CN104732303A - Oil field output prediction method based on dynamic radial basis function neural network - Google Patents

Abstract

The invention provides an oil field output prediction method based on a dynamic radial basis function neural network. The method comprises the steps that 1, factors which affect the output are determined according to oil field situations, and historical data are obtained and divided into a training data set and a test data set; 2, unitization processing is conducted on the data sets through a deviation standardization method; 3, an RBF neural network structure is adjusted in a dynamic mode through a sensitivity method, and a temporary RBF neural network prediction model is established; 4, a model error is corrected through a state transition probability matrix, and a stable RBF neural network oil output prediction model is obtained; 5, verification is conducted on the model through the test data sets obtained in the first step to judge whether the model meets expectations or not; 6 oil field output prediction is conducted through the output prediction model which meets the expectations and obtained in the fifth step. According to the oil field output prediction method based on the dynamic radial basis function neural network, the problem that the hidden layer neurons are too many or too small is avoided. and the obtained model has an adaptive adjustment function; second correction is conducted on a prediction error, and the prediction result is more accurate and reasonable.

Description

A kind of oilfield production forecast method based on dynamic radial basis function neural network

Technical field

The present invention relates to a kind of oilfield production forecast method based on dynamic radial basis function neural network, particularly relate to a kind of structure being carried out dynamic optimization radial basis function neural network by susceptibility method, bonding state transition probability method correction residual error, realizes oilfield production forecast method.

Background technology

Oil is as the lifeblood of national economy, and the height of its output directly has influence on the economic development of country.For field produces, a good economic benefit be ensured, just must have the stable oil yield that high.Guarantee that oil field stable high yield is the central task that oil-field development is produced.Therefore, to the Accurate Prediction of oil field oil yield be one of the important research task of oil-field development worker always.

The factor affecting oil field oil yield is divided into geologic agent and the large class of human factor two substantially.Geologic agent is unmodifiable change in other words is to some degree small.And the variation range of human factor wants much wide, adopt intensity from mining type, well pattern, well spacing, note, beat adjust well, closing down of old well scrap or metaideophone turns and adopts, to every artificial measures (comprise pressure break, acidifying, perforations adding tune layer, change electric pump, fluid operated pump, overhaul etc.), the change of each human factor all can have influence on the change of oil field oil yield.Therefore, the method for prediction oil field oil yield is also all based on geologic agent, human factor or the combination of the two, is mainly divided into two class methods: a class is from the view point of systematology, on the whole the oil yield in research prediction oil field; Another kind of is the increasing yield and injection effect studying single measure.First kind method mainly comprises statistical formula method (empirical method), characteristic curve method of water drive, material balance equation method, and numerical simulation for oil-gas reservoir method.But there is certain defect in statistical formula method (empirical method), characteristic curve method of water drive and material balance equation method: one is directly to consider the impact of the nonuniformity of reservoir on oil field oil yield; Two is to consider the impact of the change of every human factor on field output.In theory, numerical simulation for oil-gas reservoir method directly can consider that the change of the geologic agent of reservoir and every human factor is on the impact of field output comprehensively.But it is excessive to the dependence of geologic information, often cause because there is error to the understanding of reservoir geology situation, predicting the outcome of numerical simulation for oil-gas reservoir is had no value for use.Although, by can revise the understanding to reservoir geology to the meticulous matching of oil-field development history, but there is multi-solution in its fitting result. the Accurate Curve-fitting of oil-field development history is often required that researchist has sturdy geologic knowledge, oil reservoir knowledge, oil production technology knowledge, mathematical computations knowledge and computer literacy simultaneously, and workload is large.Isolatedly Equations of The Second Kind method studies the impact of every human factor on oil field oil yield, and this runs counter to the fact that oil-field development is an Iarge-scale system.The exploitation in oil field is a complicated nonlinear dynamic system. the prediction of oil field oil yield is a multifactor nonlinear prediction problem.Therefore, in order to ensure science and the accuracy of oilfield production forecast, a kind of new oilfield production forecast method of exigence, thus make recovery prediction result more accurate, objective, reasonable.

Summary of the invention

The present invention is from the angle of artificial intelligence, susceptibility method is utilized to optimize and revise radial basis function (RBF) neural network structure, field output influence factor data sample is utilized to train RBF neural, utilization state transition probability method is revised, and realizes recovery prediction result more accurate, objective, reasonable.

For achieving the above object, a kind of oilfield production forecast method based on dynamic radial basis function neural network being provided, mainly comprising the following steps:

A. data are obtained

According to oil field actual conditions, determine to affect field output factor index, obtain history data set and be divided into training dataset and detect data set;

B. normalized

Be normalized history data set, method for normalizing can adopt deviation standardized method, and make the data transformations of different dimension be unified processing format, transfer function is as follows:

$x^{*}=\frac{x-x_{\min }}{x_{\max }-x_{\min }}$ (formula 1)

Wherein x _maxfor the maximal value of sample data, x _minfor the minimum value of sample data;

C. the foundation of forecast model and training

In RBF neural, if K is hidden layer neuron number, x (x ₁..., x _m) be input vector, α _ka kth hidden layer neuron and the neuronic connection weights of output layer, φ _kbe the output of a kth hidden layer neuron, therefore the output of RBF neural can be described as:

$y=\underset{k=1}{\overset{K}{\mathrm{\Σ}}}{\mathrm{\α}}_k{\mathrm{\φ}}_k\left(x\right)$ (formula 2)

(1) a given hidden layer neuron is that the RBF neural of random natural number is trained, the number of times of setting training.

(2) each hidden layer neuron output valve is calculated.The output that can be obtained a kth hidden layer neuron by formula (1) is:

${\mathrm{\φ}}_k\left(x\right)=e^{\frac{-\left|\right|x-{\mathrm{\μ}}_k\left|\right|}{{\mathrm{\σ}}_k^2}}$ (formula 3)

(3) sensitivity analysis is carried out to the output of each neuron, calculate it to the contribution margin exported.The output weighted value of hidden layer neuron, as the input quantity of susceptibility method, utilizes following formula to calculate hidden layer neuron and exports the contribution done output neural network:

$S_h=\frac{{\mathrm{var}}_h\left[E\left(y\right|Z_h={\mathrm{\α}}_h{\mathrm{\φ}}_h\left(x\right))\right]}{\mathrm{var}\left(y\right)}$ (formula 4)

Wherein, Z=[Z ₁, Z ₂..., Z _k] be the input vector of susceptibility method, y is neural network output quantity, and the relation of y and Z can be expressed as y=f (Z ₁, Z ₂..., Z _k), var _h[E (y|Z _h=α _hφ _h(x))] be Z _hequal α _hφ _hx, on the impact of y variance time (), var (y) is the variance of y, S _hα _hφ _hthe one order exporting y is represented.To S _hbe normalized:

${\mathrm{ST}}_h=\frac{S_h}{\underset{i=1}{\overset{K}{\mathrm{\Σ}}}{S}_i}$ (formula 5)

(4) choose ε value and export contribution margin adjustment neural network structure according to hidden layer neuron.The value of ε is generally less than target error value, maximum and be greater than ε for contribution margin ₁hidden layer neuron divide, ε is less than for contribution margin ₂hidden layer neuron delete, ε here ₁> ε ₂, final realization adjusts neural network structure.Definition error objective function is (N is number of training):

$E=\frac{1}{2N}\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{e}_j^2$ (formula 6)

(5) according to target error function, utilize gradient descent algorithm to adjust output weights, the central value sum functions width of the hidden layer neuron of neural network:

$a_i(t+1)=a_i\left(t\right)-{\mathrm{\η}}_1{\mathrm{ST}}_i\frac{\∂E}{{\∂a}_i\left(t\right)}$ (formula 7)

${\mathrm{\μ}}_i(t+1)={\mathrm{\μ}}_i\left(t\right)-{\mathrm{\η}}_2{\mathrm{ST}}_i\frac{\∂E}{{\∂\mathrm{\μ}}_i\left(t\right)}$ (formula 8)

${\mathrm{\σ}}_i(t+1)={\mathrm{\σ}}_i\left(t\right)-{\mathrm{\η}}_3{\mathrm{ST}}_i\frac{\∂E}{{\∂\mathrm{\σ}}_i\left(t\right)}$ (formula 9)

Wherein, η ₁, η ₂, η ₃for parameter learning step-length.

(6) stop calculating when reaching anticipation error or calculation procedure.

D. residual GM

On the basis of the neural network model set up, the predicted value of training sample and actual output value are contrasted, calculates error sequence.Using error sequence as a Markovian process, carry out state demarcation, replace probability with frequency, the transition probability matrix of error of calculation state.

E. modelling verification

Use detection data the set pair analysis model to test, if the error of the predicted value exported and actual correlative value reaches the expectation of expection, neural network model is trained successfully, and model can be utilized to predict oil production rate; Otherwise model training is immature, need to re-start training.

F. oil production rate prediction

Obtain real fundamentals of forecasting data, input in the RBF neural after the optimization trained, the output of RBF neural is the predicted value of field output.

The invention has the beneficial effects as follows, more objective than general neural network prediction model basis for estimation, by the adjustment to neural network structure, the final neural network structure obtained is compact, there is good adaptive ability, make evaluation result science, accurate, fair and rational more.

Accompanying drawing explanation

Fig. 1 is the oilfield production forecast method flow diagram based on dynamic radial basis function neural network.

Embodiment

Below in conjunction with accompanying drawing, the present invention is described in further detail.

The first step: obtain data

According to oil field actual conditions, determine to affect field output factor index, obtain history data set and be divided into training dataset and detect data set;

Second step: normalized

Be normalized history data set, method for normalizing can adopt deviation standardized method, makes the data transformations of different dimension be unified processing format.Transfer function is as follows:

$x^{*}=\frac{x-x_{\min }}{x_{\max }-x_{\min }}$

Wherein x _maxfor the maximal value of sample data, x _minfor the minimum value of sample data;

3rd step: the foundation of forecast model and training

In RBF neural, if K is hidden layer neuron number, x (x ₁..., x _m) be input vector, α _ka kth hidden layer neuron and the neuronic connection weights of output layer, φ _kbe the output of a kth hidden layer neuron, therefore the output of RBF neural can be described as:

$y=\underset{k=1}{\overset{K}{\mathrm{\Σ}}}{\mathrm{\α}}_k{\mathrm{\φ}}_k\left(x\right)$

(1) a given hidden layer neuron is that the RBF neural of random natural number is trained.

(2) each hidden layer neuron output valve is calculated.The output that can be obtained a kth hidden layer neuron by formula (1) is:

${\mathrm{\φ}}_k\left(x\right)=e^{\frac{-\left|\right|x-{\mathrm{\μ}}_k\left|\right|}{{\mathrm{\σ}}_k^2}}$

(3) sensitivity analysis is carried out to the output of each neuron, calculate it to the contribution margin exported.The output weighted value of hidden layer neuron, as the input quantity of susceptibility method, utilizes following formula to calculate hidden layer neuron and exports the contribution done output neural network:

$S_h=\frac{{\mathrm{var}}_h\left[E\left(y\right|Z_h={\mathrm{\α}}_h{\mathrm{\φ}}_h\left(x\right))\right]}{\mathrm{var}\left(y\right)}$

Wherein, Z=[Z ₁, Z ₂..., Z _k] be the input vector of susceptibility method, y is neural network output quantity, and the relation of y and Z can be expressed as y=f (Z ₁, Z ₂..., Z _k), var _h[E (y|Z _h=α _hφ _h(x))] be Z _hequal α _hφ _hx, on the impact of y variance time (), var (y) is the variance of y, S _hα _hφ _hthe one order exporting y is represented.To input variable α _hφ _hcarry out Fourier transform (wherein α _hφ _hscope be [a _h, b _h]):

$Z_h=\frac{a_h+b_h}{2}+\frac{b_h-a_h}{\mathrm{\π}}\arcsin \left(\sin \left(w_{h}s\right)\right)$

Wherein, w _hbe the suitable frequency selected, get the Fourier's magnitude determinations sensitivity on fundamental frequency, formula 4 is finally deformed into conversion:

$S_h=\frac{(A_{{\mathrm{kw}}_h}^2+B_{{\mathrm{kw}}_h}^2)}{\underset{j=1}{\overset{+\∞}{\mathrm{\Σ}}}(A_j^2+B_j^2)}$

Wherein ,-π <s< π, $A_j=\frac{1}{2\mathrm{\&pi;}}{\&Integral;}_{-\&infin;}^{\&infin;}f\left(s\right)\&times;\cos \left(w_{j}s\right)\mathrm{ds},$ to S _hnormalized, note ST _hfor hidden layer neuron exports the contribution margin done output neural network:

${\mathrm{ST}}_h=\frac{S_h}{\underset{i=1}{\overset{K}{\mathrm{\&Sigma;}}}{S}_i}$

Definition error objective function is (N is number of training):

$E=\frac{1}{2N}\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}{e}_j^2$

(4) choose ε value and export contribution margin adjustment neural network structure according to hidden layer neuron.The value of ε is generally less than target error value, maximum and be greater than ε for contribution margin ₁hidden layer neuron divide, adjustment neural network structure; Suppose that the front hidden layer neuron number of division is K, the time of running is t, and contribution margin is greater than ε ₁hidden layer neuron be j, then the initial parameter of the neuron K+1 newly increased and the parameter of neuron j are:

a _K+1(t)＝λ×a _j(t)

μ _K+1(t)＝μ _j(t)

σ _K+1(t)＝σ _j(t)

$a_j^{,}\left(t\right)=(1-\mathrm{\&lambda;})\&times;a_j\left(t\right)$

${\mathrm{\&mu;}}_j^{,}\left(t\right)={\mathrm{\&mu;}}_j\left(t\right)$

${\mathrm{\&sigma;}}_j^{,}\left(t\right)={\mathrm{\&sigma;}}_j\left(t\right)$

Wherein, λ is the arbitrary constant (actual needs according to oil field sets) in (0,0.3), and the constant neuronic parameter of structure adjusts according to formula (7) ~ (9).

(5) ε is less than for contribution margin ₂hidden layer neuron delete, adjustment neural network structure; Suppose that the time of running is t, contribution margin is less than ε ₂hidden layer neuron be i, the neuron nearest with neuron i Euclidean distance is ii, and delete neuron i, the parameter of neuron ii is:

$a_{\mathrm{ii}}^{,}\left(t\right)=a_{\mathrm{ii}}\left(t\right)=a_i\left(t\right)\frac{{\mathrm{\&phi;}}_i\left(x\right)}{{\mathrm{\&phi;}}_{\mathrm{ii}}\left(x\right)}$

${\mathrm{\&mu;}}_{\mathrm{ii}}^{,}\left(t\right)={\mathrm{\&mu;}}_{\mathrm{ii}}\left(t\right)$

${\mathrm{\&sigma;}}_{\mathrm{ii}}^{,}\left(t\right)={\mathrm{\&sigma;}}_{\mathrm{ii}}\left(t\right)$

Other neuronic parameters adjust according to formula (7) ~ (9).

(6) stop calculating when reaching anticipation error or calculation procedure, model training terminates.

4th step: residual GM

On the basis of the neural network model set up, the predicted value of training sample and actual output value are contrasted, calculates error sequence.Using error sequence as a Markovian process, carry out state demarcation, replace probability with frequency, the transition probability matrix of error of calculation state.

Suppose that training dataset state status number has k, namely state has S ₁, S ₂..., S _k.Suppose to be in S now _istate, next step transfers to S _jshape probability of state is designated as P _ij, then the transfer case that state is total can by following matrix representation:

Wherein, $\underset{j=1}{\overset{k}{\mathrm{\&Sigma;}}}{p}_{\mathrm{ij}}=p_{i1}+p_{i2}+...+p_{\mathrm{ik}}=1,i=\mathrm{1,2},...,k.$

Pass between transition probability matrix is: P ^(k)=P ^(k-1)× P, wherein P is a step transition probability matrix, P ^(k)for k walks transition probability matrix.

Pass between transfering state is: S ^(k)=S ⁽⁰⁾× P ^(k), wherein S ^(k)for the state vector after the transfer of k step, S ⁽⁰⁾for initial state vector.

5th step: modelling verification

Use detection data the set pair analysis model to test, if the error of the predicted value exported and actual correlative value reaches the expectation of expection, neural network model is trained successfully, and model can be utilized to predict oil production rate; Otherwise model training is immature, needs to return to the 3rd step and train.

6th step: oil production rate is predicted

Obtain real fundamentals of forecasting data, input in the RBF neural after the optimization trained, the output of RBF neural is the predicted value of field output.

Certainly, the above-mentioned embodiment of the present invention is only can not limit the present invention to explanation of the present invention, the change that those skilled in the art do in essential scope of the present invention, remodeling, interpolation or replacement, also should belong to protection scope of the present invention.

Claims (1)

1. one kind is characterized in that based on the oilfield production forecast method of dynamic radial basis function neural network, mainly comprises the following steps:

A. data are obtained

According to oil field actual conditions, determine to affect field output factor index, obtain history data set and be divided into training dataset and detect data set;

B. normalized

Be normalized history data set, method for normalizing can adopt deviation standardized method, and make the data transformations of different dimension be unified processing format, transfer function is as follows:

$x^{*}=\frac{x-x_{\min }}{x_{\max }-x_{\min }}$ (formula 1)

Wherein x _maxfor the maximal value of sample data, x _minfor the minimum value of sample data;

C. the foundation of forecast model and training

In RBF neural, if K is hidden layer neuron number, x (x ₁..., x _m) be input vector, α _ka kth hidden layer neuron and the neuronic connection weights of output layer, φ _kbe the output of a kth hidden layer neuron, therefore the output of RBF neural can be described as:

$y=\underset{k=1}{\overset{K}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_k{\mathrm{\&phi;}}_k\left(x\right)$ (formula 2)

(1) a given hidden layer neuron is that the RBF neural of random natural number is trained, the number of times of setting training;

(2) calculate each hidden layer neuron output valve, the output that can be obtained a kth hidden layer neuron by formula (1) is:

${\mathrm{\&phi;}}_k\left(x\right)=e^{\frac{-\left|\right|x-{\mathrm{\&mu;}}_k\left|\right|}{{\mathrm{\&sigma;}}_k^2}}$ (formula 3)

(3) sensitivity analysis is carried out to the output of each neuron, calculate it to the contribution margin exported, the output weighted value of hidden layer neuron, as the input quantity of susceptibility method, utilizes following formula to calculate hidden layer neuron and exports the contribution done output neural network:

$S_h=\frac{{\mathrm{var}}_h\left[E\left(y\right|Z_h={\mathrm{\&alpha;}}_h{\mathrm{\&phi;}}_h\left(x\right))\right]}{\mathrm{var}\left(y\right)}$ (formula 4)

Wherein, Z=[Z ₁, Z ₂..., Z _k] be the input vector of susceptibility method, y is neural network output quantity, and the relation of y and Z can be expressed as y=f (Z ₁, Z ₂..., Z _k), var _h[E (y|Z _h=α _hφ _h(x))] be Z _hequal α _hφ _hx, on the impact of y variance time (), var (y) is the variance of y, S _hα _hφ _hthe one order exporting y is represented, to S _hbe normalized:

${\mathrm{ST}}_h=\frac{S_h}{\underset{i=1}{\overset{K}{\mathrm{\&Sigma;}}}{S}_i}$ (formula 5)

(4) choose ε value and export contribution margin adjustment neural network structure according to hidden layer neuron, the value of ε is generally less than target error value, maximum and be greater than ε for contribution margin ₁hidden layer neuron divide, ε is less than for contribution margin ₂hidden layer neuron delete, ε here ₁> ε ₂, final realization adjusts neural network structure, and definition error objective function is (N is number of training):

$E=\frac{1}{2N}\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}{e}_j^2$ (formula 6)

(5) according to target error function, utilize gradient descent algorithm to adjust output weights, the central value sum functions width of the hidden layer neuron of neural network:

$a_i(t+1)=a_i\left(t\right)-{\mathrm{\&eta;}}_1{\mathrm{ST}}_i\frac{\&PartialD;E}{{\&PartialD;a}_i\left(t\right)}$ (formula 7)

${\mathrm{\&mu;}}_i(t+1)={\mathrm{\&mu;}}_i\left(t\right)-{\mathrm{\&eta;}}_2{\mathrm{ST}}_i\frac{\&PartialD;E}{{\&PartialD;\mathrm{\&mu;}}_i\left(t\right)}$ (formula 8)

${\mathrm{\&sigma;}}_i(t+1)={\mathrm{\&sigma;}}_i\left(t\right)-{\mathrm{\&eta;}}_3{\mathrm{ST}}_i\frac{\&PartialD;E}{{\&PartialD;\mathrm{\&sigma;}}_i\left(t\right)}$ (formula 9)

Wherein, η ₁, η ₂, η ₃for parameter learning step-length;

(6) stop calculating when reaching anticipation error or calculation procedure;

D. residual GM

On the basis of the neural network model set up, the predicted value of training sample and actual output value are contrasted, calculates error sequence, using error sequence as a Markovian process, carry out state demarcation, replace probability with frequency, the transition probability matrix of error of calculation state;

E. modelling verification

Use detection data the set pair analysis model to test, if the error of the predicted value exported and actual correlative value reaches the expectation of expection, neural network model is trained successfully, and model can be utilized to predict oil production rate; Otherwise model training is immature, need to re-start training;

F. oil production rate prediction

Obtain real fundamentals of forecasting data, input in the RBF neural after the optimization trained, the output of RBF neural is the predicted value of field output.