Fault brine well repairing method used in well and rock salt mining

Abstract

The invention discloses a fault brine well repairing method used in well and rock salt mining, which comprises the following steps: (1) constructing a directional well at a proper position away from a fault brine well, and ensuring that a vertical well section of the directional well is vertical to the greatest extent in the construction process; (2) after the preset depth is reached, deflecting is conducted at the tail end of the straight well section, and the deflecting construction track is kept in smooth transition till the hole deviation angle is horizontal; and (3) the horizontal displacement and the azimuth angle are controlled, horizontal section construction is conducted, the well body track is kept smooth till the construction well drilling track is accurately close to the target area and communicates with the fault well solution cavity, and a new mining channel is formed. According to the method, the mining channel is re-established near the fault brine well so as to be in butt joint with the fault brine well, and the mining rate of mine well rock salt is increased.

Description

Technical Field

[0001] This invention relates to the field of salt production technology, and in particular to a method for repairing faulty brine wells in well and mineral salt mining. Background Technology

[0002] In rock salt deposits, some halite layers occur in layered or quasi-layered formations within the salt-bearing sections, maintaining a generally normal original stratigraphic sequence with the top and bottom plates, exhibiting normal stratigraphic position and occurrence. The halite layers have gentle occurrences, multiple layers, and alternate with clastic rock layers, forming rhythmic interbedded layers of equal or unequal thickness. This rhythmic layering repeats within the salt-bearing sections, constituting the multi-layered structure of the rock salt deposit.

[0003] In current well salt mining, directional drilling is commonly used. This method significantly increases the volume of the solution cavity, improving brine production and raw brine quality. Because the drilling is interconnected, trenching time is shortened, greatly increasing the recovery rate of developed reserves, and resulting in substantial improvements in both brine quality and quantity. However, it lacks effective methods to address blockages in the brine channels between convection wells caused by clastic rock collapses or other reasons. Summary of the Invention

[0004] Therefore, the present invention provides a method for repairing faulty brine wells in well salt mining, the steps of which include: (1) Construct a directional well at a suitable distance from the faulty brine well, and ensure that the vertical section of the directional well is as vertical as possible during construction. (2) After reaching the predetermined depth, the well is steered at the end of the vertical section. The steerer trajectory is kept smooth until the well inclination angle is horizontal. (3) Control the horizontal displacement and azimuth angle, carry out horizontal section construction, keep the well trajectory smooth, until the construction drilling track is close to the target area and connects with the cavity of the faulty well to form a new mining channel.

[0005] Furthermore, the vertical well section is controlled with a well inclination of ≤3.5°.

[0006] Furthermore, after the directional drilling begins, the drilling trajectory is predicted during the drilling process. The steps are as follows: S1: Obtain the characteristic parameters during the drilling process, with the well inclination angle, azimuth angle and dogleg degree as the target sequences and the characteristic sequences as the comparison sequences, obtain the gray-level correlation between each characteristic parameter and the target sequence, and determine the characteristic parameters that affect each target variable; S2: Decompose the feature parameters of each target variable within a specific sliding window into approximate components and detail components using a stationary wavelet transform. Extract the approximate components as the trend components of the target variables. Align the obtained trend components according to the sampling points within the specific sliding window and use them as input features. S3: Construct a temporal convolutional network prediction system. The prediction system includes 8 residual blocks, one stacked convolutional block layer, and one fully connected layer. Each residual block consists of a causal dilated convolutional layer, weight normalization, ReLU activation function, and spatial dropout. The input features are mapped through causal dilated convolution to obtain the corresponding feature values. In the causal dilated convolutional layer, the temporal sequence of the output needs to be aligned with the temporal sequence of the input, and zeros are padded on the left side of the sequence to ensure that future time steps are not introduced. After causal dilation convolution, weight normalization is performed, decomposing the weight vector into a direction vector and an amplitude vector. The values ​​are optimized separately, and then a nonlinear factor is introduced through the ReLU activation function to reduce the network complexity. Finally, spatial dropout is used to prevent the network from overfitting. If the input and output dimensions of the residual block are the same, the input is directly added to the output of the residual block. If the input and output dimensions are different, a 1×1 convolution is used to project the input to the same dimension as the output before adding them. The stacked convolutional block layer stacks the above multiple residual blocks and aggregates the feature relationships level by level. The output of the last time step is sliced, and the final predicted value is obtained through a fully connected layer.

[0007] Furthermore, in the stationary wavelet transform decomposition, the feature parameters are expanded into detail coefficients and approximation coefficients at different scales through a series of scaling functions and wavelet functions, wherein the expression for the approximation coefficients of the (j+1)th layer is: The expression for the detail coefficients of the (j+1)th layer is: Where i is the order, j is the layer number, k is the filter coefficient index, K is the filter length, and n is the discrete-time sampling point index. j The upsampling factor for the filter coefficients of the j-th layer; The approximate component of the first layer is selected as the trend component of the target variable.

[0008] Furthermore, a Bayesian optimization algorithm is used to optimize the kernel size and inflation factor of the temporal convolutional network prediction system during the prediction process: First, the prediction system is initialized, and candidate hyperparameter values ​​are randomly collected to build an initial surrogate model; the surrogate model is trained with existing data, and a collection function is established using a Gaussian regression model; the point that maximizes the collection function is selected, the function value of the evaluation point is obtained, added to the current evaluation point dataset, and the optimized hyperparameter value is output.

[0009] Furthermore, the temporal convolutional network prediction system is trained under optimized hyperparameter values: A root mean square error loss function is defined, and the historical data used for training is normalized using a min-max method, as shown in the formula: in The data is after normalization. This is the original historical data. It is the maximum value in the original historical data. It is the minimum value in the original historical data; The normalized data is divided into a training set, a validation set, and a test set according to the time window function. The training set and validation set are used for iterative training, and the test set is used to evaluate the performance. The training set is used to obtain predicted values ​​through the temporal convolutional network prediction system. The predicted values ​​and the true values ​​are compared to calculate the loss function and obtain the gradient of the loss with respect to the network structure model parameters. The model weights are updated according to the gradient. After completing the above training, the performance is evaluated using the validation set. The training is repeated until the loss of the validation set no longer decreases. The network structure model is saved for testing. The saved model is input with the test set to obtain its prediction accuracy and generalization.

[0010] Furthermore, the ratio of the training set, validation set, and test set is 70:15:15.

[0011] This invention also discloses a drilling trajectory prediction system during drilling operations, comprising a data acquisition layer, a data transmission and processing layer, a temporal convolutional network prediction system, and an output module. The data acquisition layer collects feature parameters during the drilling process and sends the collected data to the data transmission and processing layer. The data transmission and processing layer includes a database for storing historical data, formation characteristic data, and designed drilling trajectory data, and a data processing module for processing input features or training data. The data processing module transmits the processed data to the temporal convolutional network prediction system for prediction or training, and the predicted output value is transmitted to the output module to display the prediction result. The output module includes at least a prediction result output module for displaying the prediction result and combining it with historical trajectories and designed trajectories, and an equipment status parameter monitoring module for real-time detection and display of the equipment's operating status to ensure the normal operation of the drilling process.

[0012] Furthermore, the data acquisition layer includes at least: a well inclination angle and azimuth angle detection module for real-time detection of the well inclination angle and azimuth angle of the drilling location; a strong magnetic signal source set as a magnetic beacon source in the faulty brine well, and a magnetic signal receiving module provided on the drilling equipment for receiving signals from the magnetic beacon source and calculating relative distance and azimuth; a near-bit vibration sensor for detecting drill bit vibration and evaluating the deviation of test parameters; a near-bit torque sensor for real-time testing of drill pressure; a near-bit mechanical sensor for measuring guiding force; a temperature sensor for real-time measurement of drilling location temperature; a magnetometer for sensing changes in the strength and direction of the magnetic field in the drilling environment, determining whether the received magnetic signal is interfered with by ferromagnetic minerals, and evaluating the reliability of distance and azimuth angle monitoring; and a logging-while-drilling acoustic module for detecting formation rock physical parameters.

[0013] The beneficial effects of this invention are as follows: 1. This invention improves the salt extraction rate of mines by re-establishing a mining channel near the faulty brine well to connect with it.

[0014] 2. This invention selects the features with the strongest correlation to the target variable from a large number of drilling detection parameters. The time series of each selected key parameter is decomposed into sub-band components of different frequencies, filtering out high-frequency noise and short-term fluctuations while preserving the long-term patterns of slow parameter changes. A temporal convolutional network prediction system is then constructed, making it particularly suitable for handling complex industrial scenarios such as well-mineral-salt directional drilling, characterized by high data noise, numerous influencing factors, and non-stationary processes. It achieves accurate prediction of drilling trajectories and intelligent docking of faulty brine wells, providing operators with early warning and analysis for handling emergencies during the drilling process.

[0015] 3. This invention, through a hierarchical Gaussian process mixture model, enables the system to simultaneously capture different scale characteristics of the drilling environment, adapting to different geological conditions; and has significant advantages in improving prediction accuracy and optimizing operational efficiency. Attached Figure Description

[0016] Figure 1 This is a macroscopic schematic diagram of the construction trajectory of the present invention. Detailed Implementation

[0017] The present invention will be further described below with reference to the embodiments.

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0019] All terms used in this application (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein should be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0020] Example 1 like Figure 1 As shown, a method for repairing faulty brine wells in well salt mining includes the following steps: (1) Construct a directional well at a distance of 60m from the faulty brine well. During the construction of the vertical section of the directional well, ensure that it is as vertical as possible and control the well inclination to be ≤3.5°. Take a measurement every 20m in the vertical section and require that the position error be controlled within ±2m of the design trajectory. (2) After reaching the predetermined depth, the well is steered at the end of the vertical section. The steerer trajectory is kept smooth until the well inclination angle is horizontal. (3) Control the horizontal displacement and azimuth angle, carry out horizontal section construction, keep the well trajectory smooth, until the construction drilling track is close to the target area and connects with the cavity of the faulty well to form a new mining channel.

[0021] A drilling trajectory prediction system is constructed during the drilling process. This system includes a data acquisition layer, a data transmission and processing layer, a temporal convolutional network prediction system, and an output module. The data acquisition layer collects characteristic parameters during the drilling process and sends the collected data to the data transmission and processing layer. To provide a more intuitive understanding of the drilling operation status, the data acquisition layer can also be configured to monitor the real-time operating status of equipment and drilling environmental parameters that require technical understanding during the drilling process. These parameters are then transmitted to the output module to provide results for technical personnel. Examples of parameters affecting drill string lifespan include rotational speed, pump pressure and displacement, and equipment operating temperature. However, the data acquisition layer includes at least the following parameters that may affect the prediction of drilling trajectory: the well inclination angle and azimuth angle detection module, used to detect the well inclination angle and azimuth angle of the drilling position in real time; a strong magnetic signal source is set in the faulty brine well as a magnetic beacon source, and the drilling equipment is equipped with a magnetic signal receiving module, used to receive the signal from the magnetic beacon source and calculate the relative distance and azimuth; a near-bit vibration sensor, used to detect the vibration of the drill bit and evaluate the deviation of the test parameters; a near-bit torque sensor, used to test the bit pressure in real time; a near-bit mechanical sensor, used to measure the guiding force; a temperature sensor, used to measure the temperature at the drilling position in real time; a magnetometer, used to sense the changes in the strength and direction of the magnetic field in the drilling environment, determine whether the received magnetic signal is interfered with by ferromagnetic minerals, and evaluate the reliability of distance and azimuth angle monitoring; and a logging-while-drilling acoustic module for detecting formation rock physical parameters and collecting lithological characteristics and formation porosity data. The data transmission and processing layer includes a database for storing historical data, formation characteristic data, and designed drilling trajectory data, and a data processing module for processing input features or training data. Historical data includes the parameter values ​​tested within completed drilling trajectories, as well as the drilling parameter values ​​recorded during directional drilling of faulty brine wells. Formation characteristic data includes formation state parameters at the location of faulty brine wells, such as overburden pressure variations with depth, pore pressure variations, and fracture pressure variations with depth. Designed drilling trajectory data includes the designed vertical well section trajectory, build-up location, and the designed trajectory for the pre-build-up section. For example, the pre-build-up 50-meter trajectory can be designed and constructed according to the designed trajectory to control the approximate direction of drilling. The data processing module determines the characteristic parameters affecting each target variable from the acquired parameters and performs transformation decomposition to obtain input features. It normalizes the historical data used for training and divides it into training, validation, and test sets. The processed data is then transmitted to the temporal convolutional network prediction system for prediction or training. The predicted output value is transmitted to the output module to display the prediction result. The output module includes at least a prediction result output module for displaying prediction results and combining them with historical trajectories and design trajectories, and an equipment status parameter monitoring module for real-time detection and display of equipment operating status to ensure the normal progress of the drilling process.

[0022] After the directional drilling begins, the aforementioned prediction system is used to predict the drilling trajectory during the drilling process. The steps are as follows: S1: Obtain the characteristic parameters during the drilling process, with the well inclination angle, azimuth angle and dogleg degree as the target sequences and the characteristic sequences as the comparison sequences, to obtain the gray-level correlation between each characteristic parameter and the target sequence, and determine the characteristic parameters that affect each target variable.

[0023] Taking well inclination angle as an example: After the prediction process begins, well inclination angles are acquired every 10m to form a well inclination angle target sequence. The feature sequence formed by the feature parameters at corresponding positions Then the grey relational degree between the target sequence and each feature sequence ( The formula for calculating ) is: in, The correlation coefficient between the target sequence and the feature sequence; This represents the absolute difference between the target sequence and the feature sequence at each measurement point; This represents the number of sampling points.

[0024] According to grey relational degree ( The calculation results show that the main characteristic variables that significantly affect the well inclination angle are the well inclination angle and azimuth angle of the drilling location, drilling pressure, steering force, distance relative to the magnetic beacon source, and the formation state parameters from historical data. By extracting features highly correlated with the target variables from a large amount of collected data and performing dimensionality reduction on the feature set, unnecessary information calculations can be reduced, improving the model's generalization ability and robustness.

[0025] S2: Decompose the feature parameters of each target variable within a specific sliding window (e.g., feature parameter values ​​measured within the time range of 50m drilling in the past) into approximate components and detail components using a stationary wavelet transform. Extract the approximate components as the trend components of the target variables. Align the obtained trend components according to the sampling position points within the specific sliding window and use them as input features.

[0026] Due to the harsh drilling environment and significant mechanical vibration, the acquired data is highly noisy, and the sampled values ​​are prone to abrupt changes, generating outliers. This noise and outliers can significantly impact the prediction results. Therefore, it is necessary to perform stationary wavelet transform on specific characteristic parameters of the target variable to remove noise and outliers while preserving the original data's trends and details.

[0027] In the stationary wavelet transform decomposition, the feature parameters are expanded into detail coefficients and approximation coefficients at different scales through a series of scaling functions and wavelet functions. The expression for the approximation coefficients of the (j+1)th layer is as follows: The expression for the detail coefficients of the (j+1)th layer is: Where i is the order, j is the layer number, k is the filter coefficient index, K is the filter length, and n is the discrete-time sampling point index. j The upsampling factor for the filter coefficients of the j-th layer; Since the approximation coefficients of the first layer can preserve the trend of the original data to the greatest extent, and because the number of transformation layers is small, they also retain important low- and mid-frequency information. Therefore, the approximation coefficients of the first layer are selected as the trend components of the target variable. This effectively preserves the overall change of the original parameters and greatly eliminates noise and outliers.

[0028] Specifically, taking the aforementioned key characteristic variables that significantly influence wellbore inclination as examples, the time range taken for drilling 50m past the predicted point is used as a specific sliding window. The characteristic parameters obtained within this specific sliding window are used as transformation decomposition objects. The measured wellbore inclination and azimuth angles, drilling pressure, steering force, distance relative to the magnetic beacon source, and formation state parameters from historical data are expanded into detail coefficients and approximation coefficients at different scales using a series of scaling functions and wavelet functions. To ensure the data length remains constant, zero-insertion is used for upsampling at each transformation decomposition level, resulting in better translational properties of the transformed data. The first-level approximation coefficient components after the transformation of each characteristic parameter are selected as the trend components of the target variable. The obtained trend components are aligned according to their position points within the specific sliding window and used as input features for the prediction system.

[0029] S3: Construct a temporal convolutional network prediction system. The prediction system includes 8 residual blocks, one stacked convolutional block layer, and one fully connected layer. Each residual block consists of a causal dilated convolutional layer, weight normalization, ReLU activation function, and spatial dropout. The input features are mapped through causal dilated convolution to obtain the corresponding feature values. In the causal dilated convolutional layer, including causal convolution and dilated convolution, the temporal sequence of the output needs to be aligned with the temporal sequence of the input, and zeros are padded on the left side of the sequence to ensure that future time steps are not introduced. Therefore, the output of the causal convolution at time t can only predict based on previous information, not future information. Dilated convolution amplifies the receptive field of the convolutional layer by creating holes in the convolutional kernel. Combining causal convolution and dilated convolution can be defined as the following function: in Given the input feature sequence, It is the expansion factor. Let be the kernel function at time t. For convolution calculation, Indicates the size of the filter. This indicates the size of the convolution kernel.

[0030] Stacking multiple layers of dilated convolutions allows temporal convolutional networks to have a larger receptive field, reducing the number of layers required and better capturing long-term dependencies in time series. Specifically, taking a 4-layer causal dilated convolutional network as an example, the dilation factors d of the hidden layers are 1, 2, and 4 respectively, and the output layer has d=8, with a kernel size of 2. Passing the input feature sequence through 4 layers of causal dilated convolutions ensures that the receptive field covers time steps 1 to 16. The input feature sequence yields feature maps. After causal dilation convolution, weight normalization is performed, decomposing the weight vector into a direction vector and magnitude scalar for separate optimization. Then, a ReLU activation function is used to introduce nonlinearity, reducing network complexity. Finally, spatial dropout is used to prevent overfitting, resulting in regularized features. , in For activation function, These are the features after weight normalization.

[0031] If the input and output dimensions of the residual block are the same, then the input is directly added to the output of the residual block. ;in To output information, As a characteristic of regularization, The input is a feature sequence. If the input and output dimensions are inconsistent, a 1×1 convolution is needed to project the input to the same dimension as the output, and then they are added together. The stacked convolutional block layer stacks the above multiple residual blocks and aggregates the feature relationships level by level. The output of the last time step is sliced ​​and the final predicted value is obtained through a fully connected layer.

[0032] Generally, due to the complexity of the drilling environment, manually adjusted network hyperparameters in temporal convolutional network prediction systems often result in low prediction efficiency or significant discrepancies between predicted and actual drilling parameter values. Therefore, as a preferred solution in this embodiment, an adaptive multi-scale Bayesian optimization algorithm is used to optimize hyperparameters such as kernel size and dilation factor of the temporal convolutional network prediction system during the prediction process, obtaining the objective function with minimum error: in The hyperparameter vector matrix, where each hyperparameter represents the dilation factor, kernel size, learning rate, number of filters, and dropout rate, respectively. This represents the feasible region of each hyperparameter; This represents the loss function with hyperparameters as independent variables; N is the number of input samples; It is the mean square error function; For the actual value of the i-th sample, Let be the predicted value for the i-th sample; It is the mean absolute error function; This is a smoothness metric function to prevent drastic fluctuations in predicted values; For regularization functions; , , , These are adaptive weighting coefficients.

[0033] First, initialize the prediction system and construct a hierarchical Gaussian process mixture model. The model structure is as follows: in, express Prediction loss of a temporal convolutional network prediction system under hyperparameter configuration; Let be the number of components in the Gaussian process, and take . =3; For component indexing; It is an adaptive weighting function; Indicates the first A Gaussian process component For the first Kernel functions for each component. Due to changes in formation conditions during drilling, wear and tear on the drill bit over long-term use, and parameter fluctuations caused by drill string vibration, traditional single Gaussian processes are insufficient to capture the multi-scale characteristics of drilling parameters. Therefore, different kernel functions are needed to capture different scales. These include: Formation state scale kernel function: in, The signal variance is calculated based on historical drilling data and hyperparameters in the vicinity. Based on historical drilling data, the average loss function within each geological layer is calculated, along with the inter-layer variance, which reflects the degree of variation in loss function values ​​between different geological layers. For example, a geological stratum consists of three layers: the first layer accounts for 60% with an inter-layer variance of 0.12; the second layer accounts for 30% with an inter-layer variance of 0.35; and the third layer accounts for 10% with an inter-layer variance of 0.28. =0.6×0.12+0.3×0.35+0.1×0.28. This represents the variance of the geological similarity term, which is the increase in covariance between the corresponding loss function values ​​when two hyperparameters are configured for the same geological stratum. This represents a constant, taking the value 1 for the same geological stratum and 0 for different geological strata. This represents the spatial distance between two hyperparameter points; This represents the characteristic scale of geological strata, with a value of 2.5 for historical drilling data.

[0034] Wave scaling kernel function: in, The signal variance is calculated based on historical drilling data and hyperparameters in the vicinity. The stationary segment in the historical data is taken, the trend is fitted, and the residual is calculated. The short-term fluctuation variance is calculated using the residual, which is the signal variance value. This represents the spatial distance between two hyperparameter points; The characteristic scale representing real-time fluctuations is set to 0.18 based on historical drilling data. This represents the number of cycles in the fluctuation.

[0035] Train the proxy model using existing data, and establish the data collection function using a Gaussian regression model: in, For safety probability estimation, that is, the probability that all constraints are satisfied at the same time; The difference between the minimum loss under the optimal temporal convolutional network prediction system configuration and the predicted loss of hyperparameters in the Gaussian process; It is the standard normal cumulative distribution function; It is the standard normal probability density function; denoted as the standard deviation of the hyperparameter prediction during the Gaussian process; It is an enhancement factor obtained based on historical drilling data. It is the product of five independent Gaussian densities, each corresponding to a hyperparameter.

[0036] Select the point that maximizes the acquisition function, obtain the function value of the evaluation point, add it to the current evaluation point dataset, and output the optimized hyperparameter value.

[0037] Specifically, for example, initial hyperparameters were set as follows: dilation factors of 1, 2, 4, and 8; kernel size of 2; learning rates of 1.5, 2, 2.5, and 3; number of filters of 128; and dropout rate of 0.3. After 15 iterations using the above scheme, the system discovered a promising region: The corresponding validation loss is 0.082. When formation characteristics change, the system automatically increases the weight of the formation state scale and updates the hyperparameter prior distribution based on historical drilling data. It is found that increasing the size of the convolution kernel can better capture the slow changes in drilling data, and a higher dropout rate helps prevent overfitting. Therefore, the system updates the parameters to the optimal values. The verification loss was reduced to 0.065. When drilling through formations with high hardness, the system detected a strong data offset due to intense drill string vibration. Therefore, the system automatically increased the weight of the fluctuation scale. After 30 iterations of optimization, the optimal configuration was found to be: Smaller kernel size allows for faster response to data changes, more filters enable the learning of more complex feature interactions, a lower dropout rate maintains model capacity, and the validation loss is 0.071, resulting in excellent performance under conditions of high vibration.

[0038] Training a temporal convolutional network prediction system with optimized hyperparameter values: Define the root mean square error loss function, and perform min-max normalization on the historical data used for training, as shown in the formula: in The data is after normalization. This is the original historical data. It is the maximum value in the original historical data. It is the minimum value in the original historical data; The normalized data is divided into a training set, a validation set, and a test set according to a time window function; the ratio of the training set, validation set, and test set is 70:15:15. The training set and validation set are used for iterative training, and the test set is used to evaluate the performance. The training set is used to obtain predicted values ​​through the temporal convolutional network prediction system. The predicted values ​​are compared with the true values ​​to calculate the loss function and obtain the gradient of the loss relative to the network structure model parameters. The model weights are updated according to the gradient. After completing the above training cycle, the performance is evaluated using the validation set. Training is repeated until the loss on the validation set no longer decreases. The network structure model is then saved for testing. The prediction accuracy and generalization ability are obtained by inputting the test set into the saved model.

[0039] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0040] Furthermore, those skilled in the art will understand that although some technical solutions herein include certain features but not others included in other technical solutions, combinations of features from different technical solutions are meant to fall within the scope of this application and form different technical solutions. For example, all the above technical solutions can be used in any combination. The information disclosed in this background section is intended only to enhance the understanding of the general background of this application and should not be construed as an admission or in any way implying that such information constitutes prior art known to those skilled in the art.

Claims

1. A method for repairing faulty brine wells in well salt mining, characterized by the following steps: include: (1) Construct a directional well at a suitable distance from the faulty brine well, and ensure that the vertical section of the directional well is as vertical as possible during construction. (2) After reaching the predetermined depth, the well is steered at the end of the vertical section. The steerer trajectory is kept smooth until the well inclination angle is horizontal. (3) Control the horizontal displacement and azimuth angle, carry out horizontal section construction, keep the well trajectory smooth, until the construction drilling track is close to the target area and connects with the cavity of the faulty well to form a new mining channel.

2. The method for repairing faulty brine wells in well salt mining according to claim 1, characterized in that, The vertical well section is controlled with a well inclination of ≤3.5°.

3. The method for repairing faulty brine wells in well salt mining according to claim 1, characterized in that, After the directional drilling begins, the drilling trajectory is predicted during the drilling process. The steps are as follows: S1: Obtain the characteristic parameters during the drilling process, with the well inclination angle, azimuth angle and dogleg degree as the target sequences and the characteristic sequences as the comparison sequences, obtain the gray-level correlation between each characteristic parameter and the target sequence, and determine the characteristic parameters that affect each target variable; S2: Decompose the feature parameters of each target variable within a specific sliding window into approximate components and detail components using a stationary wavelet transform. Extract the approximate components as the trend components of the target variables. Align the obtained trend components according to the sampling points within the specific sliding window and use them as input features. S3: Construct a temporal convolutional network prediction system, which includes 8 residual blocks, one stacked convolutional block layer, and one fully connected layer; each residual block consists of a causal dilated convolutional layer, weight normalization, ReLU activation function, and spatial dropout; the input features are mapped through causal dilated convolution to obtain the corresponding feature values; In the causal dilation convolutional layer, the temporal sequence of the output needs to be aligned with that of the input, and zeros are padded on the left side of the sequence to ensure that future time steps are not introduced. After the causal dilation convolution, weight normalization is performed, decomposing the weight vector into a direction vector and magnitude scalar for optimization. Then, a nonlinear factor is introduced through the ReLU activation function to reduce network complexity. Finally, spatial dropout is used to prevent overfitting. If the input and output dimensions of the residual block are the same, the input is directly added to the output of the residual block. If the input and output dimensions are different, a 1×1 convolution is used to project the input to the same dimension as the output before addition. The stacked convolutional block layer stacks the above multiple residual blocks to aggregate feature relationships level by level. The output of the last time step is sliced ​​and the final prediction is obtained through a fully connected layer.

4. A method for repairing faulty brine wells in well salt mining according to claim 3, characterized in that, In the stationary wavelet transform decomposition, the feature parameters are expanded into detail coefficients and approximation coefficients at different scales through a series of scaling functions and wavelet functions. The expression for the approximation coefficients of the (j+1)th layer is as follows: The expression for the detail coefficients of the (j+1)th layer is: Where i is the order, j is the layer number, k is the filter coefficient index, K is the filter length, and n is the discrete-time sampling point index. j The upsampling factor for the filter coefficients of the j-th layer; The approximate component of the first layer is selected as the trend component of the target variable.

5. A method for repairing faulty brine wells in well salt mining according to claim 3, characterized in that, The Bayesian optimization algorithm is used to optimize the kernel size and inflation factor of the temporal convolutional network prediction system during the prediction process: First, the prediction system is initialized, and candidate hyperparameter values ​​are randomly collected to build an initial surrogate model; the surrogate model is trained with existing data, and a collection function is established using a Gaussian regression model; the point that maximizes the collection function is selected, the function value of the evaluation point is obtained, added to the current evaluation point dataset, and the optimized hyperparameter value is output.

6. A method for repairing faulty brine wells in well salt mining according to claim 5, characterized in that, Training a temporal convolutional network prediction system with optimized hyperparameter values: Define the root mean square error loss function, and perform min-max normalization on the historical data used for training, as shown in the formula: in The data is after normalization. This is the original historical data. It is the maximum value in the original historical data. It is the minimum value in the original historical data; The normalized data is divided into a training set, a validation set, and a test set according to the time window function. The training set and validation set are used for iterative training, and the test set is used to evaluate the performance. The training set is used to obtain predicted values ​​through the temporal convolutional network prediction system. The predicted values ​​and the true values ​​are compared to calculate the loss function and obtain the gradient of the loss with respect to the network structure model parameters. The model weights are updated according to the gradient. After completing the above training, the performance is evaluated using the validation set. The training is repeated until the loss of the validation set no longer decreases. The network structure model is saved for testing. The saved model is input with the test set to obtain its prediction accuracy and generalization.

7. A method for repairing faulty brine wells in well salt mining according to claim 1, characterized in that, The ratio of the training set, validation set, and test set is 70:15:

15.

8. A system for predicting drilling trajectory during drilling operations, characterized in that, The system comprises a data acquisition layer, a data transmission and processing layer, a temporal convolutional network prediction system, and an output module. The data acquisition layer collects characteristic parameters during the drilling process and sends the collected data to the data transmission and processing layer. The data transmission and processing layer includes a database for storing historical data, formation characteristic data, and designed drilling trajectory data, and a data processing module for processing input features or training data. The data processing module transmits the processed data to the temporal convolutional network prediction system for prediction or training, and the predicted output value is transmitted to the output module to display the prediction result. The output module includes at least a prediction result output module for displaying the prediction result and combining it with historical and designed trajectories, and an equipment status parameter monitoring module for real-time detection and display of the equipment's operating status to ensure the normal operation of the drilling process.

9. A drilling trajectory prediction system during construction drilling as described in claim 8, characterized in that, The data acquisition layer includes at least: a well inclination angle and azimuth angle detection module for real-time detection of the well inclination angle and azimuth angle of the drilling location; a strong magnetic signal source set as a magnetic beacon source in the faulty brine well, and a magnetic signal receiving module on the drilling equipment for receiving signals from the magnetic beacon source and calculating relative distance and azimuth; a near-bit vibration sensor for detecting drill bit vibration and evaluating the deviation of test parameters; a near-bit torque sensor for real-time testing of drill pressure; a near-bit mechanical sensor for measuring guiding force; a temperature sensor for real-time measurement of drilling location temperature; a magnetometer for sensing changes in the strength and direction of the magnetic field in the drilling environment, determining whether the received magnetic signal is interfered with by ferromagnetic minerals, and evaluating the reliability of distance and azimuth angle monitoring; and a logging-while-drilling acoustic module for detecting formation rock physical parameters.

Images