CN115660221A

CN115660221A - Oil and gas reservoir economic recoverable reserve assessment method and system based on hybrid neural network

Info

Publication number: CN115660221A
Application number: CN202211560212.6A
Authority: CN
Inventors: 张剑; 郝翱枭; 杨云; 李坤; 盛行; 李梓涵
Original assignee: Southwest Petroleum University
Current assignee: Southwest Petroleum University
Priority date: 2022-12-07
Filing date: 2022-12-07
Publication date: 2023-01-31
Anticipated expiration: 2042-12-07
Also published as: CN115660221B

Abstract

The invention discloses an oil and gas reservoir economic recoverable reserve assessment method and system based on a hybrid neural network, wherein the method comprises the steps of obtaining original data in the development process of an oil and gas reservoir; respectively preprocessing the acquired original data according to the data types; analyzing the importance of each type of feature data set by adopting a neural network model; constructing a hybrid neural network model, and performing model training on the hybrid neural network model by taking main characteristic variables and composite characteristic variables of various characteristic data sets as input characteristic variables; and predicting the economic recoverable reserves of the oil and gas reservoir according to the trained mixed neural network model. According to the method, various types of characteristic parameters are taken as influence factors, characteristic selection and characteristic extraction are carried out in a mode suitable for different types of data, and various types of data are obtained through characteristic fusion, so that the accuracy of the reserve evaluation result can be effectively improved.

Description

Oil-gas reservoir economic recoverable reserve evaluation method and system based on hybrid neural network

Technical Field

The invention relates to the technical field of economic recoverable reserve assessment of oil and gas reservoirs, in particular to an oil and gas reservoir economic recoverable reserve assessment method and system based on a hybrid neural network.

Background

Reserve assessment is a very important task in the development and management of oil and gas reservoirs, and particularly, economic recoverable reserves directly affect the economic benefit and profitability of oil field companies. The method is particularly important for scientifically and accurately evaluating the economical and collectable reserves, and has important practical significance for making medium-long term development plans, improving development benefits and ensuring sustainable development.

Economic reserves refer to reserves recognized by economic evaluation as having commercial benefits over a period of time. The method is generally characterized in that the evaluation is carried out according to the international oil and gas prices issued by famous oil and gas fields with similar oil and gas properties and the current market conditions in the evaluation period, the recoverable reserves are confirmed to be feasible in mining technology, reasonable in economy, allowed by other conditions such as environment and the like, and the reserves and the gains can meet the requirement of investment return in the evaluation period.

Reserve assessment is similar to yield prediction, and conventional methods mainly include statistical analysis methods such as yield decreasing curve method and water drive curve method, including conventional Arps decreasing prediction model and decreasing analysis model improved on the basis of the conventional Arps decreasing prediction model. However, the curve fitting method is performed on the basis of many assumptions and experiences, has more subjective factors, does not consider the influence of various objective conditions, and the obtained result lacks stability and generalization along with the change of experience or human factors, and the result may have deviation when applied to different oil wells. In addition, the oil reservoir numerical simulation method is a main method for predicting oil reservoir development indexes at home and abroad at present, is a typical physical driving data analysis method, can consider more factors more carefully, and has a more objective prediction result than an oil reservoir engineering. In addition, another type of method for economically recoverable reserves is a formula method that considers only economic influencing factors, such as the cash flow method, the economic limit method, the well pattern density method, the marginal cost method, the analog method, and the like. Although these methods have some feasibility, they are subjectively influenced by human subjects and have differences in the applicability of different reservoirs, which limits their applicability to problems involving mass data sets.

Machine learning and neural network technology are the mainstream development direction in the field of artificial intelligence, and the analysis of oil and gas data by using related algorithms becomes a research hotspot. Although some algorithms have been successfully applied to the fields of lithology identification, well log interpretation, oil price prediction, etc., the application of such algorithms to economic recoverable reserve assessment has been inadequate and typically only a single type of data is considered as input. The economic recoverable reserve evaluation is a complex multivariable nonlinear system, has multiple data types and large data volume, relates to parameters in multiple aspects such as oil reservoir geological characteristics, mining processes, historical yield, economic factors and the like, and a single prediction system may have poor effect.

Disclosure of Invention

Aiming at the defects in the prior art, the invention provides an oil and gas reservoir economic recoverable reserve evaluation method and system based on a hybrid neural network.

In order to achieve the purpose of the invention, the invention adopts the technical scheme that:

in a first aspect, the invention provides a method for evaluating the economic recoverable reserves of an oil and gas reservoir based on a hybrid neural network, which comprises the following steps:

s1, acquiring original data in an oil and gas reservoir development process;

s2, respectively preprocessing the acquired original data according to data types to obtain various feature data sets;

s3, analyzing the importance of each type of feature data set by adopting a neural network model, dividing feature variables in each type of feature data set into a main feature variable, a secondary feature variable and an invalid feature variable according to the importance, and selecting the secondary feature variable to construct a composite feature variable;

s4, constructing a hybrid neural network model, and performing model training on the hybrid neural network model by using main characteristic variables and composite characteristic variables of various characteristic data sets as input characteristic variables;

and S5, predicting the economic recoverable reserves of the oil and gas reservoirs according to the trained mixed neural network model.

Optionally, the raw data comprises:

geological parameter data, process parameter data, yield history data, economic parameter data and logging curve data.

Optionally, step S2 specifically includes the following sub-steps:

s2-1, performing data cleaning, data filling and standardization processing on numerical data of geological parameter data and process parameter data, coding classification labels of the geological parameter data and the process parameter data by adopting one-hot coding, and combining the classification labels with an economic and collectable reserve value to obtain a first-class characteristic data set;

s2-2, performing data cleaning, data filling and standardization processing on the yield historical data and the economic parameter data, and combining the data with the economic recoverable reserve value to obtain a second type characteristic data set;

and S2-3, carrying out data cleaning, data filling and standardization processing on the logging curve data, and combining the data with the economic recoverable reserve value to obtain a third type characteristic data set.

Optionally, step S3 specifically includes the following sub-steps:

s3-1, constructing a full-connection neural network model, setting hyper-parameters, and training and testing the full-connection neural network model by using a first class characteristic data set to obtain model parameters when the iteration of the full-connection neural network model is finished;

s3-2, calculating the importance characteristic value of each characteristic variable according to the model parameters obtained in the step S3-1;

s3-3, dividing all the characteristic variables into a main characteristic variable, a secondary characteristic variable and an invalid characteristic variable according to the importance characteristic value calculated in the step S3-2, and constructing a composite characteristic variable according to the secondary characteristic variable;

and S3-4, processing the economic parameters in the second type characteristic data set by adopting the methods from the steps S3-1 to S3-3 to obtain main characteristic variables and composite characteristic variables corresponding to the economic parameters, taking the yield historical data in the second type characteristic data set and the main characteristic variables corresponding to the economic parameters as the main characteristic variables of the second type characteristic data set, and taking the composite characteristic variables corresponding to the economic parameters as the composite characteristic variables of the second type characteristic data set.

Optionally, the model parameters include:

an input layer-hidden layer connection weight, a hidden layer-output layer connection weight, a value of an input layer neuron, an output value of a hidden layer neuron, an output value of an output layer neuron.

Optionally, step S3-2 specifically includes the following sub-steps:

s3-2-1, selecting a first number of fully-connected neural network models with optimal performance from test results of the fully-connected neural network models;

s3-2-2, calculating a total connection weight value corresponding to each characteristic variable according to the input layer-hidden layer connection weight and the hidden layer-output layer connection weight corresponding to each characteristic variable

；

S3-2-3, calculating the connection weight product corresponding to each characteristic variable according to the input layer-hidden layer connection weight and the hidden layer-output layer connection weight corresponding to each characteristic variable

；

S3-2-4, connecting weights of the input layer and the hidden layer according to the input layer and the hidden layer corresponding to each characteristic variable

Input layer neuron values

And output values of hidden layer neurons

Calculate input layer numberiThe first neuron to the hidden layerjInfluence value of individual neuron

；

S3-2-5, connecting weights of hidden layers and output layers according to the hidden layer-output layer corresponding to each characteristic variable

Hidden layer neuron output values

And output values of output layer neurons

Computing hidden layer numberjInfluence value of individual neurons on output layer neurons

；

S3-2-6, calculating the input layer number corresponding to each characteristic variable by adopting the following formulaiInfluence value of each neuron on the output value:

wherein the content of the first and second substances,Mthe number of neurons in the hidden layer;

s3-2-7, connecting the absolute value of the weight according to the input layer-hidden layer corresponding to each characteristic variable

Absolute value of hidden layer-output layer connection weight

Calculating the product thereof

；

S3-2-8, calculating the importance of each characteristic variable to each neuron of the hidden layer based on the weight absolute value by adopting the following formula

Wherein, in the step (A),Nthe number of neurons in the input layer;

s3-2-9, calculating the importance characteristic value of each characteristic variable to the output layer based on the weight absolute value by adopting the following formula:

wherein the content of the first and second substances,Mthe number of hidden layer neurons.

Optionally, step S3-3 specifically includes the following sub-steps:

s3-3-1, respectively calculating according to the calculation results of the step S3-2

、

、

、

Average value of (2)

，

，

，

；

S3-3-2, respectively calculating the relative importance characteristic value of each characteristic variable by adopting the following formula:

wherein the content of the first and second substances,nis a characteristic variable serial number;

s3-3-3, respectively pairing all characteristic variables

，

，

，

Four times of sorting is carried out according to the sizes to generate four relative importance sorting tables;

s3-3-4, dividing the four relative importance ranking tables into a main characteristic variable, a secondary characteristic variable and an invalid characteristic variable according to a threshold value;

and S3-3-5, calculating the mean value of the four relative importance characteristic values of each secondary characteristic variable and the weight mean value proportion of the four relative importance characteristic values in all the secondary characteristic variables, and performing weighted summation on all the secondary characteristic variables to construct a composite characteristic variable.

Optionally, the hybrid neural network model specifically includes:

the device comprises a parallel feature extraction channel and a full-connection neural network, wherein the parallel feature extraction channel consists of a first feature extraction channel, a second feature extraction channel and a third feature extraction channel;

the first feature extraction channel is used for inputting main feature variables and composite feature variables of the first-class feature data set, and extracting numerical values of geological parameters and process parameters and feature vectors of classification labels;

the second feature extraction channel is used for inputting main feature variables and composite feature variables of a second type of feature data set and extracting feature vectors of economic parameters and yield historical data;

the third feature extraction channel is used for inputting a third type of feature data set and extracting feature vectors of logging curve data;

and the fully-connected neural network is used for performing characteristic cascade on the characteristic vectors extracted by the parallel characteristic extraction channel connection to serve as input characteristic vectors and predicting the economic recoverable reserves of the oil and gas reservoirs.

Optionally, the hybrid neural network model is iteratively trained by using an optimized adaptive momentum method, where an iterative update formula is:

wherein the content of the first and second substances,

is the network weight at the next iteration,

for the network weight of the current iteration,

、

in order to adapt the hyper-parameters adaptively,

is a matrix of partial derivatives of the network error against the weights,

is that

The transpose matrix of (a) is,

is a scale factor, m is a penalty factor,Iis an identity matrix, E is a network error vector,

is the network weight at the last iteration,kis the current iteration number.

In a second aspect, the present invention provides a system for evaluating economic recoverable reserves of hydrocarbon reservoirs based on a hybrid neural network, which applies the method described above, and comprises:

the data acquisition module is used for acquiring original data in the development process of the oil and gas reservoir;

the data preprocessing module is used for respectively preprocessing the acquired original data according to data types to obtain various feature data sets;

the characteristic extraction module is used for analyzing the importance of each type of characteristic data set by adopting a neural network model, dividing characteristic variables in each type of characteristic data set into a main characteristic variable, a secondary characteristic variable and an invalid characteristic variable according to the importance, and selecting the secondary characteristic variable to construct a composite characteristic variable;

the model training module is used for constructing a hybrid neural network model and performing model training on the hybrid neural network model by taking the main characteristic variables and the composite characteristic variables of various characteristic data sets as input characteristic variables;

and the data prediction module is used for predicting the economic recoverable reserves of the oil and gas reservoir according to the trained hybrid neural network model.

The invention has the following beneficial effects:

according to the method, various types of characteristic parameters are taken as influence factors, characteristic selection and characteristic extraction are carried out in a mode suitable for different types of data, and various types of data are obtained through characteristic fusion, so that the accuracy of the reserve evaluation result can be effectively improved.

Drawings

FIG. 1 is a schematic flow chart of a method for evaluating the economic recoverable reserves of a hydrocarbon reservoir based on a hybrid neural network in embodiment 1 of the present invention;

FIG. 2 is a schematic structural diagram of a hybrid neural network model in embodiment 1 of the present invention;

fig. 3 is a schematic structural diagram of a system for evaluating an economic recoverable reserve of a hydrocarbon reservoir based on a hybrid neural network in embodiment 2 of the present invention.

Detailed Description

The following description of the embodiments of the present invention is provided to facilitate the understanding of the present invention by those skilled in the art, but it should be understood that the present invention is not limited to the scope of the embodiments, and it will be apparent to those skilled in the art that various changes may be made without departing from the spirit and scope of the invention as defined and defined in the appended claims, and all matters produced by the invention using the inventive concept are protected.

Example 1

As shown in fig. 1, an embodiment of the present invention provides a method for evaluating an economically recoverable reserve of a hydrocarbon reservoir based on a hybrid neural network, including the following steps S1 to S5:

s1, acquiring original data in an oil and gas reservoir development process;

in an alternative embodiment of the present invention, in considering the problem of economic recoverable reserve evaluation, it is desirable to determine the characteristics of the data relevant to the evaluation, and how to select and use the data. New sensor technologies are able to stream large amounts, multi-scale and high dimensional reservoir data into databases in real time. The data types are extremely diverse, the data are learned from multi-source and multi-type data, more details are included, the corresponding relation among various data can be obtained, in order to embody all '7V' characteristics of the big data, namely quantity, speed, diversity, variability, authenticity, visualization and value, how to combine the various types of data, how to process different levels of noise and how to process lost data, the capacity of representing the data in a meaningful mode is important for the problem of economic and recoverable reserve assessment, and different data processing modes are required for different types of data.

The invention uses various sources and various types of mass data sets in the economic recoverable reserves evaluation problem, relates to parameters in multiple aspects such as oil deposit geological characteristics, mining process, historical yield, economic factors and the like, is comprehensive and objective, considers various types of characteristic parameters as input, extracts characteristics in a mode suitable for different types of data, and then performs characteristic fusion, and various types of data contribute to the accuracy of reserves evaluation results.

The initial characteristic selection is based on the oil well production theory and factors which most possibly influence the oil well production, the real reserve related data of a certain oil field is obtained, the reserve related data comprises reserve data of a plurality of evaluation units, and the characteristic variables collected by each oil well unit are divided into the following three types of corresponding input characteristic data.

The first type is: values and classification labels for geological and process parameters. The method comprises information such as well blocks, horizons, porosity, permeability, original formation pressure, crude oil viscosity, oil-containing area, oil-containing saturation, crude oil geological storage, the number of oil production wells put into production in the current year, production days, reservoir lithology, physical property classification, production mode and the like. Such data is discrete and represents the overall effect of the reservoir.

The second type: production history data and economic parameters. In the actual production process of an oil field, a lot of time series data are generated, and the data record one or more variable quantities at the past moment. The historical data of the yield is counted according to the month and comprises the information of oil yield, gas yield, water injection amount, well opening number and the like of each month; the economic parameters comprise information such as operation cost, fixed cost ratio, oil price, gas price, exchange rate, special income, tax, well construction cost, ton oil cost, input-output ratio, net cash flow, internal income rate, financial net present value, recovery period, million tons of capacity investment and the like.

In the third category: and (3) a logging curve needs to be extracted according to the depth range of each reservoir, and according to the limitation of actual measurement, only a part of the measurement is accepted as input. In the normal section of the borehole, a good mapping relation exists between each logging curve and the formation characteristics, the detection depth is deep, the more curves which are not influenced by the collapse of the borehole are, and the better the modeling effect is. The normal section of the well diameter is selected as sample data, and the well logging data quality is high to carry out experiments. The correspondence between these log data is correct. And the logging curve comprises information such as acoustic time difference, natural gamma, resistivity, density, neutrons, natural potential and the like. The structured data of the well logging curve contains more reservoir details and has dual information of time and space.

in an optional embodiment of the present invention, step S2 specifically includes the following sub-steps:

s2-1, carrying out data cleaning, data filling and standardization processing on numerical data of geological parameter data and technological parameter data, coding classification labels of the geological parameter data and the technological parameter data by adopting one-hot coding, and combining the classification labels with an economic recoverable reserve value to obtain a first-class characteristic data set;

Specifically, due to different data collection modes or different emergency situations of collection personnel, certain redundant data, default values and abnormal values of the data of the oil field can occur. For the repeated data, on the basis of the known content of the repeated data, one record is taken out from each repeated data and is reserved, and other repeated data is deleted. Abnormal values such as zero values and bad values are processed by a data cleaning technology and filled by a local mean value or a mode, so that the instability of a prediction result caused by the abnormal values is reduced to the maximum extent.

Data preprocessing is crucial to deep learning, and requires conversion of a prepared sampled data set into a data format that can be input into a hybrid neural network model. Different types of data are measurements of different properties, with different ranges of values. In order to eliminate the influence caused by different dimensions of different variables, the variables need to be converted into the same dimension, and the data is standardized in consideration of certain outliers of the sampled data. The normalization process formula is as follows:

wherein the content of the first and second substances,

is the average of some kind of input data,

is the standard deviation of some input data. To increase the convergence rate of the model, the present document deals with the output data (i.e., the economic recoverable reserve)The numerical value of (b) is also subjected to standardization, the reserve predicted by the model belongs to the standardized range, and the predicted value of the model also needs to be subjected to anti-standardization to obtain the real economical recoverable reserve.

It should be noted that the normalization process is used for data of numerical type, so that the numerical values of geological parameters and process parameters, yield history data and economic parameters, well logs can and need to be normalized, and the classification labels of the geological parameters and process parameters cannot be digitized, so that the normalization process cannot be adopted, and the classification labels of the geological parameters and process parameters are converted into a data format in which the model can perform learning and feature extraction by adopting a one-hot coding process. The advantage of the one-hot coding is that the classification value is mapped to the integer value, the non-continuous numerical characteristic can be processed, the one-hot coding is used for the discrete characteristic, and the distance calculation between the characteristics is more reasonable.

One-hot coding mode:

for example, physical property classification: [ "low permeability", "medium permeability", "high permeability" ], "low permeability" is mapped to 100, "medium permeability" is mapped to 010, "high permeability" is mapped to 001;

such as land-sea classification: [ "land", "beach" ], "land" is mapped to 10, and "beach" is mapped to 01;

therefore, when a sample is "medium penetration", "beach" ], the complete feature digitization results as: [0,1,0,0,1].

The invention combines the preprocessed characteristic values with the economically collectable and storable values to form a data set required by an experiment. When the hybrid neural network model is constructed and trained, a training set, a verification set and a test set are divided according to the proportion of 8. The training set is used for training the model, the verification set is used for calculating a loss function in the training process, and the test set is used for evaluating the prediction effect of the model.

S3, analyzing the importance of each type of feature data set by adopting a neural network model, dividing feature variables in each type of feature data set into a main feature variable, a secondary feature variable and an invalid feature variable according to the importance, and selecting the secondary feature variables to construct a composite feature variable;

in an optional embodiment of the present invention, step S3 specifically includes the following sub-steps:

s3-1, constructing a full-connection neural network model, setting hyper-parameters, and training and testing the full-connection neural network model by using a first class of feature data set to obtain model parameters when the iteration of the full-connection neural network model is finished; wherein the model parameters include: the input layer-hidden layer connection weight, the hidden layer-output layer connection weight, the value of the input layer neuron, the output value of the hidden layer neuron and the output value of the output layer neuron;

s3-2, calculating the importance characteristic value of each characteristic variable according to the model parameters obtained in the step S3-1; the method specifically comprises the following steps:

s3-2-2, calculating a connection weight total value corresponding to each characteristic variable according to the input layer-hidden layer connection weight and the hidden layer-output layer connection weight corresponding to each characteristic variable

；

；

S3-2-4, connecting weights according to the input layer-hidden layer corresponding to each characteristic variable

Input layer neuron values

And output values of hidden layer neurons

Calculate input layer numberiThe first neuron pair of hidden layerjInfluence value of individual neuron

；

S3-2-5, connecting weights according to hidden layer-output layer corresponding to each characteristic variable

Hidden layer neuron output values

And output values of output layer neurons

；

wherein, the first and the second end of the pipe are connected with each other,Mthe number of hidden layer neurons;

Absolute value of hidden layer-output layer connection weight

Calculating the product thereof

；

S3-2-8, calculating each characteristic variable by adopting the following formulaImportance of each neuron of hidden layer based on weight absolute value

Wherein, in the process,Nthe number of neurons in the input layer;

wherein the content of the first and second substances,Mthe number of hidden layer neurons;

s3-3, dividing all the characteristic variables into a main characteristic variable, a secondary characteristic variable and an invalid characteristic variable according to the importance characteristic value calculated in the step S3-2, and constructing a composite characteristic variable according to the secondary characteristic variable; the method specifically comprises the following steps:

、

、

、

Average value of (2)

，

，

，

；

wherein the content of the first and second substances,nserial number of the characteristic variable;

s3-3-3, respectively pairing all characteristic variables

，

，

，

s3-3-5, calculating the mean value of the four relative importance characteristic values of each secondary characteristic variable and the weight mean value proportion of the mean value in all the secondary characteristic variables, and performing weighted summation on all the secondary characteristic variables to construct a composite characteristic variable;

Specifically, since the hybrid neural network has a complex structure and many parameters, multi-source and multi-type data are not only required to be classified and respectively processed, importance analysis and feature selection are also required to be performed in each type of feature variable, a fully-connected neural network is used for predicting a target value to obtain a weight under an optimal result, the total value of the connection weight of each variable, the product of the connection weights, an influence value on an output value, and the importance comprehensive analysis based on the weight of an output layer are performed, so that variables with high relative importance are retained, variables with low relative importance are weakened and integrated, and variables with extremely low relative importance are removed, thereby improving the prediction accuracy.

When the model hyper-parameters are set, the value intervals of the model learning rate are set to be {0.001,0.003,0.005,0.01,0.03,0.05 and 0.1}, the value intervals of the hidden layer unit number are {5,6,7, \43, 35}, the data set randomly selects the ratio value intervals of {50%,55%,60%,65%,70%,75%,80%,85%,90%,95% and 100% }, the loss function is a Root Mean Square Error (RMSE) loss function, the weight initialization method is a Uniform distribution initialization method (Random uniformity), the maximum epoch number is 100, and the iteration is ended in advance when the continuous 5-time Error is less than 0.001.

In neural networks, the connection weights between neurons are the connection between input and output, and thus the connection between problem and solution. The relative importance of the independent variable characteristics to the prediction output of the neural network mainly depends on the magnitude and direction of the connection weight. However, in deep learning, processing high-dimensional vectors can cause a great consumption of computing resources and even cause a dimensional disaster problem. Therefore, there is a need to find ways to convert high-dimensional vectors to low-dimensional vectors without losing data features. In the conventional oil reservoir engineering, the characteristics are preliminarily analyzed, and the characteristics are reduced by using intuition and experience of the petroleum production engineering, but the method has strong subjectivity. The invention provides a feature importance analysis method which is used for counting and evaluating the contribution of input variables in a neural network, improving the resolving power of mass feature parameters, identifying variables which have important contribution to network prediction and further selecting the feature variables.

Firstly, importance analysis is carried out on numerical values and classification labels of first-class characteristic variables, geological parameters and process parameters, only the first-class characteristic variables are used as input, economic and recoverable reserves are used as output, a three-layer full-connection neural network is constructed, and training and testing are carried out. Since the network weight matrix initialization is random, the model training has some randomness, and if a particular variable in the dataset performs poorly, it is possible that, although this variable does not appear to be important in the experiment, it may become more important as the experiment hyper-parameter set or dataset selection changes, thus multiple random experiments are used. The invention sets value ranges of a plurality of hyper-parameters, constructs a plurality of three-layer fully-connected neural networks, and randomly selects and pairs the learning rate, the number of hidden layer units and the random selection proportion of data sets. This is repeated a plurality of times, for example 100 times. Because each connection weight in the network has a magnitude and a direction, in order to more comprehensively evaluate the importance variable, when the fully-connected neural network of each experiment reaches a critical condition or an optimal iteration state, the following values are recorded: (1) Input layer-hidden layer connection weights

(ii) a (2) Hidden layer-to-output layer connection weights

(ii) a (3) Input layer neuron i values

(ii) a (4) Output value for hidden layer neuron j

(ii) a Wherein

The value of 1,2 is selected,

n, the number of input layer neurons; value of j

I.e. the number of hidden layer neurons; (5) Output values for output layer neurons

。

After multiple randomization experiments are finished, 10 fully-connected neural networks with the best performance are selected, and the learning rate (1) and the hidden layer unit number (2) of the fully-connected neural networks are recorded. Their following values were calculated: (1) The total value of the connection weights, which includes the sum of all the connection weights of each variable from the input layer to the hidden layer and then to the output layer, is

(ii) a (2) Connecting the weight product including the connection weight of each variable input layer and hidden layer

Connecting weight value with hidden layer and output layer of each hidden layer neuron

The sum of the products of (a) is made to be

(ii) a (3) Influence value of ith neuron of input layer on jth neuron of hidden layer

Influence value of j-th neuron of hidden layer on neuron of output layer

(ii) a Calculating the output value of the ith neuron pair of the input layer

Influence value of

(ii) a (4) Absolute value of connection weight of each variable input layer and hidden layer

Absolute value of connected weights of hidden layer and output layer for each hidden layer neuron

Calculating the product thereof

Then calculating the importance of each variable to each neuron of the hidden layer based on the weight absolute value

Importance characteristic value of each variable to output layer based on weight absolute value

. Wherein i takes on a value

I.e. the number of input layer neurons; value of j

I.e., the number of hidden layer neurons. This in turn allows the statistical significance of the importance of each variable to be derived.

Screening out 10 full-connection gods with optimal performance in model experimentVia a network model. Find the 10

、

、

、

Average value of (2)

，

，

，

Calculating four relative importance characteristic values of each variable

，

，

，

Comprehensively analyzing the influence of the connection weight on the variables and the output from four angles, and respectively making all the variables pair

，

，

，

Four sorts by size and generates four sort tables of relative importance.

All variables are classified into primary characteristic variables (i.e., variables of higher importance), secondary characteristic variables (i.e., variables of lower importance), and null characteristic variables (i.e., variables of lowest importance) according to a certain criterion.

Since the output variable value is the result of the common influence of all the characteristic variables, the secondary characteristic variables contribute less, but not completely useless, with respect to the primary variable. Therefore, the invention does not abandon all the secondary characteristic variables, but weakens the connection weight of the secondary characteristic variables, and more clearly analyzes the influence of each input characteristic variable on the output variable.

Two percentage thresholds are set, one larger and one smaller. In four sorting processes, all the characteristic variables with relative importance exceeding a larger threshold are taken out and four sets are formed

The intersection of the four sets

All the variables in (2) are defined as main characteristic variables, all the characteristic variables with relative importance lower than a smaller threshold value are taken out, and four sets are formed

The intersection of the four sets

All the feature variables in (1) are defined as invalid feature variables, and the rest of feature variables are all defined as secondary feature variables. Or set a percentage thresholdThe feature variable reaching this threshold 2 to 4 times is defined as a primary feature variable, the feature variable reaching this threshold 1 time is defined as a secondary feature variable, and the feature variable reaching this threshold 0 time is defined as an invalid feature variable.

Averaging the four relative importance of each sub-feature variable

And their weight-mean ratios in all sub-feature variables

And L is the number of the secondary characteristic variables. And according to the weight proportion, carrying out weighted summation on the secondary characteristic variables and integrating the secondary characteristic variables into a new composite characteristic variable. The weight proportion of each secondary characteristic variable in the composite characteristic variables is determined, and the weights are not trained in the subsequent mixed neural network model training process. And taking the composite characteristic variable and the main characteristic variable together as input characteristic variables of the hybrid neural network model. Too many feature variable inputs complicate the model and increase the parameters of the neural network. In order to simplify the model, more data types are effectively utilized while the prediction precision is ensured, so that not only can all collected characteristic data be more fully considered to exert the value of each characteristic variable, but also the connection between low-importance characteristic variables and a neural network is reduced, the number of parameters during model training is reduced, and the model structure is optimized.

According to the invention, the main characteristic variables and the new characteristic variables obtained by analyzing the importance of the first-class characteristic data set are defined as the first-class main characteristic variables and the first-class composite characteristic variables.

And (4) performing importance analysis on the second type characteristic data set, the yield historical data and the economic parameters, wherein the yield historical data and the reserves have the most direct relationship and are all reserved, so that only the second type characteristic variables are taken as economic parameter input and the economically obtainable reserves are output, the method is similar to the steps S3-1 to S3-3, and finally the second type main characteristic variables (the yield historical data and the economic parameters with higher relative importance) and the second type composite characteristic variables (the composite variables of the economic parameters with lower relative importance) are obtained.

For the third type of characteristic data set, because the logging curve sequence is easy to have more missing values in practice, several curves with high logging data quality are screened out in the data preprocessing stage, and the curves comprise information such as acoustic wave time difference, natural gamma, resistivity, density, neutron, natural potential and the like, so that the importance analysis is not performed on the data, and all the curves are defined as third type main characteristic variables.

in an optional embodiment of the invention, a hybrid neural network (HDNN) model is provided for economic recoverable reserve evaluation in consideration of various types of data such as geological parameters, technological parameters, economic parameters, yield historical data and the like of a single well in the development process of a hydrocarbon reservoir.

The hybrid neural network model constructed by the invention specifically comprises the following steps:

the system comprises a first feature extraction channel, a second feature extraction channel, a third feature extraction channel, a full-connection neural network and a second feature extraction channel, wherein the first feature extraction channel, the second feature extraction channel and the third feature extraction channel form a parallel feature extraction channel;

The hybrid neural network model of the present invention processes each input type separately with a suitable network for feature learning. All learned features are then concatenated and integrated into a feature set. This integration contains valid information from different inputs, which are then input to subsequent neural networks for overall target learning.

The input data can be continuous or discrete, it is inconvenient to input all the data into one network, and in order to process mixed input, the invention applies various types of neural networks to different data types and integrates the data types.

Aiming at some geological parameters which are numerical values such as porosity, permeability, original formation pressure, crude oil viscosity, oil saturation and the like, and some geological parameters which are classification labels such as reservoir lithology, physical property classification and the like; some technological parameters also include the number of oil wells put into production in the current year, the number of days in production and other numerical values, and the classification labels such as the production mode and the like. The values and classification labels are discrete and represent the overall effect of the reservoir. Therefore, the invention uses a Fully Connected Neural Network (FCNN) to perform feature processing on the numerical values of the geological parameters and the technological parameters and the input of the classification labels.

The historical data of production and economic parameters are important data of the oil field for a long time, the data contain time series information, and the data record one or more variable quantities at the past moment, so the invention uses a Gated Current Unit (GRU) to extract the characteristics and mine the relationship of the historical data of economic parameters and production.

For some structural measurements, such as well logs, that contain more detail, a variety of sequence curves can be provided, reflecting cumulative effects in the reservoir, including both temporal and spatial information. Therefore, the invention uses a one-dimensional Convolutional Neural Network (CNN) which comprises a Conv1D Convolutional layer and a pooling layer, and a cyclic Convolutional Neural network combined with a Gated Recurrent Unit (GRU) to process the logging curve and perform feature extraction and relationship mining.

The method respectively performs characteristic learning on data with different formats, and then performs characteristic cascade on the output of the FCNN, the GRU and the cyclic convolution neural network to realize final evaluation on the economical and recoverable reserves.

Converting numerical values and classification labels of geological parameters and process parameters, economic parameters and yield historical data and logging curves into a characteristic vector A, a characteristic vector B and a characteristic vector C, inputting the characteristic vectors into a nonlinear mapping equation of learning characteristics, and respectively expressing the three data as follows:

wherein, the first and the second end of the pipe are connected with each other,

in order to be the FCNN,

in order to gate the circulation network, the control circuit is provided with a control circuit,

in order to be a circular convolutional neural network,

、

、

numerical and classification labels for geological and technological parameters, respectivelyThe features learned in the FCNN, the economic parameters and the production history data are input into the GRU, and the logging curves are input into the cyclic convolution neural network.

Performing feature cascade on all learned features to obtain an integrated feature, combining several features according to specific data types, and then representing the integrated feature as the input of a final economic recoverable reserve assessment model

Wherein Y represents the value of the economically recoverable reserve obtained by the final model,

is a layer of fully connected network.

On the basis of the characteristic extraction, the hybrid neural network (HDNN) model provided by the invention realizes an end-to-end deep learning model expressed as

Wherein F represents the nonlinear mapping from multi-source multi-class input to target value, that is, the HDNN model provided by the invention is formed by connecting FCNN, GRU and cyclic convolution neural network in parallel and then connecting with the target value

And (4) combining models after series connection. When the HDNN model is integrally trained, the weight W, the weight V and the threshold of the model are continuously optimized through a neural network training algorithm, and the model is trained to be the model with the optimal effect.

When the model training is carried out, the learning rate is set as the average value of the learning rate values of 10 fully-connected neural networks with optimal performance in the random result of the value range of the set model learning rate

And the average value of learning rate values of 10 fully-connected neural networks with optimal performance when the importance analysis is carried out on the economic parameters in the second class of characteristic variables

Average value of

(ii) a The number of the hidden layer units at the FCNN end is the average value (rounded down if decimal) of the number of the hidden layer units of 10 fully-connected neural networks with optimal performance in the random result of the value interval of the number of the hidden layer units, and the number of the hidden layer units at the GRU end is the average value (rounded down if decimal) of the number of the hidden layer units of 10 fully-connected neural networks with optimal performance when the importance analysis is carried out on the economic parameters in the second type of characteristic variables; the loss function is Root Mean Square Error (RMSE) loss function, the weight initialization method is a Uniform distribution initialization method (Random uniformity), the maximum epoch number is 100, and iteration is ended in advance when the Error is less than 0.001 for 5 consecutive times.

In order to avoid the over-fitting condition and increase the probability that the network model converges to the global optimal value, the invention adopts an optimized adaptive momentum method and carries out model training by introducing simple weight attenuation regularization. The variation of the weights is confined to the hyperellipse. The conjugacy is also maintained between all subsequent descent directions, where only the descent direction that sufficiently reduces the objective function is accepted, and the iterative update formula is expressed as:

wherein the content of the first and second substances,

is the network weight at the next iteration,

for the network weight of the current iteration,

、

in order to adapt the hyper-parameters adaptively,

is a matrix of partial derivatives of the network error against the weights,

is that

The transpose matrix of (a) is,

is the network weight at the last iteration,kis the current iteration number.

When model training is carried out, a first class main characteristic variable and a first class composite characteristic variable are input to an FCNN end in the proposed HDNN, a second class main characteristic variable and a second class composite characteristic variable are input to a GRU end in the proposed HDNN, a third class main characteristic variable is input to a circular convolution neural network end in the proposed HDNN, and a model architecture is shown in figure 2. Training by adopting an optimized adaptive momentum method (OLMAM) until the maximum iteration number is reached or the error requirement is reached; economic recoverable reserve assessment is a regression problem, and therefore Root Mean Square Error (RMSE) loss functions are used in the training process as criteria for evaluating the predicted effect during network training. The root mean square error is an index used to describe the "systematic error", and the smaller the value, the closer to 0, the better the performance of the model. The root mean square error loss function is calculated as follows:

wherein, the total number of samples is T,

is the first onetThe actual value of the economically exploitable reserves of an individual sample,

is the firsttAnd predicting the economic recoverable reserves of the samples.

When the model test is carried out, the test data set is input into the trained model, and the performance of the model is evaluated. Because all data is normalized, the root mean square error is calculated during the training process using the normalized data. Therefore, when the performance of the model is evaluated, the actual prediction effect of the model is reflected, firstly, the output of the model is subjected to anti-standardization processing to obtain a predicted economic recoverable reserve value, and then, the root mean square error is calculated by using the predicted value and the actual value.

Example 2

As shown in fig. 3, the present invention provides a hybrid neural network-based economic recoverable reservoir assessment system applying the above method based on the hybrid neural network-based economic recoverable reservoir assessment method described in embodiment 1, including:

and the data prediction module is used for predicting the economic recoverable reserves of the oil and gas reservoir according to the trained mixed neural network model.

The present invention has been described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each flow and/or block of the flowchart illustrations and/or block diagrams, and combinations of flows and/or blocks in the flowchart illustrations and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.

These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks.

These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.

The principle and the implementation mode of the invention are explained by applying specific embodiments in the invention, and the description of the embodiments is only used for helping to understand the method and the core idea of the invention; meanwhile, for a person skilled in the art, according to the idea of the present invention, there may be variations in the specific embodiments and the application scope, and in summary, the content of the present specification should not be construed as a limitation to the present invention.

It will be appreciated by those of ordinary skill in the art that the embodiments described herein are intended to assist the reader in understanding the principles of the invention and are to be construed as being without limitation to such specifically recited embodiments and examples. Those skilled in the art can make various other specific changes and combinations based on the teachings of the present invention without departing from the spirit of the invention, and these changes and combinations are within the scope of the invention.

Claims

1. A method for evaluating the economic recoverable reserves of oil and gas reservoirs based on a hybrid neural network is characterized by comprising the following steps:

s1, acquiring original data in an oil and gas reservoir development process;

s4, constructing a hybrid neural network model, and performing model training on the hybrid neural network model by using the main characteristic variables and the composite characteristic variables of various characteristic data sets as input characteristic variables;

2. The method of claim 1, wherein the raw data comprises:

3. The method for evaluating the economic recoverable reserves of the oil and gas reservoirs based on the hybrid neural network as claimed in claim 2, wherein the step S2 comprises the following sub-steps:

4. The method for evaluating the economic recoverable reserves of the oil and gas reservoirs based on the hybrid neural network as claimed in claim 3, wherein the step S3 comprises the following sub-steps:

s3-1, constructing a full-connection neural network model, setting hyper-parameters, and training and testing the full-connection neural network model by using a first class of feature data set to obtain model parameters when the iteration of the full-connection neural network model is finished;

5. The method of claim 4, wherein the model parameters comprise:

6. The method for evaluating the economic recoverable reserves of the oil and gas reservoirs based on the hybrid neural network as claimed in claim 5, wherein the step S3-2 comprises the following sub-steps:

；

；

Input layer neuron values

And output values of hidden layer neurons

；

Hidden layer neuron output values

And output values of output layer neurons

；

Absolute value of hidden layer-output layer connection weight

Calculating the product thereof

；

Wherein, in the step (A),Nthe number of neurons in the input layer;

7. The method for evaluating the economic recoverable reserves of the oil and gas reservoirs based on the hybrid neural network as claimed in claim 6, wherein the step S3-3 comprises the following sub-steps:

、

、

、

Average value of (2)

，

，

，

；

s3-3-3, respectively pairing all characteristic variables

，

，

，

and 3-3-5, calculating the mean value of the four relative importance characteristic values of each secondary characteristic variable and the weight-mean ratio of the mean value in all the secondary characteristic variables, and performing weighted summation on all the secondary characteristic variables to construct and obtain a composite characteristic variable.

8. The method for evaluating the economic recoverable reserves of the oil and gas reservoirs based on the hybrid neural network as claimed in claim 7, wherein the model of the hybrid neural network specifically comprises:

the first feature extraction channel is used for inputting main feature variables and composite feature variables of the first class of feature data sets, and extracting numerical values of geological parameters and technological parameters and feature vectors of classification labels;

9. The method for evaluating the economic recoverable reserves of the oil and gas reservoirs based on the hybrid neural network as claimed in claim 8, wherein the hybrid neural network model adopts an optimized adaptive momentum method for iterative training, wherein an iterative updating formula is as follows:

is the network weight at the next iteration,

is the network weight for the current iteration,

、

in order to adapt the hyper-parameters adaptively,

is a matrix of partial derivatives of the network error against the weights,

is that

The transpose matrix of (a) is,

is the network weight at the last iteration,kis the current iteration number.

10. A hybrid neural network-based reservoir economic recoverable reserve evaluation system employing the method of any one of claims 1 to 9, comprising: