CN104732303A

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

Info

Publication number: CN104732303A
Application number: CN201510165643.6A
Authority: CN
Inventors: 李克文; 王义龙; 苏玉亮
Original assignee: China University of Petroleum East China
Current assignee: China University of Petroleum East China
Priority date: 2015-04-09
Filing date: 2015-04-09
Publication date: 2015-06-24

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

Oil field yield prediction method based on dynamic radial basis function neural network

Technical Field

The invention relates to an oil field yield prediction method based on a dynamic radial basis function neural network, in particular to a method for dynamically optimizing the structure of the radial basis function neural network by a sensitivity method, correcting residual errors by combining a state transition probability method and realizing the oil field yield prediction.

Background

Petroleum is used as the life line of national economy, and the economic development of the country is directly influenced by the yield of petroleum. For oil field production, a high and stable oil production is necessary to ensure a good economic efficiency. Ensuring high and stable yield of the oil field is a central task of oil field development and production. Therefore, accurate prediction of oil production from an oil field has been one of the important research tasks for oil field developers.

Factors affecting oil field oil production are largely divided into two categories, geological factors and human factors. Geological factors are, to some extent, unchangeable or slightly altered. The variation range of human factors is much wider, and the variation of each human factor influences the variation of oil yield of an oil field from the exploitation mode, well pattern, well spacing, injection and production strength, well drilling adjustment, shutdown and abandonment of old wells or transfer injection and transfer production and the like to various manual measures (including fracturing, acidizing, hole repairing and layer adjusting, electric pump modification, hydraulic pump modification, overhaul and the like). Therefore, the methods for predicting oil production of oil fields are also based on geological factors, human factors or a combination of the two, and are mainly divided into two types: one is from the viewpoint of system theory, the oil yield of the oil field is researched and predicted as a whole; the other is to study the stimulation effect of a single measure. The first category of methods mainly includes statistical formula methods (empirical methods), water flooding characteristic curve methods, material balance equation methods, and reservoir numerical simulation methods. However, the statistical formula method (empirical method), the water flooding characteristic curve method and the material balance equation method have certain defects: firstly, the influence of the heterogeneity of a reservoir on the oil yield of an oil field cannot be directly considered; secondly, the influence of the change of various human factors on the oil field yield cannot be considered. Theoretically, the numerical reservoir simulation method can comprehensively and directly consider the influence of geological factors of a reservoir and changes of various human factors on the yield of an oil field. However, the dependence on geological data is too large, and errors exist in the recognition of reservoir geological conditions, so that the prediction result of numerical simulation of the oil and gas reservoir has no use value. Accurate fitting of an oil field development history often requires researchers to have both solid geological knowledge, reservoir knowledge, oil recovery process knowledge, mathematical computation knowledge, and computer knowledge, and is also labor intensive. The second method is to study the effect of various human factors on oil production of an oil field in isolation, which is against the fact that oil field development is a large system. Oil field development is a complex nonlinear power system, and prediction of oil field oil yield is a multi-factor nonlinear prediction problem. Therefore, in order to ensure the scientificity and accuracy of oil field yield prediction, a new oil field yield prediction method is urgently needed, so that the yield prediction result is more accurate, objective and reasonable.

Disclosure of Invention

From the artificial intelligence perspective, the method optimizes and adjusts the Radial Basis Function (RBF) neural network structure by using a sensitivity method, trains the RBF neural network by using oilfield output influence factor data samples, and corrects the RBF neural network by using a state transition probability method, so that the output prediction result is more accurate, objective and reasonable.

In order to achieve the purpose, the oil field yield prediction method based on the dynamic radial basis function neural network mainly comprises the following steps:

A. obtaining data

Determining factors and indexes affecting the oil field yield according to the actual conditions of the oil field, acquiring a historical data set, and dividing the historical data set into a training data set and a detection data set;

B. normalization process

The historical data set is normalized, the normalization method can adopt a dispersion normalization method, so that data of different dimensions are converted into a uniform processing format, and the conversion function is as follows:

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

(formula 1)

Wherein x_maxIs the maximum value of the sample data, x_minIs the minimum value of the sample data;

C. establishment and training of prediction model

In the RBF neural network, let K be the number of hidden layer neurons, x (x)₁,…,x_m) Is an input vector, α_kIs the connecting weight, phi, of the kth hidden layer neuron and the output layer neuron_kIs the output of the kth hidden layer neuron, so the output of the RBF neural network can be described as:

(formula 2)

(1) And (3) giving a hidden layer neuron to train the RBF neural network of any natural number, and setting the training times.

(2) Each hidden layer neuron output value is computed. The output of the kth hidden layer neuron from equation (1) is:

(formula 3)

(3) And carrying out sensitivity analysis on each neuron output, and calculating the contribution value of each neuron output to the output. And (3) taking the output weighted value of the hidden layer neuron as the input quantity of the sensitivity method, and calculating the contribution of the hidden layer neuron output to the output of the neural network by using the following formula:

(formula 4)

Wherein Z ═ Z₁,Z₂,...,Z_k]Is the input vector of the sensitivity method, y is the output quantity of the neural network, and the relation between y and Z can be expressed as y ═ f (Z)₁,Z₂,...,Z_k)，var_h[E(y|Z_h＝α_hφ_h(x))]Is Z_hIs equal to alpha_hφ_h(x) The influence of time on the variance of y, var (y) is the variance of y, S_hIs alpha_hφ_hFirst order sensitivity representation to output y. To S_hAnd (3) carrying out normalization treatment:

(formula 5)

(4) Selecting values and adjusting the neural network structure according to the hidden layer neuron output contribution values. Is generally less than the target error value, and is greatest for the contribution value and greater than₁Is split for contribution values less than₂Deletion of hidden layer neurons, where₁>₂And finally, adjusting the neural network structure. Defining an error objective function as (N is the number of training samples):

(formula 6)

(5) According to the target error function, the output weight, the central value and the function width of the hidden layer neuron of the neural network are adjusted by using a gradient descent algorithm:

<math> <mrow> <msub> <mi>a</mi> <mi>i</mi> </msub> <mrow> <mo>(</mo> <mi>t</mi> <mo>+</mo> <mn>1</mn> <mo>)</mo> </mrow> <mo>=</mo> <msub> <mi>a</mi> <mi>i</mi> </msub> <mrow> <mo>(</mo> <mi>t</mi> <mo>)</mo> </mrow> <mo>-</mo> <msub> <mi>η</mi> <mn>1</mn> </msub> <msub> <mi>ST</mi> <mi>i</mi> </msub> <mfrac> <mrow> <mo>&PartialD;</mo> <mi>E</mi> </mrow> <mrow> <msub> <mrow> <mo>&PartialD;</mo> <mi>a</mi> </mrow> <mi>i</mi> </msub> <mrow> <mo>(</mo> <mi>t</mi> <mo>)</mo> </mrow> </mrow> </mfrac> </mrow> </math>

(formula 7)

<math> <mrow> <msub> <mi>μ</mi> <mi>i</mi> </msub> <mrow> <mo>(</mo> <mi>t</mi> <mo>+</mo> <mn>1</mn> <mo>)</mo> </mrow> <mo>=</mo> <msub> <mi>μ</mi> <mi>i</mi> </msub> <mrow> <mo>(</mo> <mi>t</mi> <mo>)</mo> </mrow> <mo>-</mo> <msub> <mi>η</mi> <mn>2</mn> </msub> <msub> <mi>ST</mi> <mi>i</mi> </msub> <mfrac> <mrow> <mo>&PartialD;</mo> <mi>E</mi> </mrow> <mrow> <msub> <mrow> <mo>&PartialD;</mo> <mi>μ</mi> </mrow> <mi>i</mi> </msub> <mrow> <mo>(</mo> <mi>t</mi> <mo>)</mo> </mrow> </mrow> </mfrac> </mrow> </math>

(formula 8)

<math> <mrow> <msub> <mi>σ</mi> <mi>i</mi> </msub> <mrow> <mo>(</mo> <mi>t</mi> <mo>+</mo> <mn>1</mn> <mo>)</mo> </mrow> <mo>=</mo> <msub> <mi>σ</mi> <mi>i</mi> </msub> <mrow> <mo>(</mo> <mi>t</mi> <mo>)</mo> </mrow> <mo>-</mo> <msub> <mi>η</mi> <mn>3</mn> </msub> <msub> <mi>ST</mi> <mi>i</mi> </msub> <mfrac> <mrow> <mo>&PartialD;</mo> <mi>E</mi> </mrow> <mrow> <msub> <mrow> <mo>&PartialD;</mo> <mi>σ</mi> </mrow> <mi>i</mi> </msub> <mrow> <mo>(</mo> <mi>t</mi> <mo>)</mo> </mrow> </mrow> </mfrac> </mrow> </math>

(formula 9)

Wherein eta is₁,η₂,η₃Step sizes are learned for the parameters.

(6) The calculation is stopped when the desired error or calculation step is reached.

D. Residual correction

On the basis of the established neural network model, the predicted value of the training sample is compared with the actual yield value, and an error sequence is calculated. And taking the error sequence as a Markov process, carrying out state division, replacing the probability with frequency, and calculating a transition probability matrix of the error state.

E. Model validation

The model is checked by using the detection data set, if the error between the output predicted value and the actual comparison value reaches the expected expectation, the neural network model is successfully trained, and the model can be used for predicting the petroleum yield; on the contrary, the model training is not mature and needs to be retrained.

F. Oil production prediction

And acquiring real prediction basic data, inputting the real prediction basic data into the trained and optimized RBF neural network, wherein the output of the RBF neural network is the predicted value of the oil field yield.

Compared with a common neural network prediction model, the method has the advantages that the judgment basis is more objective, the finally obtained neural network structure is more compact through the adjustment of the neural network structure, the self-adaptive capacity is good, and the evaluation result is more scientific, accurate, fair and reasonable.

Drawings

FIG. 1 is a flow chart of a method for predicting oilfield production based on a dynamic radial basis function neural network.

Detailed Description

The present invention is described in further detail below with reference to the attached drawings.

The first step is as follows: obtaining data

the second step is that: normalization process

And (3) carrying out normalization processing on the historical data set, wherein the normalization method can adopt a dispersion normalization method, so that the data of different dimensions are converted into a uniform processing format. The transfer function is as follows:

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

the third step: establishment and training of prediction model

(1) a hidden layer neuron is given to train an RBF neural network of any natural number.

wherein Z ═ Z₁,Z₂,...,Z_k]Is the input vector of the sensitivity method, y is the output quantity of the neural network, and the relation between y and Z can be expressed as y ═ f (Z)₁,Z₂,...,Z_k)，var_h[E(y|Z_h＝α_hφ_h(x))]Is Z_hIs equal to alpha_hφ_h(x) The influence of time on the variance of y, var (y) is the variance of y, S_hIs alpha_hφ_hFirst order sensitivity representation to output y. For input variable alpha_hφ_hFourier transform (where α_hφ_hIn the range of [ a_h,b_h]):

<math> <mrow> <msub> <mi>Z</mi> <mi>h</mi> </msub> <mo>=</mo> <mfrac> <mrow> <msub> <mi>a</mi> <mi>h</mi> </msub> <mo>+</mo> <msub> <mi>b</mi> <mi>h</mi> </msub> </mrow> <mn>2</mn> </mfrac> <mo>+</mo> <mfrac> <mrow> <msub> <mi>b</mi> <mi>h</mi> </msub> <mo>-</mo> <msub> <mi>a</mi> <mi>h</mi> </msub> </mrow> <mi>π</mi> </mfrac> <mi>arcsin</mi> <mrow> <mo>(</mo> <mi>sin</mi> <mrow> <mo>(</mo> <msub> <mi>w</mi> <mi>h</mi> </msub> <mi>s</mi> <mo>)</mo> </mrow> <mo>)</mo> </mrow> </mrow> </math>

Wherein, w_hThe sensitivity is calculated by taking the fourier amplitude at the fundamental frequency, and the final deformation of equation 4 is a transformation:

wherein-pi<s<π，

<math> <mrow> <msub> <mi>A</mi> <mi>j</mi> </msub> <mo>=</mo> <mfrac> <mn>1</mn> <mrow> <mn>2</mn> <mi>π</mi> </mrow> </mfrac> <msubsup> <mo>&Integral;</mo> <mrow> <mo>-</mo> <mo>∞</mo> </mrow> <mo>∞</mo> </msubsup> <mi>f</mi> <mrow> <mo>(</mo> <mi>s</mi> <mo>)</mo> </mrow> <mo>×</mo> <mi>cos</mi> <mrow> <mo>(</mo> <msub> <mi>w</mi> <mi>j</mi> </msub> <mi>s</mi> <mo>)</mo> </mrow> <mi>ds</mi> <mo>,</mo> </mrow> </math>

To S_hNormalization processing, ST_hFor the contribution value made by the hidden layer neuron output to the neural network output:

defining an error objective function as (N is the number of training samples):

(4) selecting values and adjusting the neural network structure according to the hidden layer neuron output contribution values. Is generally less than the target error value, and is greatest for the contribution value and greater than₁The hidden layer neuron splits and adjusts the neural network structure; assuming that the number of neurons in the pre-division hidden layer is K, the running time is t, and the contribution value is greater than₁The hidden layer neuron ofj, then the initial parameters of the newly added neuron K +1 and the parameters of the 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)

wherein λ is any constant (set according to the actual needs of the oil field) in (0, 0.3), and the parameters of the structurally invariant neurons are adjusted according to equations (7) to (9).

(5) For contribution values less than₂Deleting the hidden layer neuron and adjusting the neural network structure; assuming that the operation time is t, the contribution value is less than₂The hidden layer neuron is i, the neuron with the Euclidean distance to the neuron i is ii, the deleted neuron i, and the parameters of the neuron ii are as follows:

the parameters of other neurons are adjusted according to equations (7) to (9).

(6) And stopping calculation when the expected error is achieved or the calculation step is completed, and finishing the model training.

The fourth step: residual correction

Suppose that the number of states of the training data set is k, i.e. the state has S₁,S₂,…,S_k. Suppose now is at S_iState, next shifts to S_jThe probability of a state is denoted P_ijThen the total transition situation of the state can be represented by the following matrix:

wherein,

the relationship between the transition probability matrices is: p^(k)＝P^(k-1)xPP, where P is a one-step transition probability matrix, P^(k)A transition probability matrix for k steps.

The relationship between the transition states is: s^(k)＝S⁽⁰⁾×P^(k)In which S is^(k)For the state vector after k-step transfer, S⁽⁰⁾Is an initial state vector.

The fifth step: model validation

The model is checked by using the detection data set, if the error between the output predicted value and the actual comparison value reaches the expected expectation, the neural network model is successfully trained, and the model can be used for predicting the petroleum yield; otherwise, the model training is not mature, and the training needs to return to the third step again.

And a sixth step: oil production prediction

It is to be understood that the above-described embodiments of the present invention are illustrative only and not intended to be limiting, and that various changes, modifications, additions or substitutions which may be made by those skilled in the art without departing from the spirit and scope of the invention are intended to be included within the scope of the invention.

Claims

1. The oilfield production prediction method based on the dynamic radial basis function neural network is characterized by mainly comprising the following steps of:

A. obtaining data

B. normalization process

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

(formula 1)

C. establishment and training of prediction model

(formula 2)

(1) Giving a hidden layer neuron as an RBF neural network of any natural number for training, and setting the training times;

(2) calculating the output value of each hidden layer neuron, wherein the output of the kth hidden layer neuron can be obtained by the formula (1):

(formula 3)

(3) Carrying out sensitivity analysis on each neuron output, calculating a contribution value of each neuron output to the output, taking an output weighted value of a hidden layer neuron as an input quantity of a sensitivity method, and calculating the contribution of the hidden layer neuron output to the neural network to the output by using the following formula:

(formula 4)

Wherein Z ═ Z₁,Z₂,...,Z_k]Is the input vector of the sensitivity method, y is the output quantity of the neural network,the relationship between y and Z may be expressed as y ═ f (Z)₁,Z₂,...,Z_k)，var_h[E(y|Z_h＝α_hφ_h(x))]Is Z_hIs equal to alpha_hφ_h(x) The influence of time on the variance of y, var (y) is the variance of y, S_hIs alpha_hφ_hFirst order sensitivity representation to output y, for S_hAnd (3) carrying out normalization treatment:

(formula 5)

(4) Selecting value and adjusting the neural network structure according to the contribution value output by the hidden layer neuron, wherein the value is generally smaller than the target error value, and the contribution value is maximum and larger than₁Is split for contribution values less than₂Deletion of hidden layer neurons, where₁>₂Finally, adjusting the neural network structure, and defining an error objective function as (N is the number of training samples):

(formula 6)

(formula 7)

(formula 8)

(formula 9)

Wherein eta is₁,η₂,η₃Learning a step size for the parameter;

(6) stopping the calculation when the expected error is reached or the calculation step is completed;

D. residual correction

On the basis of the established neural network model, comparing a predicted value and an actual yield value of a training sample, calculating an error sequence, taking the error sequence as a Markov process, carrying out state division, replacing probability with frequency, and calculating a transition probability matrix of the error state;

E. model validation

The model is checked by using the detection data set, if the error between the output predicted value and the actual comparison value reaches the expected expectation, the neural network model is successfully trained, and the model can be used for predicting the petroleum yield; on the contrary, the model training is immature and needs to be retrained;

F. oil production prediction