Deepwater ultra-shallow gas layer horizontal well gas invasion temperature and pressure neural posterior calibration method and system

Abstract

The present application relates to the technical field of well control safety, in particular to a deepwater ultra-shallow gas layer horizontal well gas invasion temperature and pressure neural network posterior calibration method and system. The method selects key mechanism parameters as calibration objects based on a gas invasion temperature and pressure coupling forward model, and decomposes model prediction bias into segmented error characteristic vectors of different physical mechanisms. An error component and mechanism parameter physical prior matrix are constructed through parameter sensitivity disturbance to establish error attribution constraints. A conditional neural spline flow network is used to learn the joint posterior distribution of key mechanism parameters, and a physically credible filtering mechanism is used to screen compliant parameters. Finally, the forward model is back-substituted to complete calibration. The present application effectively solves the problems of existing methods, such as dependence on empirical values, global fitting prone to non-physical compensation, lack of uncertainty evaluation under small samples, and high cost of repeated forward calculation, and can effectively improve model prediction accuracy and working condition generalization ability, providing reliable technical support for deepwater ultra-shallow gas layer horizontal well temperature and pressure prediction and well control risk evaluation.

Description

Technical Field

[0001] This invention relates to the field of well control safety technology, specifically to a method and system for post-test calibration of gas intrusion temperature, pressure, and nerve signals in deep-water ultra-shallow gas layer horizontal wells. Background Technology

[0002] The Lingshui 36-1 gas field in the South China Sea is a representative example of ultra-deepwater and ultra-shallow gas fields with large natural gas reserves. These fields are characterized by deep water, shallow burial, weak reservoir cementation, high porosity and permeability, and the coexistence of natural gas and hydrates, making them of significant development value. However, deepwater shallow gas has long been considered a high-risk geological hazard source in offshore drilling, easily triggering complex accidents such as shallow gas intrusion, well kicks, blowouts, gas escape, wellbore instability, and wellbore integrity failure. In recent years, with the development of deepwater drilling, well control, and wellbore stabilization technologies, the resource attributes of shallow gas have gradually gained attention, and safe and efficient drilling of deepwater and ultra-shallow gas layers has become a crucial technical issue in deepwater oil and gas development.

[0003] Drilling in these types of gas fields simultaneously faces challenges such as shallow gas escape, insufficient formation bearing capacity, narrow safety density windows, low-temperature heat transfer in long seawater sections (in deep-water ultra-shallow gas layer horizontal well drilling operations, this refers to an ultra-long tubing string section spanning the entire seabed layer, extending from the subsea blowout preventer assembly on the surface drilling platform to the riser at the seabed wellhead), the risk of natural gas hydrate formation, poor wellbore stability, and complex prediction of the wellbore temperature and pressure field. Especially during the horizontal well development stage, the complex wellbore trajectory, increased contact length between the horizontal section and the gas layer, and the coupling of shallow gas intrusion, annular gas content changes, reduced bottomhole pressure safety margin, and low-temperature flow assurance risks further increase the difficulty of well control safety and temperature and pressure prediction. Therefore, establishing a gas intrusion temperature and pressure model suitable for deep-water ultra-shallow gas layer horizontal wells, and improving the reliability, prediction accuracy, and versatility of this model under different well conditions, is of great significance for ensuring drilling safety in deep-water ultra-shallow gas layers, identifying well control risks, and subsequent efficient development.

[0004] Gas intrusion temperature-pressure coupled models for horizontal wells in deep-water ultra-shallow gas formations typically need to simultaneously consider factors such as heat transfer in the long seawater section, heat transfer in shallow formations below the mudline, pressure loss in the build-up and near-horizontal sections, shallow gas intrusion, gas release during drilling rock breaking, gas-liquid slip, and changes in drilling fluid temperature, pressure, and physical properties. While these models can perform forward modeling predictions, the calculation results are significantly affected by key mechanistic parameters. These key mechanistic parameters are difficult to obtain directly through field testing. Current technologies often rely on empirical values, literature parameter transfer, manual calculations, or fitting of a single error objective to determine these parameters, which has the following shortcomings: First, key mechanistic parameters rely on empirical values, leading to a high degree of subjectivity in model calibration. Existing methods for predicting temperature and pressure in deepwater wellbores and simulating gas intrusion typically require specifying key parameters such as heat transfer coefficient, gas source release efficiency, gas-liquid slip parameters, and gas source distribution length. For deepwater ultra-shallow gas-bearing horizontal wells, the combined effects of low-temperature heat transfer in the long seawater section, heat transfer in the shallow formation, gas intrusion in the horizontal gas-bearing section, and gas-liquid slip in the near-horizontal section cause the model parameters to exhibit significant well condition correlations, making it difficult for empirical parameters to adapt to the actual conditions of different target wells.

[0005] Second, existing temperature and pressure model calibrations often employ global error fitting, which easily leads to non-physical compensation. The controlling mechanisms differ across different sections of deep-water, ultra-shallow gas-bearing horizontal wells. If only the temperature or pressure error of the entire well section is used as the optimization objective, the algorithm might reduce the overall error by using parameters unrelated to local error sources. For example, gas source parameters might be used to compensate for heat transfer errors in the seawater section, and heat transfer parameters to compensate for pressure errors in the gas-bearing section. While such calibration results may have smaller mathematical errors, the physical meaning of the parameters is distorted, hindering cross-condition generalization prediction.

[0006] Third, existing model optimization methods are ill-suited to small sample data conditions and lack evaluation of parameter uncertainties. Field-measured temperature and pressure data from deep-water, ultra-shallow gas-bearing horizontal wells, as well as data from adjacent wells, are typically limited, and some key state variables are difficult to measure directly. Traditional optimization methods usually only output a set of optimal parameters, failing to provide posterior parameter distributions, confidence intervals, and identifiability evaluations. In situations with insufficient observational data or strong multi-parameter coupling, it is difficult to determine the stability and reliability of calibration results.

[0007] Fourth, existing calibration methods lack an attribution constraint process for error components and mechanistic parameters. Existing calibration methods typically fail to transform the physical prior information obtained from sensitivity analysis into calibration constraints and fail to establish a physical correlation matrix between "error components and mechanistic parameters," leading to a calibration process that easily degenerates into a black-box fitting and makes it difficult to prevent parameter coupling interference and non-physical calibration results.

[0008] Fifth, existing multi-parameter tuning methods require repeated calls to the forward model, resulting in low computational efficiency. Complex gas intrusion temperature-pressure coupled forward modeling requires repeated runs during multi-parameter tuning. After each parameter update, the temperature field, pressure field, and gas content distribution need to be recalculated, leading to high tuning costs and making it difficult to meet the engineering requirements for rapid updates of drilling data. Summary of the Invention

[0009] To address the shortcomings of existing methods and the needs of practical applications, this invention provides a post-hoc calibration method for gas intrusion temperature, pressure, and neural networks in deep-water ultra-shallow gas-bearing horizontal wells, comprising the following steps: A forward model for gas intrusion temperature-pressure coupling in a horizontal well in a deep-water ultra-shallow gas layer was established. A set of key mechanistic parameters to be calibrated was selected, and the boundaries of the physical feasible region were set. An observation dataset was constructed based on the target well's observation data. The observation data was interpolated and matched to the computational grid nodes of the forward model. The forward simulation results were compared with the target well's observation data. Error calculation intervals were divided according to the physical mechanism attributes of the well section, and multiple types of piecewise error components were extracted to construct piecewise error feature vectors. Through parameter sensitivity perturbation trials, the sensitivity of each error component to each key mechanistic parameter was calculated. An error component-mechanistic parameter physical prior matrix was constructed, and constraint rules were set to suppress low-correlation parameters and prioritize strong-correlation parameters. Multiple sets of candidate parameter combinations were sampled and generated within the parameter physical feasible region. The wellbore temperature and pressure response corresponding to each set of parameters was calculated using the forward model. A neural posterior estimation training sample library and a neural posterior calibration model incorporating physical prior constraints are constructed. The fusion features of the piecewise error feature vector and the physical prior matrix are used as conditional inputs to obtain the joint posterior probability distribution of key mechanism parameters. The neural posterior calibration model is trained using the training sample library, and the piecewise error feature vector corresponding to the measured data of the target well is calculated. The conditional posterior distribution of the key mechanism parameters of the target well is obtained using the trained neural posterior calibration model, and multiple sets of candidate calibration parameters are generated. Based on these candidate calibration parameters, a set of reliable parameters conforming to the physical laws of the wellbore is selected. The optimal recommended calibration parameters are then substituted back into the gas intrusion temperature-pressure coupled forward model to calculate the temperature-pressure field and gas content distribution of the wellbore after calibration. By comparing the error indices before and after calibration, the calibration results and well control risk assessment parameters are obtained.

[0010] Optionally, the set of key mechanism parameters includes: Equivalent heat transfer correction factor for the long sea section, equivalent heat transfer correction factor for the shallow strata section below the mudline, comprehensive pressure field correction factor, drilling rock breaking gas release efficiency, equivalent connection length of pressure differential seepage gas, distribution length of drilling rock breaking gas source, and gas-liquid slip correction parameters for the near-horizontal section.

[0011] Optionally, the segmented error feature vector includes: The errors include temperature error in the seawater section, temperature error in the formation section below the mudline, bottom hole pressure error, pressure error in the gas layer section, gas content error, and temperature-pressure curve shape error. Among these, the temperature-pressure curve shape error is obtained by weighting the temperature profile slope error, the minimum temperature point position offset error, and the temperature curve correlation error according to preset weights.

[0012] Optionally, the construction of the error component-mechanism parameter physical prior matrix and the setting of constraint rules for suppressing low correlation parameters and prioritizing strong correlation parameters include the following steps: For each key mechanism parameter to be calibrated, positive and negative proportional perturbations are applied based on the initial values. The values ​​of various error components corresponding to the perturbations are obtained using a forward model. The sensitivity index of each parameter to each type of error component is calculated, and the sensitivity index is normalized by the error dimension to obtain the correlation strength coefficient between the parameter and the error component. A physical prior matrix is ​​constructed with the error component as the row and the key mechanism parameter as the column. A low correlation threshold is set, and matrix elements with correlation strength coefficients below the threshold are set to zero to suppress the compensation effect of irrelevant parameters on local errors. The weights of strongly correlated parameters are increased, including the correction factor for seawater temperature error and equivalent heat transfer in long seawater section, the correction factor for formation temperature error and equivalent heat transfer in shallow formation section, the correction factor for pressure error and comprehensive pressure field, the correction factor for gas intrusion error and gas source parameters, and the correction parameter for curve morphology error and gas-liquid slip in near-horizontal section.

[0013] Optionally, the neural posterior calibration model that integrates physical prior constraints adopts a conditional neural spline flow network structure, satisfying: Using a random variable following a standard multivariate normal distribution as the base distribution, a conditionally invertible spline transformation controlled by neural network parameters maps the base distribution to the posterior distribution of normalized key mechanism parameters. The conditional input includes: the original piecewise error feature vector and the parameter-specific error features obtained by gating the mapping through the physical prior matrix. The two types of features are concatenated as the conditional input. The output includes: the joint posterior probability density function, posterior mean, maximum posterior estimate, confidence interval of each key mechanism parameter, and correlation coefficient between parameters.

[0014] Optionally, the neural posterior calibration model is trained using a training sample library, satisfying the following condition: the total training loss is composed of a weighted sum of negative log-likelihood loss and a physical prior penalty term. The physical prior penalty term includes: The parameter out-of-bounds penalty term is used to penalize parameter outputs that exceed the physical feasible region; the physical monotonicity violation penalty term is used to penalize parameter combinations that violate the physical monotonicity of the parameter-response relationship; the forward model non-convergence penalty term is used to penalize parameter combinations that cause the forward model to fail to converge; the three penalty terms are weighted and summed after setting weight coefficients for each term.

[0015] Optionally, the step of selecting a set of reliable parameters that conform to the physical laws of the wellbore based on multiple sets of candidate calibration parameters includes performing physical reliability filtering and reliability verification on the candidate calibration parameters: Boundary constraint verification: Each candidate parameter is checked to see if it is within the physically feasible region, and parameters that exceed the boundary are eliminated; Monotonicity verification: After applying a small local perturbation to the candidate parameters, they are input into the forward model to verify that the direction of response change conforms to the physical laws of wellbore heat transfer, pressure recursion, gas source release, and gas-liquid slip, and parameters that violate monotonicity are eliminated; Model convergence verification: The candidate parameters are input into the forward model. If the model cannot meet the temperature and pressure convergence thresholds within the set maximum number of iterations, the parameters are eliminated; Identifiability evaluation: Based on the parameter samples that have passed all verifications, the posterior mean, standard deviation, and confidence interval of each parameter are calculated, and the identifiability of the parameters is evaluated based on the width of the confidence interval.

[0016] Optionally, the step of establishing a coupled forward model of gas intrusion temperature and pressure in a horizontal well in a deep-water ultra-shallow gas layer, screening and determining the set of key mechanism parameters to be calibrated, and setting the boundary of the physically feasible domain includes the following steps: Input basic data, including wellbore structure data, drilling parameters, environmental parameters, formation parameters, and gas layer parameters; using the measured well depth as the calculation coordinate, discretize the wellbore into multiple grid nodes, assigning a well section label to each node, which includes at least the seawater section, the shallow formation section below the mudline, the directional drilling section, and the horizontal gas layer section; set corresponding boundary conditions for different well sections: the seawater section uses the seawater temperature profile as the external thermal boundary, the formation section below the mudline uses the geothermal profile as the external thermal boundary, and the horizontal gas layer section activates the gas source calculation label; construct a composite gas intrusion term calculation module to calculate the standard volumetric flow rate of pressure differential seepage gas and drilling rock-breaking release gas, respectively, and convert them to units. The well depth gas mass source term is superimposed to obtain the composite gas intrusion source term; a pressure field calculation module is constructed, which calculates the gas phase volume fraction and gas-liquid mixing density based on the drift flow model, derives the annular pressure gradient, and solves the pressure field by combining the gas-liquid two-phase mass conservation equation; a temperature field calculation module is constructed, which adopts a three-node heat transfer model of drill string fluid, drill string wall, and annular fluid, and calls the corresponding external heat transfer boundary according to the well section mark to solve the wellbore temperature field; a temperature-pressure coupled iterative process is established, which alternately updates the gas source term, two-phase flow parameters, pressure field, and temperature field in each time step until the temperature and pressure simultaneously meet the convergence conditions, and outputs the wellbore temperature field, pressure field, and gas content distribution results.

[0017] Optionally, constructing the neural posterior estimation training sample library includes the following steps: Within the physical feasible domain of the parameters, the prior distribution of each parameter is defined: length-type parameters adopt a log-uniform distribution, and correction factor-type parameters adopt a truncated normal distribution; multiple sets of candidate parameter combinations are generated using the Latin hypercube sampling method, and invalid samples that exceed the physical feasible domain are eliminated; for each set of valid candidate parameters, the forward model is called to calculate the wellbore temperature and pressure response and gas content distribution; the simulation results of each set of parameters are compared with the target observation data, and the corresponding piecewise error feature vector is calculated to form a mechanism parameter-piecewise error feature vector sample pair, and all valid samples constitute the training sample library.

[0018] Secondly, to efficiently execute the gas invasion temperature, pressure, and neural post-calibration method for deep-water ultra-shallow gas layer horizontal wells provided by this invention, this invention also provides a deep-water ultra-shallow gas layer horizontal well gas invasion temperature, pressure, and neural post-calibration system, including: an input device, an output device, a processor, and a memory. The input device, output device, processor, and memory are interconnected. The memory stores program instructions used in the deep-water ultra-shallow gas layer horizontal well gas invasion temperature, pressure, and neural post-calibration method. This deep-water ultra-shallow gas layer horizontal well gas invasion temperature, pressure, and neural post-calibration system of this invention has a compact structure and stable performance, and can stably execute the deep-water ultra-shallow gas layer horizontal well gas invasion temperature, pressure, and neural post-calibration method provided by this invention, further improving the overall applicability and practical application capability of this invention.

[0019] Beneficial Effects: This invention is based on a gas-invasion temperature-pressure coupled forward model. It automatically infers mechanistic parameters from the temperature-pressure response of the target well using a neural posterior estimation method, replacing traditional empirical values ​​and manual calculations. This results in stronger parameter adaptability and clearer physical meaning. Model bias is decomposed into piecewise error feature vectors corresponding to different physical mechanisms. The attribution relationship between errors and parameters is constrained by a physical prior matrix, ensuring that local errors are preferentially calibrated by the corresponding mechanistic parameters, avoiding cross-compensation by irrelevant parameters, and guaranteeing the mechanistic interpretability of the calibration results. The joint posterior distribution and confidence interval of the output parameters using neural posterior estimation can evaluate parameter identifiability and calibration result stability under conditions of limited measured data, overcoming the shortcomings of traditional single-point optimization. A training sample library is generated through offline sampling. The trained neural posterior model can quickly output the posterior distribution of target well parameters without repeatedly calling the forward model for iterative optimization online, improving calibration efficiency and facilitating applications in drilling scenarios. By ensuring the rationality of calibration parameters through multiple physical constraints, the calibrated model can not only be adapted to the target well data, but also extended to well conditions with similar water depths, well types and gas reservoirs. It can provide reliable model support for pre-drilling design, well control risk assessment and drilling parameter optimization of horizontal wells in deep water and ultra-shallow gas reservoirs. Attached Figure Description

[0020] Figure 1 A flowchart of a post-hoc calibration method for gas intrusion temperature, pressure, and neural pathways in a deep-water ultra-shallow gas layer horizontal well, provided as an embodiment of the present invention; Figure 2 A framework diagram of a deep-water ultra-shallow gas layer horizontal well gas invasion temperature-pressure neural post-test calibration system provided for an embodiment of the present invention. Detailed Implementation

[0021] Specific embodiments of the present invention will now be described in detail. It should be noted that the embodiments described herein are for illustrative purposes only and are not intended to limit the invention. In the following description, numerous specific details are set forth in order to provide a thorough understanding of the invention. However, it will be apparent to those skilled in the art that these specific details are not necessary to practice the invention. In other instances, well-known circuits, software, or methods have not been specifically described to avoid obscuring the invention.

[0022] Throughout this specification, references to "an embodiment," "an embodiment," "an example," or "an example" mean that a particular feature, structure, or characteristic described in connection with that embodiment or example is included in at least one embodiment of the invention. Therefore, the phrases "in an embodiment," "in an embodiment," "an example," or "an example" appearing in various places throughout the specification do not necessarily refer to the same embodiment or example. Furthermore, specific features, structures, or characteristics can be combined in one or more embodiments or examples in any suitable combination and / or sub-combination. Moreover, those skilled in the art will understand that the illustrations provided herein are for illustrative purposes and are not necessarily drawn to scale.

[0023] Please see Figure 1 This invention provides a post-test calibration method for gas intrusion temperature, pressure, and nerve energy in horizontal wells in deep-water ultra-shallow gas formations, comprising the following steps: S1. Establish a forward model of gas intrusion temperature-pressure coupling in a horizontal well in a deep-water ultra-shallow gas layer, screen and determine the set of key mechanism parameters to be calibrated, and set the boundary of the physical feasible domain.

[0024] In this embodiment, the basic data of the target well is first acquired, including well structure data, drilling parameters, environmental parameters, formation parameters, and gas reservoir parameters.

[0025] Wellbore structure data includes measurements of well depth Vertical depth , well inclination angle Seawater depth Location of mudline, start and end depths of build-up section, start and end depths of horizontal section, and well diameter. The drill string outer diameter, drill string inner diameter, and annular hydraulic diameter.

[0026] Drilling parameters include drilling fluid displacement Drilling fluid inlet temperature Drilling fluid density, plastic viscosity, dynamic shear force, and mechanical drilling rate Cycle time, wellhead pressure, or annular outlet pressure.

[0027] Environmental parameters include seawater temperature profile, mudline temperature, and original formation temperature.

[0028] Formation parameters, including geothermal gradient, formation pressure gradient, and formation thermal conductivity. Heat transfer coefficient of the wellbore side in the formation section .

[0029] Gas layer parameters include gas layer pressure air layer temperature Porosity Gas saturation Permeability, gas deviation coefficient, gas relative density, and standard state pressure and temperature.

[0030] Furthermore, a wellbore computational grid is established, specifically for measuring well depth. To calculate the coordinates, the wellbore is divided into... With discrete grid nodes, we obtain:

[0031] In the formula, Represents the set of computational grids for the wellbore. , For the first Each node measures the well depth. For vertical depth, The well inclination angle, For spatial step size, This is a well section marker.

[0032] Well section markers should include at least:

[0033] in, Indicates the first Well segment markers for each grid node, Indicates well section, Indicates the seawater section. This indicates the shallow strata below the mudline. Indicates the inclined section. This indicates a horizontal gas layer segment.

[0034] Then, set the external thermal boundary and the gas source activation flag.

[0035] For seawater segment nodes, the external ambient temperature is taken from the seawater temperature profile. The external heat exchange boundary is set as the seawater section heat exchange boundary.

[0036] For nodes in the strata below the mudline, the external stratum temperature is taken as:

[0037] In the formula, Indicates the first External formation temperature at each grid node This refers to the mudline temperature. For geothermal gradient, For the first Vertical depth of each grid node; This represents the depth of the seawater, specifically the vertical depth corresponding to the mudline position.

[0038] For horizontal gas layer nodes, set a gas source activation flag. For non-gas-bearing segment nodes, set .

[0039] Next, we will establish a module for calculating the composite gas intrusion source term.

[0040] For gas-bearing sections, calculate the near-wellbore formation pressure differential seepage gas and drilling rock-breaking gas release separately.

[0041] The standard state volumetric flow rate of the pressure differential seepage gas is expressed using a binomial seepage relationship as follows:

[0042] when At that time, take:

[0043] In the formula, For gas layer pressure, To correspond to the annular pressure of the well section, and These are the binomial coefficients for pressure differential seepage.

[0044] Convert the pressure differential seepage gas standard state volumetric flow rate to a mass source term per unit measurement well depth:

[0045] In the formula, The density of a gas under standard conditions. It represents the equivalent connection length of the pressure differential seepage gas.

[0046] The standard state volumetric flow rate of gas released during drilling rock breaking is:

[0047] In the formula, To improve the efficiency of gas release during drilling and rock breaking. The standard state gas deviation coefficient, This represents the gas deviation coefficient under gas layer conditions.

[0048] Convert the standard state volumetric flow rate of drilling rock-breaking gas to a mass source term per unit measurement well depth:

[0049] In the formula, The length is allocated for drilling rock-breaking gas sources.

[0050] Therefore, the complex gas intrusion source term is:

[0051] In the formula, For the first Gas mass source term at each node unit measurement well depth.

[0052] In this embodiment, a pressure field calculation module is also established, specifically, to calculate the gas phase volume fraction based on the gas-liquid two-phase flow parameters:

[0053] In the formula, This refers to the gas phase volume fraction. For the apparent velocity of the gas phase, For the apparent velocity of the liquid phase, For distribution parameters, This is the drift velocity coefficient. This is the well inclination correction index.

[0054] Therefore, the mixing density is expressed as:

[0055] In the formula, The density of the gas-liquid mixture. For gas phase density, This refers to the drilling fluid density.

[0056] Furthermore, the annular pressure gradient is expressed as:

[0057] In the formula, For annular pressure, The coefficient of friction between the two phases is denoted as . For mixed flow rates, The annular hydraulic diameter, This is the comprehensive correction factor for the pressure field.

[0058] The circumferential air phase satisfies the mass conservation law:

[0059] In the formula, For the circular flow area, The velocity is the gas phase velocity.

[0060] Next, a temperature field calculation module is established. Specifically, a three-node heat transfer model of the fluid inside the drill string, the drill string wall, and the annular fluid is used to calculate the wellbore temperature field. The temperature field calculation module calls different external heat transfer boundaries according to the well section markings.

[0061] For the annular fluid in the seawater section, the following is adopted:

[0062] in:

[0063] In the formula, The annular fluid temperature, This refers to the temperature of the drill string wall. For seawater temperature, The heat transfer coefficient between the outer wall of the drill string and the annular fluid. The heat exchange perimeter of the drill string's outer wall. The outer heat exchange perimeter of the annulus. The overall heat transfer coefficient for the seawater section is... The equivalent heat transfer coefficient of the riser or outer structure. The heat transfer coefficient on the seawater side is... The equivalent heat transfer effect function is related to the cycle time and water depth. For the seawater section heat source item, This refers to the drilling fluid circulation time. Seawater depth; For reference cycle time; For reference seawater depth; and These are the circulation time influence coefficient and the seawater depth influence coefficient, respectively. Both are non-negative preset empirical coefficients that can be determined based on data from similar wells and numerical calculation results. If there is no relevant data or results, the equivalent heat transfer influence function can be taken as 1.

[0064] For annular fluids in shallow strata below the mudline, skew sections, and horizontal gas layers, the following approach is adopted:

[0065] in:

[0066] In the formula, The comprehensive heat transfer coefficient of the strata below the mudline. The equivalent heat transfer coefficient on the well wall side is... The diameter of the wellbore. The equivalent thermal conductivity of the formation, Let be the function of unsteady thermal conduction influence of the formation. The external stratum temperature, This is the thermal effect term caused by air intrusion.

[0067] Drilling fluid density is calculated using a temperature-pressure coupled density relationship:

[0068] In the formula, For local temperature and local pressure Drilling fluid density under the specified conditions; Reference temperature and reference absolute pressure The density of the drilling fluid below; The drilling fluid temperature coefficient; This represents the drilling fluid pressure coefficient.

[0069] Gas density is calculated using the real gas law:

[0070] In the formula, The density of the gas under local temperature and pressure conditions; This refers to the absolute pressure of the gas. The molar mass of the gas; This is the gas deviation coefficient; This is the universal gas constant; This is the absolute temperature of the gas.

[0071] The effective apparent viscosity of the drilling fluid was calculated using the Bingham plastic fluid model:

[0072] In the formula, The effective apparent viscosity of the drilling fluid; The plastic viscosity of the drilling fluid; For drilling fluid dynamic shear force; The annular hydraulic diameter; The apparent velocity of local gas-liquid mixing.

[0073] Finally, a temperature-pressure coupling iterative process is established, which is executed sequentially at each time step: (1) Update the differential pressure seepage gas source term based on the pressure field of the previous iteration step; (2) Update the drilling rock breaking and gas release item according to the mechanical drilling rate and gas layer parameters; (3) Calculate the composite gas intrusion source term; (4) Calculate the gas holdup, mixing density, and two-phase flow parameters; (5) Update drilling fluid density, effective apparent viscosity and gas density according to local temperature and pressure; (6) Calculate the annular pressure field; (7) Calculate the heat transfer coefficient; (8) Solve for the temperature fields of the fluid inside the drill string, the drill string wall, and the annular fluid; (9) Determine the convergence condition.

[0074] In this embodiment, the convergence condition is:

[0075]

[0076] In the formula, For the current iteration step, This is the temperature convergence threshold. This is the pressure convergence threshold.

[0077] Output the forward modeling result after convergence is achieved:

[0078] In the formula, Forward modeling of gas intrusion temperature and pressure coupling, Input data for the target well. For the key mechanism parameters to be calibrated, The model outputs results including temperature field, pressure field, and gas content distribution.

[0079] In another embodiment, based on the important error sources of the gas intrusion temperature and pressure model for deep-water ultra-shallow gas layer horizontal wells, the following parameters are preferred as key mechanism parameters to be calibrated:

[0080] in, It is the equivalent heat transfer coefficient of the long seawater section, used to describe the external heat transfer intensity of the seawater section; It is the equivalent heat transfer coefficient of the shallow strata below the mudline, used to describe the limited heat recharge and formation-side heat transfer capacity of the shallow strata; This is a comprehensive correction factor for the pressure field, used to describe the combined effects of annular frictional pressure loss, well inclination, and changes in drilling fluid temperature, pressure, and physical properties on the pressure field calculation. The drilling rock-breaking gas release efficiency is used to describe the proportion of gas that enters the wellbore after the gas-bearing formation is broken during the drilling process. The equivalent connectivity length of the differential pressure gas flow is used to describe the distribution scale of the formation differential pressure gas flow source along the wellbore. The length allocated to the drilling rock-breaking gas source is used to describe the distribution range of the drilling gas release within the wellbore grid; It is the gas-liquid slip coefficient in the near-horizontal section, used to describe gas slip, retention, and gas content distribution in the inclined section and near-horizontal section.

[0081] Furthermore, the physical feasible domain of the parameters is set, defining a physical feasible range for each parameter:

[0082] in:

[0083] The parameter range is determined by the literature on similar operating conditions, experimental results, data from adjacent wells, or sensitivity analysis results.

[0084] To facilitate neural network training, the parameters to be calibrated also need to be normalized:

[0085] in, , .

[0086] S2. Construct an observation dataset based on the target well observation data, interpolate and match the observation data to the computational grid nodes of the forward model, compare the forward simulation results with the target well observation data, divide the error calculation interval according to the physical mechanism attributes of the well section, and extract multiple types of piecewise error components to construct a piecewise error feature vector.

[0087] In this embodiment, the target observation data may come from measured drilling data, historical data from adjacent wells, digitized data from literature charts, or standard calibration operating condition data, including annular temperature, drill string temperature, annular pressure, bottom hole pressure, gas content, or gas intrusion response data.

[0088] The target observation data is recorded as follows:

[0089] In the formula, For annular temperature observation data, This is data on temperature observation inside the drill string. For annular pressure observation data, For bottom hole pressure observation data, For gas content observation data, This is observational data on the gas intrusion response.

[0090] When some observation data is missing, the missing data item is not included in the calculation of the corresponding error component, or its influence is set to zero through observation weights.

[0091] Furthermore, by unifying spatial coordinates, the observation data is interpolated onto the established computational grid according to the measurement well depth:

[0092] In the formula, For interpolation operators, The original observation data, For the first Each computing node measures the well depth.

[0093] Then set the observation weights, which can be adjusted according to the reliability of different observation data:

[0094] in, As weights for temperature data, Weighting of stress data, As weights for air intrusion response data, Weights are assigned to the curve shape data.

[0095] Next, outlier handling is performed, including identifying abnormal points in the observation data. If the deviation between a certain measuring point and its adjacent measuring points exceeds a set threshold and does not conform to changes in the well structure or physical boundaries, then the point is marked as an outlier and its weight is reduced or it is removed from the error calculation.

[0096] Next, calculate the forward output under the current parameters, and set the initial parameters... Input the forward model of gas intrusion temperature and pressure, and obtain the prediction results:

[0097] In the formula, This includes annular temperature, drill string temperature, annular pressure, bottom hole pressure, gas content distribution, and gas intrusion response results.

[0098] Based on this, the error calculation interval is divided according to the well section markings.

[0099] Define the set of nodes for the seawater segment:

[0100] Define the set of nodes for a stratigraphic segment:

[0101] Define the set of nodes for the gas segment:

[0102] Then, calculate the temperature error of the seawater section:

[0103] In the formula, For the first Calculated annular temperature values ​​for each node. For the first Observed annular temperature values ​​at each node, This is the normalized scale for temperature error.

[0104] Calculate the temperature error of the strata below the mudline:

[0105] Calculate the bottom hole pressure error:

[0106] In the formula, This is the calculated value of the bottom hole pressure. These are bottom hole pressure observation values. This is the normalized scale for pressure error.

[0107] Calculate the pressure error in the gas layer section:

[0108] Calculate the gas content or gas intrusion response error: If gas content observation data exists, then calculate:

[0109] In the formula, For the first Calculated gas content of each node. For the first Gas content observations at each node, This serves as the normalization scale for gas content error.

[0110] If there is no gas cut observation data, but there are indirect indicators such as gas intrusion flow rate, cumulative gas intrusion volume, return outlet flow rate change, wellhead return gas response, or pressure response, then select at least one of these gas intrusion response observations and calculate the gas intrusion response error:

[0111] In the formula, For the first Calculated values ​​of air intrusion response corresponding to each observation time or observation location; For the first Observed values ​​of air intrusion response corresponding to each observation time or observation location; It indicates a pre-selected gas invasion response index, including at least one of the following: gas invasion flow rate, cumulative gas invasion volume, change in return outlet flow rate, wellhead return gas response, or pressure response indirect index; The number of effective gas intrusion response observation data; This is the normalized scale for the gas intrusion response error.

[0112] When both gas content observation data and gas intrusion response observation data exist, the gas intrusion error is expressed as:

[0113] And satisfy:

[0114] In the formula, This is the weighting factor for the gas content error. The weights are the weights for the air intrusion response error, which are preset based on the reliability of the observation data, the number of valid data, or the measurement accuracy.

[0115] When only gas content observation data is available, set as follows:

[0116] and with As an error related to air intrusion.

[0117] When only air intrusion response observation data exists, set as follows:

[0118] and with As an error related to air intrusion.

[0119] When neither gas content observation data nor gas invasion response observation data is available, the observation weight corresponding to the gas invasion error is set to zero, so that this error component does not participate in the comprehensive error calculation, the training loss calculation of the neural posterior calibration model, or the posterior inference of the target well conditions. To maintain the dimension of the piecewise error feature vector unchanged, the gas invasion error component is set to zero, and the corresponding observation data availability flag is set to zero.

[0120] The shape error of the temperature-pressure curve is calculated. The shape error of the temperature-pressure curve includes the slope error of the curve, the position offset of the low temperature point, and the trend correlation error.

[0121] The slope error of the curve can be expressed as:

[0122] The offset of the low temperature point position can be expressed as:

[0123] Trend correlation error refers to the correlation coefficient between the temperature curve and the observed temperature curve.

[0124] Furthermore, the error in the shape of the temperature-pressure curve can be taken as follows:

[0125] In the formula, To calculate the location of the lowest temperature on the temperature curve, To observe the lowest temperature position on the temperature curve, The length of the wellbore. To calculate the correlation coefficient between the temperature curve and the observed temperature curve, These are weighting coefficients, which can be 0.4, 0.3, and 0.3 respectively.

[0126] Finally, the above error components are combined into a piecewise error feature vector:

[0127] S3. Through parameter sensitivity perturbation trial calculation, calculate the sensitivity of various error components to each key mechanism parameter, construct the error component-mechanism parameter physical prior matrix, and set constraint rules for suppressing low correlation parameters and prioritizing strong correlation parameters.

[0128] First, parameter sensitivity perturbation calculations are performed on the parameters to be calibrated. Perform positive and negative perturbations:

[0129] In the formula, For the first The initial values ​​of the parameters, This represents the parameter perturbation ratio. Then, respectively... and Input a forward model and calculate the corresponding error characteristics. and .

[0130] Next, the sensitivity index is calculated:

[0131] In the formula, Indicates the first The mechanistic parameter for the first The influence strength of each error component.

[0132] Furthermore, the normalized sensitivity index:

[0133] In the formula, To prevent tiny constants with a denominator of zero.

[0134] Then, the error component-mechanism parameter physical prior matrix is ​​formed:

[0135] matrix The rows correspond to six types of error features, and the columns correspond to seven types of key mechanism parameters to be calibrated. This matrix is ​​used to represent the strength of the physical correlation between different error feature components and different mechanism parameters.

[0136] In this embodiment, a low correlation parameter suppression rule is also set: when When this happens, the correlation weight between this parameter and the error component is reduced, or it is set to zero:

[0137] In the formula, The threshold for low correlation is [value].

[0138] By using low correlation parameter suppression rules, the compensating effect of irrelevant parameters on local errors can be reduced, preventing calibration results where the overall error is reduced but the physical meaning of the parameters is distorted.

[0139] In another embodiment, a priority rule for strongly correlated parameters is also set: For seawater temperature error ,improve The association weight; For formation temperature error ,improve The association weight; For bottom hole pressure error Pressure error in the gas layer ,improve The association weight; For pressure error in gas layer section and gas content or gas intrusion response error ,improve , and The association weight; For curve shape error ,improve The association weight.

[0140] S4. Sample and generate multiple sets of candidate parameter combinations within the physical feasible domain of the parameters. Calculate the wellbore temperature and pressure response corresponding to each set of parameters through a forward model. Construct a neural posterior estimation training sample library and a neural posterior calibration model that integrates physical prior constraints. Use the fusion features of the piecewise error feature vector and the physical prior matrix as conditional inputs to obtain the joint posterior probability distribution of key mechanism parameters.

[0141] In one embodiment, a prior distribution is defined within the parametric physically feasible region:

[0142] Among them, length parameters with large value ranges adopt a logarithmic uniform distribution, while correction factors that fluctuate around empirical values ​​adopt a truncated normal distribution.

[0143] Then, the Latin hypercube sampling method is used to sample candidate parameter combinations from the prior distribution:

[0144] In the formula, This represents the number of candidate samples.

[0145] Next, a preliminary screening of physical feasibility is conducted for each set of candidate parameters. Check if it meets the following requirements:

[0146] If the conditions are not met, the candidate sample is removed.

[0147] Then, for each set of candidate parameters Run the forward model:

[0148] The corresponding temperature field, pressure field, gas content distribution, and gas intrusion response results are obtained.

[0149] Will With target observation data Compare and calculate the piecewise error eigenvector:

[0150] Finally, a training sample library is formed:

[0151] In the formula, This represents the number of valid samples that have passed the physical feasibility screening and completed the forward modeling calculation.

[0152] In one embodiment, establishing a neural posterior calibration model that incorporates physical prior constraints includes the following steps: Construct a conditional input vector, and combine the piecewise error feature vector. With the physical prior matrix By concatenating or embedding, the conditional input can be obtained:

[0153] In the formula, Construct a constructor for conditional features.

[0154] Physical prior matrix Flattened into a vector, with error characteristics splicing:

[0155] Using the physical prior matrix Gating error characteristics:

[0156] The gating error characteristics corresponding to each mechanism parameter are obtained:

[0157] And As input to the neural posterior calibration model.

[0158] Through the above gating process, the neural posterior calibration model can preferentially utilize error feature components that are physically correlated with the target parameters, thereby reducing the interference of irrelevant errors on the posterior inference of parameters.

[0159] Furthermore, a neural density estimation network is established, and a neural density estimation model that can output probability distributions is used to represent the posterior distribution of the parameters:

[0160] In the formula, For network parameters The approximate posterior distribution is represented by Let be the posterior distribution of the true parameters given the error characteristics and the physical prior matrix.

[0161] In this embodiment, the neural density estimation model employs a conditional neural spline flow model. The piecewise error feature vector is used as the basis for this estimation. Using physical prior gating features as conditional inputs and seven types of normalized mechanism parameters as output variables, the joint posterior distribution of key mechanism parameters is learned through invertible spline transformation.

[0162] Let the basic random variable be:

[0163] In the formula, Let be the basic random variable that follows a standard normal distribution.

[0164] Input under given conditions In this case, the normalized parameters to be calibrated are obtained through conditional invertible spline transformation:

[0165] In the formula, This is the normalized vector of key mechanism parameters. For neural network parameters Controlled conditions for invertible spline transformation functions.

[0166] According to the variable transformation formula, the conditional posterior probability density is expressed as:

[0167] in, For the inverse transformation of the conditionally invertible spline transformation, the determinant term is used to characterize the change in probability density from the space of the basic random variables to the parameter space.

[0168] Finally, the output should be set so that the output of the neural posterior calibration model includes at least the following: (1) Joint posterior distribution of key mechanism parameters:

[0169] (2) Posterior mean:

[0170] (3) Maximum a posteriori estimate:

[0171] (4) Confidence intervals or ranges of confidence for each key mechanism parameter; (5) Correlation or identifiability information between parameters.

[0172] S5. Train the neural posterior calibration model using the training sample library, calculate the piecewise error feature vector corresponding to the measured data of the target well, use the trained neural posterior calibration model to obtain the conditional posterior distribution of the key mechanism parameters of the target well, and sample to generate multiple sets of candidate calibration parameters.

[0173] Input the training sample library into the neural posterior calibration model:

[0174] Then, the negative log-likelihood loss is calculated. The training objective is to maximize the conditional probability of the parameters, which is equivalent to minimizing the negative log-likelihood loss.

[0175] In the formula, It is the negative log-likelihood loss function.

[0176] To prevent the network from outputting non-physical parameter combinations, a physical feasibility penalty is also added:

[0177] in, For parameter out-of-bounds penalty terms, This is a penalty term for violating the physical monotonicity. This is a penalty term for non-convergence in the forward model. These are weighting coefficients, which can be 1.0, 0.5, and 0.1 respectively.

[0178] The total loss function is:

[0179] By incorporating a physical prior penalty term, the neural posterior calibration model can not only learn the parameter distribution but also avoid outputting parameter combinations that violate the laws of wellbore heat transfer, pressure recursion, gas source release, or gas-liquid slip.

[0180] Next, gradient descent or its improved algorithm is used to update the network parameters. Until one of the following conditions is met: (1) The decrease in training loss is less than the set threshold; (2) The negative log-likelihood of the validation set no longer decreases; (3) Reach the maximum number of training rounds; (4) The posterior prediction error meets the set accuracy requirement; (5) The posterior sampling parameters, after being substituted back into the forward model, meet the physical feasibility requirements.

[0181] In this embodiment, several parameter combinations are also sampled from the trained posterior distribution:

[0182] Input it into the forward model for verification. If the overall error after calibration satisfies:

[0183] If the parameter combination satisfies the physical feasible region constraint, monotonicity constraint, and model convergence constraint, then the neural posterior calibration model training is considered effective.

[0184] Based on this, the current observation data of the target well is compared with the calculation results of the initial forward model, and the error of the target well is calculated:

[0185] target well error characteristics and physical prior matrix Input the trained neural posterior calibration model:

[0186] The conditional posterior distribution of the key mechanism parameters of the target well is obtained.

[0187] Then sample from the posterior distribution Group parameters:

[0188] Finally, the optimal recommended calibration parameters are obtained by using the sample with the smallest overall error within the physically feasible region:

[0189] The comprehensive error function is as follows:

[0190] In the formula, These are the weighting coefficients for each error component.

[0191] It should be understood that the posterior distribution of key mechanistic parameters of the target well and the optimal recommended calibration parameters... It is not a simple single optimal fitting parameter, but a calibration result that includes information on posterior distribution, confidence range, and parameter uncertainty.

[0192] S6. Based on multiple sets of candidate calibration parameters, a set of reliable parameters that conform to the physical laws of the wellbore is selected. The optimal recommended calibration parameters are substituted back into the gas intrusion temperature-pressure coupled forward model to calculate the temperature-pressure field and gas content distribution of the wellbore after calibration. The error indexes before and after calibration are compared to obtain the calibration results and well control risk assessment parameters.

[0193] First, perform boundary constraint verification to determine whether the candidate parameters meet the requirements:

[0194] Candidate parameters that do not meet the physical feasible region constraints are eliminated.

[0195] Next, a monotonicity check is performed by inputting the candidate parameters into the forward model after local perturbation, and judging whether the response direction conforms to physical laws.

[0196] (1) Increase Afterwards, the heat exchange intensity of the seawater section should be enhanced; (2) Increase Subsequently, the equivalent heat transfer capacity between the shallow strata below the mudline and the strata should be enhanced; (3) Increase Subsequently, the pressure profile or bottom hole pressure response should reflect the pressure field correction effect and should not show changes that are significantly opposite to the direction of frictional pressure loss or physical property correction. (4) Increase Subsequently, the gas source intensity or gas content of the gas layer should not decrease. (5) Change Subsequently, the equivalent range of the differential pressure seepage gas source along the wellbore direction should change accordingly; (6) Change Afterwards, the distribution range of gas released during drilling and rock breaking should change accordingly; (7) Change Subsequently, the gas content and pressure patterns in the near-horizontal section should change accordingly.

[0197] If a candidate parameter violates physical monotonicity, it is discarded or its credibility is reduced.

[0198] Secondly, the model convergence is verified by inputting the candidate parameters into the forward model. If the model cannot meet the temperature and pressure convergence condition within the set maximum number of iterations (50 times in this embodiment), the candidate parameter is considered infeasible and is eliminated.

[0199] Finally, confidence intervals and identifiability evaluations are performed, and parameter samples that meet the physical confidence filtering conditions are statistically analyzed. The posterior mean, standard deviation, and confidence interval are then calculated.

[0200] If the posterior confidence interval of a parameter is too wide, it indicates that the parameter is not sufficiently identifiable under the current observation data. It is necessary to increase the corresponding observation data, reduce the calibration weight of the parameter, or treat it as a weak constraint parameter.

[0201] If a parameter has a concentrated posterior distribution and can stably reduce the corresponding segment error eigenvector pattern after backsubstitution and forward modeling, it indicates that the parameter has high identifiability under the current target well data conditions.

[0202] Based on this, the calibration parameters are substituted back and the model improvement effect is evaluated. The recommended parameters that have passed the validation are then used. Back to the forward model of temperature and pressure intrusion:

[0203] In the formula, This is the output result of the calibrated forward model.

[0204] Recalculate the calibrated piecewise error eigenvector:

[0205] Recalculate the calibration accuracy improvement rate:

[0206] In the formula, This represents the overall error under the initial parameters. This represents the overall error under the calibration parameters.

[0207] Finally, the calibration results are output, including: (1) Combination of key mechanism parameters after calibration; (2) Posterior mean, maximum posterior estimate and confidence interval of each key mechanism parameter; (3) Evaluation results of parameter identifiability; (4) Temperature error of seawater section, temperature error of formation section, bottom hole pressure error, pressure error of gas section, gas content error and curve shape error before and after calibration; (5) Calibrated annular temperature profile, drill string temperature profile, annular pressure profile, and gas content distribution; (6) The improvement rate of prediction accuracy of the calibrated model; (7) Well control risk assessment results, including bottom hole pressure safety margin, minimum annular temperature, gas invasion risk response, and low temperature risk well section.

[0208] By substituting the key mechanism parameters, which have undergone posterior calibration and physical reliability filtering, back into the forward model, a generalized calibration model is formed that can be used for subsequent temperature and pressure prediction, well control risk assessment, and drilling parameter design in similar deep-water ultra-shallow gas layer horizontal wells.

[0209] It should be noted that the specific implementation methods described above, such as image processing, numerical simulation, and the construction and training of machine learning models, can all be accomplished by the processor by calling the corresponding computer program instructions stored in memory. Those skilled in the art can implement the above functions using algorithms and tools known in the prior art, according to actual needs.

[0210] Please see Figure 2 In this embodiment, to efficiently execute the gas invasion temperature, pressure, and neural post-calibration method for horizontal wells in deep-water ultra-shallow gas formations provided by this invention, the present invention also provides a gas invasion temperature, pressure, and neural post-calibration system for horizontal wells in deep-water ultra-shallow gas formations, comprising: an input device 1, an output device 2, a processor 3, and a memory 4. The input device 1, output device 2, processor 3, and memory 4 are interconnected. The memory 4 stores program instructions used to execute the steps of the gas invasion temperature, pressure, and neural post-calibration method for horizontal wells in deep-water ultra-shallow gas formations. The gas invasion temperature, pressure, and neural post-calibration system for horizontal wells in deep-water ultra-shallow gas formations provided by this invention has a compact structure and stable performance, and can stably execute the gas invasion temperature, pressure, and neural post-calibration method for horizontal wells in deep-water ultra-shallow gas formations provided by this invention, further improving the overall applicability and practical application capability of this invention.

[0211] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention 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. 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 the present invention, and they should all be covered within the scope of the present invention.

Claims

1. A method for post-test calibration of gas intrusion temperature, pressure, and nerve energy in deep-water ultra-shallow gas layer horizontal wells, characterized in that, Includes the following steps: A forward model of gas intrusion temperature-pressure coupling in a horizontal well in a deep-water ultra-shallow gas layer was established, and the set of key mechanism parameters to be calibrated was selected and the boundary of the physical feasible domain was set. An observation dataset is constructed based on the observation data of the target well. The observation data is interpolated and matched to the computational grid nodes of the forward model. The forward simulation results are compared with the observation data of the target well. The error calculation interval is divided according to the physical mechanism properties of the well section. Multiple types of piecewise error components are extracted to construct a piecewise error feature vector. Through parameter sensitivity perturbation trials, the sensitivity of various error components to each key mechanism parameter is calculated, an error component-mechanism parameter physical prior matrix is ​​constructed, and constraint rules for suppressing low-correlation parameters and prioritizing strong-correlation parameters are set. Multiple sets of candidate parameter combinations are generated by sampling within the physical feasible domain of the parameters. The wellbore temperature and pressure response corresponding to each set of parameters is calculated by forward modeling. A neural posterior estimation training sample library and a neural posterior calibration model that integrates physical prior constraints are constructed. The fusion feature of the piecewise error feature vector and the physical prior matrix is ​​used as the condition input to obtain the joint posterior probability distribution of key mechanism parameters. The neural posterior calibration model is trained using a training sample library. The piecewise error feature vector corresponding to the measured data of the target well is calculated. The conditional posterior distribution of the key mechanism parameters of the target well is obtained using the trained neural posterior calibration model, and multiple sets of candidate calibration parameters are sampled and generated. Based on the selection of multiple sets of candidate calibration parameters, a set of reliable parameters that conform to the physical laws of the wellbore is selected. The optimal recommended calibration parameters are then substituted back into the gas intrusion temperature-pressure coupled forward model to calculate the temperature-pressure field and gas content distribution of the wellbore after calibration. The error indexes before and after calibration are compared to obtain the calibration results and well control risk assessment parameters. The construction of the error component-mechanism parameter physical prior matrix, and the setting of constraint rules for suppressing low correlation parameters and prioritizing strong correlation parameters, includes the following steps: For each key mechanism parameter to be calibrated, positive and negative proportional perturbations are applied based on the initial value, and the values ​​of various error components corresponding to the perturbations are obtained using a forward model. Calculate the sensitivity index of each parameter to each type of error component, normalize the sensitivity index according to the error dimension, and obtain the correlation strength coefficient between the parameter and the error component. A physical prior matrix is ​​constructed with error components as rows and key mechanism parameters as columns; Setting a low correlation threshold sets matrix elements with correlation strength coefficients below the threshold to zero, thus suppressing the compensating effect of irrelevant parameters on local errors. Increase the weight of strongly correlated parameter pairs, which include the correction factor for temperature error in the seawater section and equivalent heat transfer in the long seawater section, the correction factor for temperature error in the shallow strata section below the mudline and equivalent heat transfer in the shallow strata section below the mudline, the correction factor for pressure errors and comprehensive pressure field correction factor, the correction factor for gas intrusion errors and gas source parameters, and the correction parameter for gas-liquid slip in the near-horizontal section.

2. The method for post-test calibration of gas intrusion temperature, pressure, and nerve energy in deep-water ultra-shallow gas layer horizontal wells according to claim 1, characterized in that, The set of key mechanism parameters includes: Equivalent heat transfer correction factor for the long sea section, equivalent heat transfer correction factor for the shallow strata section below the mudline, comprehensive pressure field correction factor, drilling rock breaking gas release efficiency, equivalent connection length of pressure differential seepage gas, distribution length of drilling rock breaking gas source, and gas-liquid slip correction parameters for the near-horizontal section.

3. The method for post-test calibration of gas intrusion temperature, pressure, and nerve energy in deep-water ultra-shallow gas layer horizontal wells according to claim 1, characterized in that, The segmented error feature vector includes: Temperature error in the seawater section, temperature error in the shallow formation section below the mudline, bottom hole pressure error, pressure error in the gas layer section, gas content error, and temperature-pressure curve morphology error.

4. The method for post-test calibration of gas intrusion temperature, pressure, and nerve energy in deep-water ultra-shallow gas layer horizontal wells according to claim 1, characterized in that, The neural posterior calibration model that integrates physical prior constraints adopts a conditional neural spline flow network structure and satisfies: Using random variables that follow a standard multivariate normal distribution as the base distribution, the base distribution is mapped to the posterior distribution of normalized key mechanism parameters through conditional invertible spline transformation controlled by neural network parameters. The conditional input includes: the original piecewise error feature vector and the parameter-specific error features obtained by gating mapping through the physical prior matrix. The two types of features are concatenated and used as the conditional input. The output includes: the joint posterior probability density function, posterior mean, maximum posterior estimate, confidence interval of each key mechanism parameter, and correlation coefficient between parameters.

5. The method for post-test calibration of gas intrusion temperature, pressure, and nerve energy in deep-water ultra-shallow gas layer horizontal wells according to claim 1, characterized in that, The neural posterior calibration model is trained using a training sample library, satisfying the following condition: the total training loss consists of a weighted sum of negative log-likelihood loss and physical prior penalty term. The physical prior penalty term includes: The parameter out-of-bounds penalty term is used to penalize parameter outputs that exceed the physical feasible region; Physical monotonicity violation penalty term, used to penalize parameter combinations that violate the physical monotonicity of the response parameter; The non-convergence penalty term for the forward model is used to penalize the combination of parameters that causes the forward model to fail to converge. The three penalty items are weighted and summed after each is assigned a weight coefficient.

6. The method for post-test calibration of gas intrusion temperature, pressure, and nerve energy in deep-water ultra-shallow gas layer horizontal wells according to claim 1, characterized in that, The process of selecting a set of reliable parameters that conform to the physical laws of the wellbore based on multiple sets of candidate calibration parameters includes performing physical reliability filtering and reliability verification on the candidate calibration parameters: Boundary constraint verification: Determine whether each candidate parameter is within the physically feasible region and remove out-of-bounds parameters; Monotonicity verification: After applying a small local perturbation to the candidate parameters, input them into the forward model to verify that the direction of response change conforms to the physical laws of wellbore heat transfer, pressure recursion, gas source release and gas-liquid slip, and eliminate parameters that violate monotonicity; Model convergence verification: Input the candidate parameters into the forward model. If the model cannot meet the temperature and pressure convergence thresholds within the set maximum number of iterations, then the set of parameters is removed. Distinguishingness evaluation: Based on the parameter samples that have passed all validations, the posterior mean, standard deviation and confidence interval of each parameter are calculated, and the distinguishingness of the parameters is evaluated based on the width of the confidence interval.

7. The method for post-test calibration of gas intrusion temperature, pressure, and nerve energy in deep-water ultra-shallow gas layer horizontal wells according to claim 1, characterized in that, The establishment of a coupled forward model for gas intrusion temperature and pressure in horizontal wells in deep-water ultra-shallow gas layers, the screening and determination of the key mechanism parameter set to be calibrated, and the setting of the physical feasible domain boundary include the following steps: Input basic data, including well structure data, drilling parameters, environmental parameters, formation parameters, and gas reservoir parameters; Using the measured well depth as the calculation coordinate, the wellbore is discretized into multiple grid nodes, and each node is assigned a well section label. The well section label includes at least the seawater section, the shallow strata section below the mudline, the directional drilling section, and the horizontal gas layer section. Set corresponding boundary conditions for different well sections: the seawater section uses the seawater temperature profile as the external thermal boundary, the formation section below the mudline uses the geothermal profile as the external thermal boundary, and the horizontal gas layer section activates the gas source calculation flag. A composite gas intrusion source term calculation module is constructed to calculate the standard volumetric flow rate of differential pressure seepage gas and drilling rock-breaking release gas, respectively, convert them into gas mass source terms per unit well depth, and superimpose them to obtain the composite gas intrusion source term; A pressure field calculation module is constructed, which calculates the gas phase volume fraction and gas-liquid mixing density based on the drift flow model, derives the annular pressure gradient, and solves the pressure field by combining the gas-liquid two-phase mass conservation equation. A temperature field calculation module is constructed, which adopts a three-node heat transfer model of fluid inside the drill string, drill string wall and annular fluid, and calls the corresponding external heat transfer boundary according to the well section mark to solve the wellbore temperature field. A temperature-pressure coupled iterative process is established, in which the gas source term, two-phase flow parameters, pressure field and temperature field are alternately updated in each time step until the temperature and pressure simultaneously meet the convergence conditions, and the wellbore temperature field, pressure field and gas content distribution results are output.

8. The method for post-test calibration of gas intrusion temperature, pressure, and nerve energy in deep-water ultra-shallow gas layer horizontal wells according to claim 1, characterized in that, The construction of the neural posterior estimation training sample library includes the following steps: Define the prior distribution of each parameter within the physical feasible region: the length class parameters adopt a log uniform distribution, and the correction factor class parameters adopt a truncated normal distribution; The Latin hypercube sampling method is used to generate multiple sets of candidate parameter combinations, and invalid samples that are outside the physical feasible region are eliminated. For each set of valid candidate parameters, the forward model is called to calculate the wellbore temperature and pressure response and gas content distribution; The simulation results of each set of parameters are compared with the target observation data, and the corresponding piecewise error feature vector is calculated to form a sample pair of mechanism parameters-piecewise error feature vector. All valid samples constitute the training sample library.

9. A post-test calibration system for gas intrusion temperature, pressure, and nerve stimulation in deep-water ultra-shallow gas-bearing horizontal wells, characterized in that... The deep-water ultra-shallow gas layer horizontal well gas invasion temperature-pressure neural post-test calibration system includes: an input device, an output device, a processor, and a memory. The input device, output device, processor, and memory are interconnected. The memory stores program instructions, which are used to execute the deep-water ultra-shallow gas layer horizontal well gas invasion temperature-pressure neural post-test calibration method according to any one of claims 1-8.

Images