CN115059448A - Stratum pressure monitoring method based on deep learning algorithm - Google Patents

Abstract

The invention relates to a formation pressure monitoring method based on a deep learning algorithm. The technical scheme comprises the following steps: step 1, selecting proper adjacent target block well positions according to actual engineering requirements, and collecting well position logging information: the depth, the mechanical drilling speed, the torque and the mud circulation density, and preprocessing the data, wherein the more detailed well position geological logging information is, the higher the selection priority is; step 2, establishing a GA-BP model, and training the GA-BP neural network model by using adjacent well logging information; the invention has the beneficial effects that: compared with the traditional Dc index method, the Eaton method has more accurate prediction result; meanwhile, the prediction effect is good no matter on a small data set or a large data set, and the prediction effect is better along with the perfection of a logging data set; the method can accurately predict the formation pore pressure and the change trend thereof, and is not limited by the well position of the block.

Description

Stratum pressure monitoring method based on deep learning algorithm

Technical Field

The invention relates to a prediction method of formation pore pressure in the field of petroleum drilling, in particular to a formation pressure monitoring method based on a deep learning algorithm.

Background

In the petroleum industry, the formation pore pressure is taken as a geological parameter and plays an important role in oil and gas exploration, drilling engineering and oil and gas development, and the pore pressure is an essential important parameter for realizing rapid, safe and economic drilling in the aspect of the drilling engineering, so that the accurate pre-drilling prediction and the monitoring while drilling of the pore pressure are very important.

The abnormal high pressure is a very important factor influencing the drilling safety, once the density of the slurry used in the abnormal high pressure stratum is unreasonable, the downhole complex problems such as well kick, overflow and the like are easily caused, even blowout is caused, and the drilling safety is seriously influenced. Especially, when deep water and deep drilling are carried out, the safe mud density window is narrow, the underground condition is complex, and the accurate prediction of the formation pore pressure is particularly important.

Therefore, in the drilling design, the drilling fluid density design must be carried out on the basis of the formation pore pressure, but because the pore pressure prediction before drilling is mainly carried out on the seismic interval velocity, and the accuracy of the seismic interval velocity is low, the inaccuracy of the formation pressure prediction is easily caused, and great trouble is brought to actual drilling. Therefore, in the actual drilling process, the drilling parameters and the formation information obtained while drilling are used for monitoring the formation pore pressure in real time, so as to ensure that the drilling safety formation pore pressure has a close relationship with the vertical depth, the pump pressure, the drilling time and the rotating speed. Most of the traditional Dc index method, the Sigma method and the like have regional limitations, and meanwhile, the Dc index method requires high well bottom cleaning degree, otherwise, the monitoring error is very large and is influenced by the density and the pressure difference of the drilling fluid; the Sigma method is influenced by the change of the hole diameter, and the prediction result has obvious deviation, and if a simple linear model is selected, the formation pore pressure cannot be accurately predicted.

Disclosure of Invention

Aiming at the defects in the prior art, the invention provides a stratum pressure monitoring method based on a deep learning algorithm, which adopts a GA-BP neural network to predict the pore pressure of a stratum so as to determine reasonable safe density of drilling fluid, thereby ensuring the safety, high efficiency and economy of a drilling process.

The invention provides a stratum pressure monitoring method based on a deep learning algorithm, which adopts the technical scheme that the method comprises the following steps:

step 1, selecting proper adjacent target block well positions according to actual engineering requirements, and collecting well position logging information: the method comprises the following steps of (1) preprocessing data including vertical depth, mechanical drilling speed, torque and mud circulation density, wherein the more detailed well position geological logging information is, the higher the selection priority is;

step 2, establishing a GA-BP model, and training the GA-BP neural network model by using adjacent well logging information;

the method for establishing the GA-BP model comprises the following steps:

step 2-1, well position logging information collected according to the selected block well positions: the vertical depth, the mechanical drilling speed, the torque and the mud circulation density, and data are preprocessed to eliminate missing values and abnormal values; carrying out Pearson correlation coefficient analysis on factors of the logging data influencing the formation pore pressure to obtain the correlation between each logging element and the formation pore pressure, and selecting optimal parameters as input parameters;

2-2, determining a topological structure of the BP neural network, determining characteristic parameters of the population size, the evolution iteration times, the gene coding length and the population fitness of a genetic algorithm, dividing a data set into a training set and a test set, and normalizing the data set;

step 2-3, encoding initialization population is carried out on the weight and the threshold value of the BP neural network, and prediction error is obtained through calculation;

step 2-4, the adaptive weight is substituted into a genetic algorithm to obtain fitness, and codes meeting the fitness are decoded to output optimal weight and threshold;

and 2-5, updating the weight and the threshold of the BP neural network to obtain a result output by the network, and performing inverse normalization to obtain the output predicted formation pore pressure.

Preferably, in step 2-2, a normalization formula is adopted to perform normalization processing on the formation pore pressure and the logging elements, and the normalization formula is as follows:

wherein, X _scale And X is the raw data, and Xmax and Xmin are respectively the maximum value and the minimum value of the raw data set.

Preferably, in step 2-2, when the GA-BP neural network model is trained and the topology of the BP neural network is determined, the step signal is processed by using an activation function, where the activation function is:

where x is the input to the neuron.

Preferably, in the training of the GA-BP neural network model, binary coding is adopted in the variable coding.

Preferably, roulette selection is used in the selection of population individuals in the training of the GA-BP neural network model.

Preferably, the roulette selection method includes the following steps:

step 1, calculating the fitness f (i =1,2, …, M) of each individual in the population, wherein M is the size of the population;

step 2, calculating the probability of each individual inheritance to the next generation:

step 3, calculating the cumulative probability of each individual:

is chromosome x [ i](i =1,2, …, n);

step 4, generating a pseudo-random number r in the interval of [0,1 ];

step 5, if r < q [1], selecting an individual 1, otherwise selecting an individual k, so that: q [ k-1] < r ≦ q [ k ], and repeat step 4, step 5M times.

Compared with the prior art, the invention has the following beneficial effects:

the method aims at predicting the formation pore pressure of logging data, and establishes a connection between the formation pressure and logging elements so as to establish a formation pore pressure prediction model, so that the formation pore pressure can be accurately predicted, the method is not limited by well positions of blocks, has a good prediction effect under both complex geological conditions and normal compaction conditions, has higher prediction precision compared with a standard BP neural network model, and is more accurate in Eaton method prediction result compared with a traditional Dc index method. Meanwhile, the prediction effect is good no matter on a small data set or a large data set, and the prediction effect is better along with the perfection of a logging data set; the method can accurately predict the formation pore pressure and the change trend thereof, and is not limited by the well position of the block.

Drawings

FIG. 1 is a schematic diagram of a neuron model;

FIG. 2 is a GA-BP neural network training flow diagram;

FIG. 3 is a cloud chart of a matrix of correlation coefficients of logging elements;

FIG. 4 is a graph of the predicted effect of the BP neural network model;

FIG. 5 is a graph of the prediction effect of the GA-BP neural network model.

Detailed Description

The preferred embodiments of the present invention will be described in conjunction with the accompanying drawings, and it will be understood that they are described herein for the purpose of illustration and explanation and not limitation.

In embodiment 1, the invention provides a formation pressure monitoring method based on a deep learning algorithm, which has the technical scheme that the method comprises the following steps:

step 1, selecting proper adjacent target block well positions according to actual engineering requirements, and collecting well position logging information: the depth, the mechanical drilling speed, the torque and the mud circulation density, and preprocessing the data, wherein the more detailed well position geological logging information is, the higher the selection priority is;

step 2, establishing a GA-BP model, and training the GA-BP neural network model by using adjacent well logging information;

the method for establishing the GA-BP model comprises the following steps:

step 2-1, well position logging information collected according to the selected block well positions: the vertical depth, the mechanical drilling speed, the torque and the mud circulation density, and data are preprocessed to eliminate missing values and abnormal values; carrying out Pearson correlation coefficient analysis on factors of the logging data influencing the formation pore pressure to obtain the correlation between each logging element and the formation pore pressure, and selecting optimal parameters as input parameters;

2-2, determining a topological structure of the BP neural network, determining characteristic parameters of the population size, the evolution iteration times, the gene coding length and the population fitness of a genetic algorithm, dividing a data set into a training set and a test set, and normalizing the data set;

step 2-3, encoding initialization population is carried out on the weight and the threshold value of the BP neural network, and prediction error is obtained through calculation;

step 2-4, the adaptive weight is substituted into a genetic algorithm to obtain fitness, and codes meeting the fitness are decoded to output optimal weight and threshold;

and 2-5, updating the weight and the threshold of the BP neural network to obtain a result output by the network, and performing inverse normalization to obtain the output predicted formation pore pressure.

In step 2-2, normalization processing is performed on the formation pore pressure and the logging elements by using a normalization formula, wherein the normalization formula is as follows:

wherein, X _scale And X is the raw data, and Xmax and Xmin are respectively the maximum value and the minimum value of the raw data set.

In addition, in step 2-2, when the GA-BP neural network model is trained and the topology structure of the BP neural network is determined, the step signal is processed by using an activation function, where the activation function is:

where x is the input to the neuron.

Preferably, in the training of the GA-BP neural network model, binary coding is adopted in the variable coding.

In addition, when the GA-BP neural network model is trained, a roulette selection method and a roulette selection method are adopted when population individuals are selected, and the steps are as follows:

step 1, calculating the fitness f (i =1,2, …, M) of each individual in the population, wherein M is the size of the population;

step 2, calculating the probability of each individual inheritance to the next generation:

step 3, calculating the cumulative probability of each individual:

is chromosome x [ i](i =1,2, …, n);

step 4, generating a pseudo-random number r in the interval of [0,1 ];

step 5, if r < q [1], selecting an individual 1, otherwise selecting an individual k, so that: q [ k-1] < r ≦ q [ k ], and repeat step 4, step 5M times.

Embodiment 2, the invention provides a formation pressure monitoring method based on a deep learning algorithm, which has the technical scheme that the method comprises the following steps:

selecting a block well position, and collecting historical well position logging information: vertical depth, pump pressure, drilling time, rotation speed, formation pore pressure and the like.

Step two, data processing:

(1) correlation analysis

By logging nine sets of data: and (3) carrying out Pearson coefficient correlation analysis on the vertical depth, the mechanical drilling speed, the drilling time, the drilling pressure, the rotating speed, the torque, the mud circulation density, the hanging weight and the pumping pressure of the drilling mud and the formation pore pressure, and selecting parameters with strong correlation as input layer parameters.

TABLE 1 logging element and formation pore pressure related coefficient table

And PP is the formation pore pressure.

Referring to fig. 3, in combination with the table of coefficients relating logging elements to formation pore pressure, it can be seen that the correlation between each input feature and the formation pore pressure is ordered as: mud circulation density > vertical depth > drilling time > hanging weight > rotating speed > drilling pressure > pumping pressure > torque > mechanical drilling speed. The formation pore pressure is in positive correlation with mud circulation density, vertical depth, drilling time, drilling pressure, rotating speed and hanging weight, and in negative correlation with drilling rate, pumping pressure and torque, wherein the mud circulation density has the strongest correlation and is 0.948. Finally, combining with professional knowledge, comprehensively considering the data sample acquisition difficulty and the correlation, selecting the vertical depth, the mud circulation density, the mechanical drilling speed and the torque as input parameters, and solving the variable formation pore pressure as an output parameter.

(2) Data feature normalization

Because of the sensitivity of neural networks to data scaling, it is necessary to normalize the logging information and formation pore pressure data. Meanwhile, normalization processing is carried out according to the formula (1) in order to reduce the difference of the magnitude of each data and avoid large error of a network prediction model caused by the difference of input and output quantities.

(1)

X _scale Is normalized data, X is raw data, X _max 、X _min Respectively, the maximum and minimum values of the original data set.

(3) Activating a function

The neural network is an arithmetic mathematical model which imitates the behavior characteristics of the animal neural network and performs distributed parallel information processing. The simulation of neurons of the human brain receives stimulation of input signals and generates output. The neuron model is shown in fig. 1. X ₁ ,X ₂ ,X ₃ Is an input to a neuron, W ₁ ,W ₂ ,W ₃ Is the neuron weight and b is the neuron bias. Neuron output is as in formula (2):

(2)

the activation function is a key parameter endowing the model with nonlinear characteristics in the network model, the selection of the activation function directly influences the convergence and performance of the model, and the tanh activation function (formula (3)) is selected in order to accelerate the convergence speed and reduce the iteration number:

(3)

where x is the input to the neuron.

(4) Inverse normalization

Because the data is uniformly scaled in the [0,1] interval after normalization, in order to obtain the predicted output formation pore pressure, the data needs to be subjected to inverse normalization and the original data size unit characteristics are returned.

The flow chart of the GA-BP neural network training method is shown in figure 2. The BP neural network part is formed by feedforward network propagation of signals and reverse network propagation of errors and comprises an output layer, a hidden layer and an output layer, wherein the hidden layer can be expanded and optimized, and neurons between adjacent layers form a full link. After initializing input parameters, inputting training samples, carrying out feedforward propagation by the network and calculating network output values and output errors, carrying out backward propagation by the network and calculating backward errors, updating weights and thresholds, returning to feedforward operation again when output is not completed and the maximum iteration number is reached, repeating the operation until the samples are output one by one, obtaining data meeting error precision by a trial and error method, and finishing network training.

The genetic algorithm part expresses the problem to be solved as a chromosome to be coded, then selects an initial population, namely an initial sample, randomly generates a group of populations with proper quantity and size, continuously selects the best sample to be added into the initial population, and iterates for multiple times until the preset number of evolution iterations is reached; setting a fitness function to select a proper sample, and carrying out selection operation on the next step in order to directly inherit or cross-pair the optimized individual to the next generation; and then carrying out cross operation for the recombination of the chromosome genes, and finally carrying out mutation operation, namely carrying out genome change on the chromosomes of individual individuals in the whole population to finally obtain a next generation population, wherein the optimal solution is the individual output with the highest fitness, the operation is stopped, the output result is the optimal weight and the threshold value of the BP neural network, and the optimal weight and the threshold value are brought into the BP neural network for retraining, and then the test is carried out to output the prediction result.

Finally, through multiple experiments, the GA-BP neural network structure is determined to be three layers, the number of input layer layers is 1, the number of neurons is 4, the number of hidden layer layers is 1, the number of neurons is 10, the number of output layer layers is 1, and the number of neurons is 1.

Thirdly, predicting the formation pore pressure by using the trained GA-BP neural model

And (3) taking the logging data of the LD10-1-2 well 3105 group as a training set, and predicting the formation pore pressure of the test set samples of the LD10-1-1 well 3123 group by using the trained GA-BP neural network model. The regression evaluation index selects MSE mean square error as formula (4):

（4）

wherein

In order to predict the value of the sample,

the true value of the sample, and m is the number of samples.

From fig. 4, the mean square error of the GA-BP neural network model test set is 0.001, and the mean square error of the standard BP neural network test set is 0.006. The difference between the predicted value and the actual value of the formation pore pressure of the GA-BP neural network model is small, and accurate data can be provided for prediction of the formation pore pressure. Compared with the traditional stratum pressure prediction model, the GA-BP neural network model can consider the action of the neglected influence factors on the stratum pore pressure, establish a more comprehensive nonlinear regression model and have better prediction effect.

The prediction results and errors of the partial BP neural network model and the GA-BP neural network model are shown in Table 2.

TABLE 2 GA-BP model and BP model stratum pore pressure prediction error analysis table

The above description is only a few of the preferred embodiments of the present invention, and any person skilled in the art may modify the above-described embodiments or modify them into equivalent ones. Therefore, the technical solution according to the present invention is subject to corresponding simple modifications or equivalent changes, as far as the scope of the present invention is claimed.

Claims (6)

1. A stratum pressure monitoring method based on a deep learning algorithm is characterized by comprising the following steps:

step 1, selecting proper adjacent target block well positions according to actual engineering requirements, and collecting well position logging information: the depth, the mechanical drilling speed, the torque and the mud circulation density, and preprocessing the data, wherein the more detailed well position geological logging information is, the higher the selection priority is;

step 2, establishing a GA-BP model, and training the GA-BP neural network model by using adjacent well logging information;

the method for establishing the GA-BP model comprises the following steps:

step 2-1, well position logging information collected according to the selected block well positions: the vertical depth, the mechanical drilling speed, the torque and the mud circulation density, and data are preprocessed to eliminate missing values and abnormal values; carrying out Pearson correlation coefficient analysis on factors of the logging data influencing the formation pore pressure to obtain the correlation between each logging element and the formation pore pressure, and selecting optimal parameters as input parameters;

2-2, determining a topological structure of the BP neural network, determining characteristic parameters of the population size, the evolution iteration times, the gene coding length and the population fitness of a genetic algorithm, dividing a data set into a training set and a test set, and normalizing the data set;

step 2-3, encoding initialization population is carried out on the weight and the threshold value of the BP neural network, and prediction error is obtained through calculation;

step 2-4, the fitness is substituted into a genetic algorithm to obtain fitness, and codes meeting the fitness are decoded to output optimal weight and threshold;

and 2-5, updating the weight and the threshold of the BP neural network to obtain a result output by the network, and performing inverse normalization to obtain the output predicted formation pore pressure.

2. The method for monitoring formation pressure based on deep learning algorithm as claimed in claim 1, wherein: in step 2-2, normalization processing is carried out on the formation pore pressure and logging factors by adopting a normalization formula, wherein the normalization formula is as follows:

wherein, X _scale And X is the raw data, and Xmax and Xmin are respectively the maximum value and the minimum value of the raw data set.

3. The method for monitoring formation pressure based on deep learning algorithm as claimed in claim 2, wherein: in step 2-2, when a GA-BP neural network model is trained and the topological structure of the BP neural network is determined, step signals are processed by using an activation function, wherein the activation function is as follows:

where x is the input to the neuron.

4. The method for monitoring formation pressure based on deep learning algorithm as claimed in claim 2, wherein: when the GA-BP neural network model is trained, binary coding is adopted when variables are coded.

5. The method for monitoring formation pressure based on deep learning algorithm as claimed in claim 1, wherein: in the training of GA-BP neural network model, roulette selection method is adopted in the selection of population individuals.

6. The method of claim 5 for monitoring formation pressure based on deep learning algorithm, which is characterized in that: the roulette wheel selection method comprises the following steps:

step 1, calculating the fitness f (i =1,2, …, M) of each individual in the population, wherein M is the size of the population;

step 2, calculating the probability of each individual inheritance to the next generation:

step 3, calculating the cumulative probability of each individual:

is chromosome x [ i](i =1,2, …, n);

step 4, generating a pseudo-random number r in the interval of [0,1 ];

step 5, if r < q [1], selecting an individual 1, otherwise selecting an individual k, so that: q [ k-1] < r ≦ q [ k ], and repeat step 4, step 5M times.

Images