>
<mi>S</mi>
<mo>=</mo>
<munderover>
<mo>&Sigma;</mo>
<mrow>
<mi>i</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
<mi>n</mi>
</munderover>
<msup>
<mrow>
<mo>&lsqb;</mo>
<msub>
<mi>f</mi>
<mrow>
<mi>f</mi>
<mi>i</mi>
<mi>t</mi>
</mrow>
</msub>
<mrow>
<mo>(</mo>
<msubsup>
<mi>T</mi>
<mn>2</mn>
<mo>&prime;</mo>
</msubsup>
<mo>)</mo>
</mrow>
<mo>-</mo>
<mi>f</mi>
<mrow>
<mo>(</mo>
<msubsup>
<mi>T</mi>
<mn>2</mn>
<mo>&prime;</mo>
</msubsup>
<mo>)</mo>
</mrow>
<mo>&rsqb;</mo>
</mrow>
<mn>2</mn>
</msup>
</mrow>
<mrow>
<mi>T</mi>
<mi>S</mi>
<mi>S</mi>
<mo>=</mo>
<munderover>
<mo>&Sigma;</mo>
<mrow>
<mi>i</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
<mi>n</mi>
</munderover>
<msup>
<mrow>
<mo>&lsqb;</mo>
<mi>f</mi>
<mrow>
<mo>(</mo>
<msubsup>
<mi>T</mi>
<mn>2</mn>
<mo>&prime;</mo>
</msubsup>
<mo>)</mo>
</mrow>
<mo>-</mo>
<mover>
<mrow>
<mi>f</mi>
<mrow>
<mo>(</mo>
<msubsup>
<mi>T</mi>
<mn>2</mn>
<mo>&prime;</mo>
</msubsup>
<mo>)</mo>
</mrow>
</mrow>
<mo>&OverBar;</mo>
</mover>
<mo>&rsqb;</mo>
</mrow>
<mn>2</mn>
</msup>
</mrow>
Wherein ffit(T’2) it is match value or predicted value, f (T '2) it is initial value,It is initial mean value, RSS is square-error
With TSS is total sum of squares;Through being successively fitted, every layer of T is obtained2The R of the matched curve of distribution2Value, R2Closer to 1, show result
It is more accurate.
A kind of 8. prediction nuclear magnetic resonance log T2 T as claimed in claim 12The method of distribution, it is characterized in that,
Data pre-processing, which includes suppressing exception value and standard normalized, in the step 4 is, firstly, it is necessary to delete initial data
In exceptional value, initial data is subjected to standard normalized, ensures that all data have same distribution scope so that all
For the scope of data all between -1 to 1, its standard normalizing equation is as follows after processing:
<mrow>
<mi>y</mi>
<mo>=</mo>
<mn>2</mn>
<mfrac>
<mrow>
<mi>x</mi>
<mo>-</mo>
<msub>
<mi>x</mi>
<mi>min</mi>
</msub>
</mrow>
<mrow>
<msub>
<mi>x</mi>
<mi>max</mi>
</msub>
<mo>-</mo>
<msub>
<mi>x</mi>
<mi>min</mi>
</msub>
</mrow>
</mfrac>
<mo>-</mo>
<mn>1</mn>
</mrow>
Wherein x is initial data, and y is standardized data;After the completion of model training, all data are inversely converted back into each original
Scope.
A kind of 9. prediction nuclear magnetic resonance log T2 T as claimed in claim 12The method of distribution, it is characterized in that,
The data of step 5 sorting and grouping, 85% data are therefrom randomly selected as training data, residue 15% is as surveying
Try data;The artificial nerve network model is divided into input layer, output layer and hidden layer, and input layer is input data, output
Layer is output data, and hidden layer is used for the relation for finding input layer and output layer, and in of the invention, artificial nerve network model is set
The hidden layer number of plies is two layers, works as input layer, when the neuron number of hidden layer and output layer is close to arithmetic progression, the nerve of model
First number, which is set, to be more adapted to, and the formula for calculating weight and deviation factor number is as follows:
Nt=(N1+1)*N2+(N2+1)*N3+(N3+1)*N4
Wherein NtIt is the sum of weight and deviation factor, NiIt is each layer of neuron number;
The algorithm that the artificial nerve network model is used to restrain object function includes, Levenberg-Ma Kuite Levenberg-
Marquardt algorithms, i.e. LM algorithms and convolution gradient Conjugate Gradient algorithms, i.e. CG algorithms;LM algorithms are applied to
The sum of weight and deviation factor and the model less than 300, CG algorithms are applied to sum and the model more than 500, when weight and
Need two algorithms being separately in model when the sum of deviation factor is between 300-500, trained by comparison model
Time determines the computational efficiency height of algorithm, from the algorithm that the required model training time is less, computational efficiency is higher;
In order to avoid over-fitting, regularization process is added in model, and punishment parameter is set, introduce punishment parameter the difference of two squares and
(Sum of Squared Errors, abbreviation SSE) formula is as follows:
<mrow>
<mi>S</mi>
<mi>S</mi>
<mi>E</mi>
<mo>=</mo>
<munderover>
<mo>&Sigma;</mo>
<mrow>
<mi>i</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
<mi>n</mi>
</munderover>
<msup>
<mrow>
<mo>(</mo>
<msub>
<mi>y</mi>
<mi>i</mi>
</msub>
<mo>-</mo>
<msub>
<mover>
<mi>y</mi>
<mo>^</mo>
</mover>
<mi>i</mi>
</msub>
<mo>)</mo>
</mrow>
<mn>2</mn>
</msup>
<mo>+</mo>
<mi>&lambda;</mi>
<munderover>
<mo>&Sigma;</mo>
<mrow>
<mi>j</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
<mi>P</mi>
</munderover>
<msubsup>
<mi>&sigma;</mi>
<mi>j</mi>
<mn>2</mn>
</msubsup>
</mrow>
Wherein n is total number of samples, and P is the neuron number of output layer, that is, in the formula for calculating weight and deviation factor number
N4, λ is punishment parameter, and yi is original output,It is to predict output valve, σj 2It is output valve variance, SSE formula are artificial god
Object function through network model.
A kind of 10. prediction nuclear magnetic resonance log T2 T as claimed in claim 12The method of distribution, it is characterized in that,
With the specific method of coefficient correlation and standard root mean square deviation NRMSE assessment prediction model exactness it is to pass through in the step 6
Using NRMSE and R2To model prediction result compared with measured data fitting result in step 3, gained prediction mould is assessed
Type accuracy, wherein, NRMSE is defined as follows:
<mrow>
<mi>R</mi>
<mi>M</mi>
<mi>S</mi>
<mi>E</mi>
<mo>=</mo>
<msqrt>
<mfrac>
<mrow>
<msubsup>
<mo>&Sigma;</mo>
<mrow>
<mi>i</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
<mi>n</mi>
</msubsup>
<msup>
<mrow>
<mo>(</mo>
<msub>
<mover>
<mi>y</mi>
<mo>^</mo>
</mover>
<mi>i</mi>
</msub>
<mo>-</mo>
<msub>
<mi>y</mi>
<mi>i</mi>
</msub>
<mo>)</mo>
</mrow>
<mn>2</mn>
</msup>
</mrow>
<mi>n</mi>
</mfrac>
</msqrt>
</mrow>
<mrow>
<mi>N</mi>
<mi>R</mi>
<mi>M</mi>
<mi>S</mi>
<mi>E</mi>
<mo>=</mo>
<mfrac>
<mrow>
<mi>R</mi>
<mi>M</mi>
<mi>S</mi>
<mi>E</mi>
</mrow>
<mrow>
<msub>
<mi>y</mi>
<mrow>
<mi>m</mi>
<mi>a</mi>
<mi>x</mi>
</mrow>
</msub>
<mo>-</mo>
<msub>
<mi>y</mi>
<mi>min</mi>
</msub>
</mrow>
</mfrac>
</mrow>
Wherein RMSE is root mean square deviation, and NRMSE span is 0 to 1, shows that prediction result is more accurate closer to 0.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710969507.1A CN107728231B (en) | 2017-10-18 | 2017-10-18 | A kind of prediction nuclear magnetic resonance log lateral relaxation time T2The method of distribution |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710969507.1A CN107728231B (en) | 2017-10-18 | 2017-10-18 | A kind of prediction nuclear magnetic resonance log lateral relaxation time T2The method of distribution |
Publications (2)
Publication Number | Publication Date |
---|---|
CN107728231A true CN107728231A (en) | 2018-02-23 |
CN107728231B CN107728231B (en) | 2019-04-12 |
Family
ID=61211724
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710969507.1A Expired - Fee Related CN107728231B (en) | 2017-10-18 | 2017-10-18 | A kind of prediction nuclear magnetic resonance log lateral relaxation time T2The method of distribution |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN107728231B (en) |
Cited By (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109343134A (en) * | 2018-11-27 | 2019-02-15 | 中煤科工集团西安研究院有限公司 | A kind of Transient Electromagnetic Method in Mine data analysis interpretation method and system |
CN111950193A (en) * | 2020-07-15 | 2020-11-17 | 中海油田服务股份有限公司 | Modeling method and device of pore network model based on reservoir |
CN113743597A (en) * | 2021-11-03 | 2021-12-03 | 中国科学院精密测量科学与技术创新研究院 | NMR (nuclear magnetic resonance) relaxation time inversion method based on untrained deep neural network |
CN113743682A (en) * | 2021-11-03 | 2021-12-03 | 中国科学院精密测量科学与技术创新研究院 | NMR (nuclear magnetic resonance) relaxation time inversion method based on supervised deep neural network |
CN113743596A (en) * | 2021-11-03 | 2021-12-03 | 中国科学院精密测量科学与技术创新研究院 | NMR (nuclear magnetic resonance) relaxation time inversion method based on unsupervised neural network |
CN114046145A (en) * | 2021-11-26 | 2022-02-15 | 中国石油大学(北京) | Reservoir fluid identification and saturation determination method and device |
CN114207550A (en) * | 2019-03-08 | 2022-03-18 | 斯伦贝谢技术有限公司 | System and method for supervised learning of permeability of a formation |
CN114594523A (en) * | 2022-01-04 | 2022-06-07 | 中国石油大学(北京) | Physical constraint and data-driven nuclear magnetic resonance T construction2Spectrum method and device |
US11680998B2 (en) | 2021-11-03 | 2023-06-20 | Innovation Academy For Precision Measurement Science And Technology, Cas | NMR relaxation time inversion method based on unsupervised neural network |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20050272158A1 (en) * | 2004-01-04 | 2005-12-08 | Galford James E | Method and apparatus for detecting hydrocarbons with NMR logs in wells drilled with oil-based muds |
US20130103627A1 (en) * | 2011-10-21 | 2013-04-25 | Eni S.P.A. | Method for predicting the properties of crude oils by the application of neural networks |
-
2017
- 2017-10-18 CN CN201710969507.1A patent/CN107728231B/en not_active Expired - Fee Related
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20050272158A1 (en) * | 2004-01-04 | 2005-12-08 | Galford James E | Method and apparatus for detecting hydrocarbons with NMR logs in wells drilled with oil-based muds |
US20130103627A1 (en) * | 2011-10-21 | 2013-04-25 | Eni S.P.A. | Method for predicting the properties of crude oils by the application of neural networks |
Non-Patent Citations (3)
Title |
---|
M R REZAEE ET AL.: "Relationships between permeability, porosity and pore throat size in carbonate rocks using regression analysis and neural networks", 《JOURNAL OF GEOPHYSICS AND ENGINEERING》 * |
周宇,等: "应用人工智能优化降低核磁共振孔隙度测量误差", 《西安石油大学学报(自然科学版)》 * |
朱林奇,等: "基于改进BPNN与T2全谱的致密砂岩储层渗透率预测", 《石油物探》 * |
Cited By (12)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109343134A (en) * | 2018-11-27 | 2019-02-15 | 中煤科工集团西安研究院有限公司 | A kind of Transient Electromagnetic Method in Mine data analysis interpretation method and system |
CN114207550A (en) * | 2019-03-08 | 2022-03-18 | 斯伦贝谢技术有限公司 | System and method for supervised learning of permeability of a formation |
CN111950193A (en) * | 2020-07-15 | 2020-11-17 | 中海油田服务股份有限公司 | Modeling method and device of pore network model based on reservoir |
CN111950193B (en) * | 2020-07-15 | 2023-03-14 | 中海油田服务股份有限公司 | Modeling method and device of pore network model based on reservoir |
CN113743597A (en) * | 2021-11-03 | 2021-12-03 | 中国科学院精密测量科学与技术创新研究院 | NMR (nuclear magnetic resonance) relaxation time inversion method based on untrained deep neural network |
CN113743682A (en) * | 2021-11-03 | 2021-12-03 | 中国科学院精密测量科学与技术创新研究院 | NMR (nuclear magnetic resonance) relaxation time inversion method based on supervised deep neural network |
CN113743596A (en) * | 2021-11-03 | 2021-12-03 | 中国科学院精密测量科学与技术创新研究院 | NMR (nuclear magnetic resonance) relaxation time inversion method based on unsupervised neural network |
CN113743596B (en) * | 2021-11-03 | 2022-02-18 | 中国科学院精密测量科学与技术创新研究院 | NMR (nuclear magnetic resonance) relaxation time inversion method based on unsupervised neural network |
WO2023077442A1 (en) * | 2021-11-03 | 2023-05-11 | 中国科学院精密测量科学与技术创新研究院 | Unsupervised neural network-based nmr relaxation time inversion method |
US11680998B2 (en) | 2021-11-03 | 2023-06-20 | Innovation Academy For Precision Measurement Science And Technology, Cas | NMR relaxation time inversion method based on unsupervised neural network |
CN114046145A (en) * | 2021-11-26 | 2022-02-15 | 中国石油大学(北京) | Reservoir fluid identification and saturation determination method and device |
CN114594523A (en) * | 2022-01-04 | 2022-06-07 | 中国石油大学(北京) | Physical constraint and data-driven nuclear magnetic resonance T construction2Spectrum method and device |
Also Published As
Publication number | Publication date |
---|---|
CN107728231B (en) | 2019-04-12 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN107728231B (en) | A kind of prediction nuclear magnetic resonance log lateral relaxation time T2The method of distribution | |
CN108303752B (en) | Conventional well logging quantitative identification method for effective glutenite reservoir | |
Wang et al. | Marcellus shale lithofacies prediction by multiclass neural network classification in the Appalachian Basin | |
Guerriero et al. | Improved statistical multi-scale analysis of fractured reservoir analogues | |
CN103336305B (en) | A kind of method dividing Sandstone Gas Reservoir high water cut based on gray theory | |
CN104749624A (en) | Method for synchronously realizing seismic lithofacies identification and quantitative assessment of uncertainty of seismic lithofacies identification | |
CN108897066A (en) | Carbonate rock crack density quantitative prediction method and device | |
CN109507733A (en) | A kind of method and device for predicting organic matter abundance in hydrocarbon source rock | |
CN105158796A (en) | Method and device for determining TOC content | |
CN111506861B (en) | Method for calculating crack strength of favorable region of target layer | |
CN111027882A (en) | Method for evaluating brittleness index by utilizing conventional logging data based on high-order neural network | |
CN103376468A (en) | Reservoir parameter quantitative characterization method based on neural network function approximation algorithm | |
CN111679318B (en) | Method and system for identifying fracture system in reservoir | |
Harding et al. | Trend modeling and modeling with a trend | |
CN111335871B (en) | Layering hole checking and supplementing technical method based on layering productivity evaluation | |
CN109655394B (en) | Nuclear magnetic resonance T2 spectrum permeability calculation method under constraint of throat structure parameters | |
CN111523795A (en) | Method for evaluating damage of shallow natural gas in region containing oil-gas basin to tunnel | |
Sarkheil et al. | The fracture network modeling in naturally fractured reservoirs using artificial neural network based on image loges and core measurements | |
La Pointe et al. | Estimation of effective block conductivities based on discrete network analyses using data from the Äspö site | |
CN107015289B (en) | Trap evaluation stock number determines method and device | |
CN116930023A (en) | Fine interpretation method and device for dense sandstone phase-control classified porosity logging | |
CN106767773A (en) | A kind of indoor earth magnetism reference map construction method and its device | |
CN106326620A (en) | Optimized selection method for diagenetic coefficient model of exploration target distribution range | |
Etminan et al. | An improved model for geostatistical simulation of fracture parameters and their effect on static and dynamic models | |
Röshoff et al. | Strategy for a Rock Mechanics Site Descriptive Model. Development and testing of the empirical approach |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant | ||
CB03 | Change of inventor or designer information | ||
CB03 | Change of inventor or designer information |
Inventor after: He Jiabo Inventor after: Yuan Bin Inventor after: Wang Wendong Inventor before: Yuan Bin Inventor before: He Jiabo Inventor before: Wang Wendong |
|
CF01 | Termination of patent right due to non-payment of annual fee | ||
CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20190412 Termination date: 20211018 |