A method for predicting mud loss before drilling in complex well conditions based on BP neural network

Abstract

The present invention discloses a method for predicting mud loss before drilling in complex well conditions based on BP neural network, the method comprising: S1, collecting historical seismic attribute data of the target area, engineering and geological data materials corresponding to the loss cases; S2, standardizing the collected data, and converting all data for the convenience of implementing the method; S3, taking the pre-processed historical seismic data as input, the loss situation as output, and the real loss state as the standard value, supervising training and optimizing to obtain a loss prediction neural network model; S4, inputting the real-time seismic data of the target area, and the model automatically judges the loss situation corresponding to each depth in this area. The present invention solves the problem that it is difficult to accurately predict the well conditions before drilling in existing complex well conditions and mud loss is easy to occur.

Description

Technical Field

[0001] The invention relates to the technical field of drilling construction, and in particular to a method for predicting mud loss before drilling in complex well conditions based on a BP neural network. Background Art

[0002] During the drilling production process, when drilling in fracture development zones and / or karst cave development zones, well leakage often occurs. If drilling is carried out in a formation prone to well leakage, it is a more complicated drilling work. Drilling in the above-mentioned formations is very likely to induce major underground accidents such as blowouts and drill sticking. More importantly, drilling engineering accidents directly affect the discovery and protection of oil and gas layers, resulting in a decline in the effectiveness of exploration and development. For example, the currently commonly used technical means in the Tazhong North Slope area of ​​the Tarim Basin is mainly to test and analyze well leakage through well logging, find the location of the leakage layer, and determine the properties of the leakage layer channel. That is, drilling work is carried out first, and only when well leakage is encountered can it be known that the drilling formation is a drilling leakage layer section. This method cannot predict well leakage in the formation before drilling, and thus cannot provide guidance for drilling and avoid drilling leakage layers. Therefore, the drilling success rate of this method is low.

[0003] Chinese patent application number CN 110941010 A, application publication date March 31, 2020, the name of the invention is a method for predicting drilling loss using seismic data, comprising the following steps: The present invention provides a method for predicting drilling loss using seismic data, comprising the following steps: S1: performing noise reduction filtering processing on the original seismic data body to obtain a filtered data body; S2: performing coherence processing on the filtered data body to obtain a coherence body; S3: fusing the coherence body with the original seismic data body to obtain a fused data body; S4: determining the distribution area of ​​the coherence body according to the profile of the fused data body, and predicting the position of the fracture development zone and / or the karst cave development zone according to the distribution area of ​​the coherence body; S5: predicting the drilling loss layer section according to the position of the fracture development zone and / or the karst cave development zone. The application proposes a technical solution for predicting drilling loss using seismic data, but its shortcomings are: first, the processing and fusion of seismic data in the method described in the application has a large error, which will affect the final prediction of the location of the fracture development zone and / or the karst cave development zone; second, the method can only preliminarily determine the location of the fracture development zone and / or the karst cave development zone, and cannot determine the loss location solely based on the location of the fracture development zone and / or the karst cave development zone. There are many factors that affect the loss, so it has great limitations and uncertainties. Therefore, a method for predicting mud loss before drilling for complex well conditions is needed to improve safety. Summary of the invention

[0004] To this end, the present invention provides a method for predicting mud loss before drilling in complex well conditions based on BP neural network to solve the problem that it is difficult to accurately predict the well conditions before drilling in complex well conditions and mud loss is likely to occur.

[0005] In order to achieve the above object, the present invention provides the following technical solutions:

[0006] The present invention discloses a method for predicting mud loss before drilling in complex well conditions based on a BP neural network. The method comprises:

[0007] S1. Collect historical earthquake attribute data in the target area and engineering and geological data corresponding to missed cases;

[0008] S2. Standardize the collected data and convert all data to facilitate the implementation of the method;

[0009] S3, using the pre-processed historical seismic data as input, the leakage situation as output, and the actual leakage state as the standard value, supervised training and optimization to obtain the leakage prediction neural network model;

[0010] S4. Input the real-time earthquake data of the target area, and the model will automatically judge the leakage situation corresponding to each depth in this area.

[0011] Furthermore, in the S1 step, by collecting, sorting and verifying the geological parameters and engineering parameters of historical drilling losses in the target block, the geological parameters mainly include the geological strata at the time of loss and the lithology of the loss section, and the engineering parameters include the mud loss rate, the cumulative mud loss, the mud density and the construction conditions at the time of mud loss, a variable set of five input parameters is determined: seismic attribute data such as original amplitude, AFE, root mean square amplitude, instantaneous frequency, and response phase. The collected true value of the loss situation of the block is set to 1 for loss and 0 for no loss, as the standard value for model training.

[0012] Furthermore, in the step S2, the method of standardization processing is to adopt the histogram method; determine the relative layer; draw the histogram; obtain the standard value of each data in the study area; and obtain the correction value of the single well curve.

[0013] Furthermore, in the determination of the relative layer, the top surface of the Silurian system in the study area is used as the relative standard layer, and the seismic identification characteristics are obvious; in the drawing of the histogram, the standard layer data of each single well is extracted, and the seismic response frequency distribution histogram is fitted by the Gaussian normal distribution function to obtain the peak readings of each seismic attribute.

[0014] Furthermore, the standard values ​​of various data in the study area are obtained, and the seismic attribute response values ​​of each single well are taken to draw the frequency distribution histogram of the original amplitude, AFE, root mean square amplitude, instantaneous frequency, and response phase of the entire area to obtain the histogram peak value of the entire area, that is, the standard value of each attribute.

[0015] Furthermore, in the obtaining of the single well curve correction value, the correction value = the whole area standard value - the single well peak value.

[0016] Furthermore, in the step S2, the data is converted to normalize the data. Data normalization is to limit the preprocessed data to the range of [-1, 1], so that all features of the data are mapped to the same scale.

[0017] Furthermore, in the step S3, the input parameter InputData of the training set is set, and the five seismic attribute parameters collected and sorted are used as data for neural network training; the output standard value OutputTarget of the training set is set, and the standard value is the leakage situation of the corresponding depth of the block measured; a BP neural network is established, the number of nodes in the input layer, hidden layer, and output layer is set, the transfer function is set, and the batch gradient descent algorithm is used to perform unconstrained nonlinear optimization on the model.

[0018] Furthermore, the method for supervised training and optimization to obtain the missing prediction neural network model is as follows: initializing the original variable parameters, setting the initial weights, thresholds and learning rates, then inputting a given sample to calculate the input and output values ​​of each layer through a sigmoid action function, the neuron state of each layer only affects the neuron state of the next layer, and gradually propagates to the hidden layer, the input of the hidden layer is equal to the weighted sum of the input layer signal, the output layer is equal to the output value of the output of the previous hidden layer after being mapped by an excitation function, and an output signal is generated at the output end. If the ideal output value cannot be obtained in the output layer, the system enters the error signal back propagation process, the error signal propagates from the output layer to the input layer and adjusts the connection weights between the layers and the bias values ​​of the neurons in each layer along the way, and executes the error function gradient descent strategy to continuously reduce the error signal, and the weights are continuously adjusted to minimize the network error function.

[0019] Furthermore, in step S4, based on the gap between the predicted value of the leakage and the measured value of the leakage, the back propagation algorithm will update the weights of the neural network parameters accordingly, and modify the neural network according to the weights. The model prediction value approaches the target value during continuous modification until a certain accuracy rate or number of training times is reached.

[0020] The present invention has the following advantages:

[0021] The present invention discloses a method for predicting mud loss before drilling in complex well conditions based on BP neural network. The predicted loss situation is used as the basis for judging the well location deployment and development, providing a more accurate and effective decision-making basis for drilling technicians and construction personnel, thereby improving the probability of avoiding complex well conditions, increasing the implementation efficiency of drilling operations, avoiding repeated operations, and saving engineering time. The pre-drilling loss situation can be predicted quickly and accurately, thereby providing decision support for drilling, and improving the probability of avoiding loss and reducing drilling time. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the implementation methods of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for the implementation methods or the description of the prior art. Obviously, the drawings in the following description are only exemplary, and for ordinary technicians in this field, other implementation drawings can be derived from the provided drawings without creative work.

[0023] The structures, proportions, sizes, etc. illustrated in this specification are only used to match the contents disclosed in the specification so as to facilitate understanding and reading by persons familiar with the technology. They are not used to limit the conditions under which the present invention can be implemented, and therefore have no substantial technical significance. Any structural modification, change in proportion or adjustment of size shall still fall within the scope of the technical contents disclosed in the present invention without affecting the effects and purposes that can be achieved by the present invention.

[0024] Figure 1 A flow chart of a method for predicting mud loss before drilling in complex well conditions based on a BP neural network provided by an embodiment of the present invention;

[0025] Figure 2 A schematic diagram of a BP network operation structure provided by an embodiment of the present invention;

[0026] Figure 3 BP network algorithm flow chart provided by the embodiment of the present invention;

[0027] Figure 4 The model training optimization curve diagram provided by the embodiment of the present invention;

[0028] Figure 5 A comparison chart of validation set loss prediction results provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0029] The following is a description of the implementation of the present invention by specific embodiments. People familiar with the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0030] Example

[0031] refer to Figure 1 This embodiment discloses a method for predicting mud loss before drilling in complex well conditions based on a BP neural network. The method is:

[0032] S1. Collect historical earthquake attribute data in the target area and engineering and geological data corresponding to missed cases;

[0033] S2. Standardize the collected data and convert all data to facilitate the implementation of the method;

[0034] S3, using the pre-processed historical seismic data as input, the leakage situation as output, and the actual leakage state as the standard value, supervised training and optimization to obtain the leakage prediction neural network model;

[0035] S4. Input the real-time earthquake data of the target area, and the model will automatically judge the leakage situation corresponding to each depth in this area.

[0036] refer to Figure 2 It is a schematic diagram of the BP network operation structure in the technical solution of the present invention, which consists of Figure 2 It can be seen that the typical BP neural network mainly refers to a forward network with three or more layers of network structure. The first and last layers of the structure are generally called input layer and output layer, and the middle layers are called hidden layers or middle layers. There is no feedback between networks, and the network layers are not connected to each other. It mainly depends on the complexity of the system to dynamically adjust the connection relationship and connection strength of the internal nodes of the network, so as to better reflect the structural situation of the system.

[0037] Figure 3It is a BP network algorithm flow chart in the technical solution of the present invention. The learning process is generally as follows: first, the original variable parameters are standardized, converted and initialized in turn, and the initial weights, thresholds and learning rates are set. Then, a given sample is input to calculate the input value and output value of each layer through the sigmoid action function. The state of each layer of neurons only affects the state of the neurons in the next layer and gradually propagates to the hidden layer. The input of the hidden layer is equal to the weighted sum of the input layer signals, and the output layer is equal to the output value of the output of the previous hidden layer after being mapped by the excitation function, and an output signal is generated at the output end. If the ideal output value cannot be obtained in the output layer, the system enters the error signal back propagation process. The error signal propagates from the output layer to the input layer and adjusts the connection weights between each layer and the bias value of each layer of neurons along the way. The error function gradient descent strategy is executed to continuously reduce the error signal, and the weights are continuously adjusted to minimize the network error function. The learning rate of the neural network is set to 0.01, the minimum target error is 0.001, the maximum number of iterations is 1000, and other parameters remain default. The number of hidden layer neurons is set to 12, the input layer is 5, and the output layer is 1. The following formula is used to calculate the number of hidden layer nodes k:

[0038]

[0039] Where m is the number of hidden layer nodes; n is the number of input layer nodes; l is the number of output layer nodes; α is an integer between 1 and 10, so k is 4-13. Different k values ​​are selected to train the model, and the number of hidden layer nodes with the best accuracy is 12.

[0040] In step S1, the geological parameters and engineering parameters of historical drilling losses in the target block are collected, sorted and verified. The geological parameters mainly include the geological stratum at the time of loss and the lithology of the loss section. The engineering parameters include mud loss velocity, cumulative mud loss, mud density and construction conditions at the time of mud loss. The variable set of five input parameters is determined: seismic attribute data such as original amplitude, AFE, root mean square amplitude, instantaneous frequency, and response phase. The actual value of the loss situation in the block collected is set to 1 for loss and 0 for no loss, which is used as the standard value for model training.

[0041] Since the seismic data of each well is measured at different times, with different depths, downhole environments, instrument series, and the operation methods of the later standard scales may not be the same, errors may occur between the seismic attributes of different blocks. Therefore, when applying these seismic attribute data, they must be standardized, and all data will also be converted to facilitate the implementation of the method.

[0042] The method of standardization is to use the histogram method; determine the relative layer; draw the histogram; obtain the standard value of each data in the study area; and obtain the correction value of the single well curve.

[0043] In determining the relative layer, the top surface of the Silurian system in the study area is used as the relative standard layer, and the seismic identification characteristics are obvious; the histogram is drawn, the standard layer data of each single well is extracted, and the seismic response frequency distribution histogram is fitted by the Gaussian normal distribution function to obtain the peak reading of each seismic attribute. The standard values ​​of each data in the study area are obtained, and the seismic attribute response values ​​of each single well are taken to draw the frequency distribution histogram of the original amplitude, AFE, root mean square amplitude, instantaneous frequency, and response phase of the whole area, and the peak value of the histogram of the whole area, that is, the standard value of each attribute, is obtained. In obtaining the correction value of the single well curve, the correction value = the standard value of the whole area - the peak value of the single well.

[0044] The conversion process is data normalization. Data normalization is to limit the preprocessed data to the range of [-1,1], so that all the features of the data are mapped to the same scale. This can avoid the dominance of certain features of the data due to different dimensions.

[0045] In step S3, the input parameter InputData of the training set is set, and the five collected seismic attribute parameters are used as the data for neural network training; the output standard value OutputTarget of the training set is set, and the standard value is the leakage situation of the corresponding depth of the block measured; a BP neural network is established, the number of nodes in the input layer, hidden layer, and output layer is set, the transfer function is set, and the batch gradient descent algorithm is used to perform unconstrained nonlinear optimization on the model.

[0046] The specific process of supervised training and optimization of the pre-drilling mud loss prediction neural network model is as follows: first, the original variable parameters are initialized, and the initial weights, thresholds, and learning rates are set, such as Figure 3 , and then input the given sample to calculate the input and output values ​​of each layer through the sigmoid action function. The state of each layer of neurons only affects the state of the next layer of neurons, and gradually propagates to the hidden layer. The input of the hidden layer is equal to the weighted sum of the input layer signal, and the output layer is equal to the output value of the output of the previous hidden layer after being mapped by the activation function, and the output signal is generated at the output end. If the ideal output value cannot be obtained in the output layer, the system will enter the error signal back propagation process. The error signal propagates from the output layer to the input layer and adjusts the connection weights between each layer and the bias values ​​of each layer of neurons along the way. The error function gradient descent strategy is executed to continuously reduce the error signal, and the weights are continuously adjusted to minimize the network error function.

[0047] To ensure the automatic initialization of weights and thresholds when the model is running, the newff function is used as the generator function of the network model. The newff function means that the first step in training a feedforward network is to create a network object, which is actually the parameter of the newff function. This command creates a network object and initializes the network weights and biases, so the network can be trained.

[0048] The batch gradient descent method with momentum is triggered by the training function traingdm;

[0049] The hyperbolic tangent function tansig is used as the transfer function between the input layer and the hidden layer; logsigz is used as the transfer function between the hidden layer and the output layer;

[0050] The traingdx function is selected as the training function of the "error back propagation" process.

[0051] In step S4, supervised training is performed using data such as the predicted value of the missing situation and the standard value. Based on the gap between the predicted value of the missing situation and the actual measured value of the missing situation, the back propagation algorithm will update the weights of the neural network parameters accordingly, and modify the neural network according to the weights. The model prediction value approaches the target value during continuous modification until a certain accuracy rate or number of training times is reached.

[0052] refer to Figure 4 This is the model training optimization curve. It can be seen from the figure that after 9 iterations, the result is less than the error of 0.001 for 6 consecutive times, which meets the training requirements and successfully establishes the neural network training model.

[0053] refer to Figure 5 This is a comparison chart of the validation set missing prediction results. It can be seen from the figure that the overall prediction compliance rate has reached more than 80%, which proves the effectiveness of using this information to predict missing. At the same time, it shows that the BP neural network is suitable for the judgment of missing in the area. The main reason for the prediction deviation may still be due to the geological conditions in different regions, the multi-solution of seismic attributes or insufficient data representativeness.

[0054] This embodiment discloses a method for predicting mud loss before drilling in complex well conditions based on a BP neural network. The predicted loss situation is used as a basis for judging the well site deployment and development, providing drilling technicians and construction personnel with a more accurate and effective decision-making basis, thereby increasing the probability of avoiding complex well conditions, increasing the implementation efficiency of drilling operations, avoiding repeated operations, and saving engineering time. The pre-drilling loss situation can be predicted quickly and accurately, thereby providing decision support for drilling and increasing the probability of avoiding loss and reducing drilling time.

[0055] Although the present invention has been described in detail above by general description and specific embodiments, it is obvious to those skilled in the art that some modifications or improvements can be made to the present invention. Therefore, these modifications or improvements made without departing from the spirit of the present invention all belong to the scope of protection claimed by the present invention.

Claims

1. A method for predicting mud loss before drilling in complex well conditions based on BP neural network, characterized in that: The method is: S1. Collect historical earthquake attribute data in the target area and engineering and geological data corresponding to missed cases; S2. Standardize the collected data and pre-process all data to facilitate the implementation of the method; S3, using the pre-processed historical seismic data as input, the leakage situation as output, and the actual leakage state as the standard value, supervised training and optimization to obtain the leakage prediction neural network model; S4. Input the real-time earthquake data of the target area, and the model will automatically judge the leakage situation corresponding to each depth in this area.

2. A method for predicting mud loss before drilling in complex well conditions based on BP neural network as claimed in claim 1, characterized in that: In the step S1, by collecting, sorting and verifying the geological parameters and engineering parameters of historical drilling loss in the target area, five input parameter variable sets are determined: original amplitude, AFE, root mean square amplitude, instantaneous frequency, and response phase. The collected true values ​​of the loss situation in the target area are set to 1 for loss and 0 for no loss as standard values ​​for model training; The geological parameters include the geological stratum at the time of leakage and the lithology of the leakage section; the engineering parameters include the mud loss rate, the cumulative mud loss, the mud density and the construction conditions at the time of mud loss.

3. The method for predicting mud loss before drilling in complex well conditions based on BP neural network according to claim 1, characterized in that: In the step S2, the method of standardization processing is to adopt the histogram method; determine the relative layer; draw the histogram; obtain the standard value of each data in the study area; and obtain the correction value of the single well curve.

4. A method for predicting mud loss before drilling in complex well conditions based on BP neural network as claimed in claim 3, characterized in that: In the determination of the relative layer, the top surface of the Silurian system in the study area is used as the relative standard layer, and the seismic identification characteristics are obvious; the histogram is drawn, the standard layer data of each single well is extracted, and the seismic response frequency distribution histogram is fitted by the Gaussian normal distribution function to obtain the peak readings of each seismic attribute.

5. The method for predicting mud loss before drilling in complex well conditions based on BP neural network as claimed in claim 3, characterized in that: The standard values ​​of various data in the study area are obtained, and the seismic attribute response values ​​of each single well are taken to draw the frequency distribution histogram of the original amplitude, AFE, root mean square amplitude, instantaneous frequency, and response phase of the entire area to obtain the histogram peak value of the entire area, that is, the standard value of each attribute.

6. The method for predicting mud loss before drilling in complex well conditions based on BP neural network as claimed in claim 3, characterized in that: In the process of obtaining the correction value of the single well curve, the correction value = the standard value of the whole area - the peak value of the single well.

7. The method for predicting mud loss before drilling in complex well conditions based on BP neural network according to claim 1, characterized in that: In the step S2, the data is transformed to normalize the data. Data normalization is to limit the preprocessed data to the range of [-1, 1], so that all features of the data are mapped to the same scale.

8. The method for predicting mud loss before drilling in complex well conditions based on BP neural network as claimed in claim 2, characterized in that: In the step S3, the input parameter InputData of the training set is set, and the collected and sorted 5 input parameter variable sets are used as the data for neural network training; the output standard value OutputTarget of the training set is set, and the standard value is the leakage situation of the corresponding depth of the measured target area; a BP neural network is established, the number of nodes in the input layer, hidden layer, and output layer is set, the transfer function is set, and the batch gradient descent algorithm is used to perform unconstrained nonlinear optimization on the model.

9. The method for predicting mud loss before drilling in complex well conditions based on BP neural network according to claim 1, characterized in that: The method for supervised training and optimization to obtain a missing prediction neural network model is as follows: initializing the original variable parameters, setting the initial weights, thresholds and learning rates, then inputting a given sample and calculating the input and output values ​​of each layer through a sigmoid action function, wherein the neuron state of each layer only affects the neuron state of the next layer and gradually propagates to the hidden layer, wherein the input of the hidden layer is equal to the weighted sum of the input layer signals, and the output layer is equal to the output value of the output of the previous hidden layer after being mapped by an excitation function, and an output signal is generated at the output end; if the ideal output value cannot be obtained at the output layer, the system enters an error signal back propagation process, wherein the error signal propagates from the output layer to the input layer and adjusts the connection weights between the layers and the bias values ​​of the neurons of each layer along the way, and executes an error function gradient descent strategy to continuously reduce the error signal, and continuously adjusts the weights to minimize the network error function.

10. The method for predicting mud loss before drilling in complex well conditions based on BP neural network according to claim 1, characterized in that: In step S4, based on the gap between the predicted value of the leakage and the measured value of the leakage, the back propagation algorithm will update the weights of the neural network parameters accordingly, and modify the neural network according to the weights. The model prediction value approaches the target value during continuous modification until a certain accuracy rate or number of training times is reached.

Images