Cement leakage risk prediction method based on half-process well cementation technology

Abstract

The invention provides a cement leakage risk prediction method based on a half-pass well cementation technology, which comprises the following steps: adopting a plurality of groups of sensors to regularly collect underground regular data at the same time, and constructing a physical model based on the regular data to simulate the flowing state of cement paste so as to obtain a pressure field, a pressure field and a temperature field; inputting the sensor data and the physical model into a machine learning model for training to obtain a cement leakage risk prediction value; designing a reinforcement learning model according to the cement leakage risk prediction value and sensor data, designing a composite reward function by taking minimization and operation stability as targets, and generating an optimal adjustment value of well cementation operation parameters; the optimal adjustment values of the well cementation operation parameters are mapped to corresponding physical control equipment, the underground pressure, the slurry flow and the cement slurry density are adjusted regularly, and the underground physical state is monitored regularly through a multi-mode sensor; and receiving the latest environment data and the field equipment operation state, constructing a multi-dimensional joint early warning model, and performing full-time-domain dynamic monitoring on the key risk points of well cementation.

Description

Technical Field

[0001] The present invention belongs to the field of cement leakage risk prediction, and in particular relates to a cement leakage risk prediction method based on a half-process cementing technology. Background Art

[0002] In the field of oil and gas exploration and development, cementing is a crucial step in the lower completion process, directly impacting well integrity, operational safety, and subsequent production stability. In recent years, mid-stroke cementing technology has gradually become a mainstream cementing process for deep, complex, and high-temperature, high-pressure wells, owing to its ability to ensure cementing quality while improving operational efficiency and shortening construction cycles. However, current mid-stroke cementing technology still faces a number of difficult-to-overcome technical challenges in its practical application. First, the unique characteristics of the cementing sleeve, tubing structure, and annular passages during mid-stroke cementing can easily lead to cement backflow and cement leakage, directly impacting cementing quality and even causing downhole safety accidents. Some existing downhole tool designs, such as conventional check sleeves and standard packers, suffer from deficiencies such as insufficient sealing performance, the inability to effectively block cement backflow after cementing, and the inability to achieve multiple repositioning or precise positioning, which in turn impacts subsequent acidizing and sand control operations. In addition, current cementing operations rely heavily on manual experience and static parameter settings, and are unable to dynamically perceive and predict regular changes in the downhole environment. This results in a lack of effective early warning and process optimization for cement leakage risks during operations. Especially in key links such as the opening and closing of cementing sleeves, the seating of packers, and the placement of rubber plugs, complex downhole conditions such as high temperature, high pressure, and high solids content make traditional tools prone to clogging and failure during multiple operations, severely reducing the reliability of cementing operations. At the same time, some tool string designs based on "cementing-sand control integration" have been developed both domestically and internationally. While these solutions can achieve simultaneous cementing and sand control operations to a certain extent, they still fail to completely resolve core technical bottlenecks such as difficulty in multiple positioning, inaccurate rubber plug placement, severe cement leakage after cementing, and the inability to effectively transmit seating pressure.

[0003] Therefore, current technologies still have technical gaps that need to be addressed in key issues such as "regular monitoring and control of cementing process in dynamic environment", "composite function integration of cementing tools", and "systematic risk prediction and optimization of cement backflow and leakage". Summary of the Invention

[0004] The purpose of the present invention is to propose a cement leakage risk prediction method based on semi-process cementing technology, which can effectively solve the problems commonly existing in the prior art, such as cementing sleeve backflow, abnormal seat seal pressure transmission, difficulty in multiple positioning of downhole tools, and unpredictable leakage risk.

[0005] An embodiment of the present invention provides a method for predicting cement leakage risk based on a mid-process cementing technology, the method comprising:

[0006] Several sets of sensors are used to collect downhole periodic data simultaneously and regularly. Based on the periodic data, a physical model is constructed to simulate the flow state of cement slurry to obtain the pressure field, pressure field and temperature field. Among them, the several sets of sensors include pressure sensors, temperature sensors, mud flow sensors and displacement sensors.

[0007] The sensor data and physical model are input into the machine learning model for training to obtain the cement leakage risk prediction value;

[0008] Based on the cement leakage risk prediction value and sensor data, a reinforcement learning model is designed. A composite reward function is designed with minimization and operation smoothness as the goal to generate optimal adjustment values ​​for cementing operation parameters. The optimal adjustment values ​​for cementing operation parameters include adjustment amounts for downhole pressure, slurry flow rate, and cement slurry density.

[0009] Mapping the optimal adjustment values ​​of cementing operation parameters to corresponding physical control devices, regularly adjusting downhole pressure, slurry flow rate, and cement slurry density, and regularly monitoring the physical conditions downhole through multimodal sensors;

[0010] Receive the latest environmental data and on-site equipment operating status, build a multi-dimensional joint early warning model, and dynamically monitor key risk points in cementing over the entire time domain;

[0011] The machine learning model is a physical information constrained deep neural network with dual-branch input, where:

[0012] The first branch inputs periodic data and enters the standard fully connected layer network for feature extraction and gradual dimensionality reduction;

[0013] The second branch inputs the pressure field, velocity field, and temperature field, and processes them through the Gaussian kernel convolution module to extract spatial distribution features; wherein the pressure field, velocity field, and temperature field are physical branch features;

[0014] The features of the first branch and the second branch are spliced ​​in the fusion layer, and the fusion is input into the multi-layer perceptron network to output the cement leakage risk prediction value.

[0015] Furthermore, the physical model is: combining a fluid mechanics model and a stress analysis model in cementing operations to simulate downhole periodic data and simulate the flow state of cement slurry.

[0016] Furthermore, the first branch and the second branch are feature spliced ​​in the fusion layer, and the fusion is input into the multi-layer perceptron network to output the cement leakage risk prediction value, which is specifically:

[0017] The physical perception weight mechanism is introduced in the fusion layer, and a weighted splicing method is adopted. The fusion ratio of physical branch features and periodic data branch features is calculated by the physical perception weight w phys and periodic data weight w data Perform dynamic balance; where w phys and w data The sum of is 1;

[0018] The multilayer perceptron network is a 5-layer MLP with 256, 128, 64, 32 and 1 neurons, respectively. It uses the ReLU activation function, combines the vortex balance constraint and the risk smoothing term, and outputs the cement leakage risk prediction value R t ∈[0,1];

[0019] The training of the machine learning model uses historical operation data and periodic data, combined with the Adam optimizer, with an initial learning rate of 1e-4, and dynamically adjusts the physical perception weight w during the training process. phys , so that the machine learning model can reasonably balance the feature fusion of data branch and physical branch at different cementing stages.

[0020] Furthermore, based on the cement leakage risk prediction value and sensor data, a reinforcement learning model is designed, and a composite reward function is designed with the goal of minimizing and ensuring operation stability to generate optimal adjustment values ​​of cementing operation parameters, specifically:

[0021] Taking the periodic data as a state space;

[0022] The adjustment amount of downhole pressure, slurry flow rate and cement slurry density is designed as the action space. Physical boundary constraints are designed through action projection to limit the action space to the range allowed by the actual process.

[0023] Combining the dual goals of risk minimization and operational stability, a compound reward system is designed. t :

[0024]

[0025] in, Is the physical boundary After that, the risk value predicted by the machine learning model is reused, and γ is the penalty factor. is the recommended value for the current stage; ΔP i is the downhole pressure adjustment, ΔQ i is the slurry flow adjustment amount.

[0026] Furthermore, the reinforcement learning model structure is:

[0027] Actor network: A 3-layer fully connected network using ReLU activation, outputting an action space consisting of three-dimensional continuous actions corresponding to adjustments to downhole pressure, slurry flow rate, and cement slurry density.

[0028] Critic network: LSTM+FC structure, input state space and action space to capture the environmental temporal dependencies of the cementing process, output state-action value, and assist in Actor network strategy optimization;

[0029] Among them, the algorithm flow of the reinforcement learning model is:

[0030] Receive state space and cement leakage risk prediction value;

[0031] The Actor network outputs the action space, which is adjusted by the physical boundaries.

[0032] Apply the adjusted action space to update the current downhole pressure, slurry flow rate, and cement slurry density to generate new operating parameters;

[0033] Apply new operating parameters to on-site operations;

[0034] Obtain a new state space and a new cement leakage risk prediction value, and update the experience replay buffer.

[0035] Furthermore, the optimal adjustment value of the cementing operation parameter is mapped to the corresponding physical control device, specifically:

[0036] Build a feedback system for the cementing operation process, including: execution layer, monitoring layer and feedback layer;

[0037] Among them, the execution layer is used to control the current downhole pressure, slurry flow rate and cement slurry density; the monitoring layer is used to regularly collect current data through sensors; and the feedback layer is used to transmit the collected current data to the edge device for preliminary data denoising and outlier detection, and feed back to the reinforcement learning model for optimization.

[0038] Furthermore, the physical boundary also includes a temperature-sensitive weight mechanism: when the current temperature is greater than the temperature warning value, the current downhole pressure is automatically increased, and the penalty weight is adjusted to suppress pressure mutations in high-temperature environments, preventing the cement slurry from solidifying prematurely or flowing back due to temperature increases.

[0039] Furthermore, the latest environmental data and on-site equipment operating status are received, a multi-dimensional joint early warning model is constructed, and full-time dynamic monitoring of key risk points of cementing is performed, specifically:

[0040] A physical risk field perception feature set is constructed based on the cement leakage risk prediction value and the predicted slurry flow rate to measure the abnormal trend of the current cementing conditions;

[0041] Dynamic warning thresholds are designed based on downhole pressure and temperature thresholds to make the warning judgment process environmentally adaptive;

[0042] Constructing a multi-dimensional joint warning model based on the physical risk field perception feature set and dynamic warning thresholds: low risk, medium risk and high risk;

[0043] If the physical risk field perception feature set crosses the dynamic warning threshold multiple times within K consecutive cycles, and the cement leakage risk prediction value is simultaneously higher than 0.8, it is determined to be a suspected cement leakage event, and the main control system is automatically advised to lower the optimization targets of slurry flow and cement slurry density, or to perform a secondary verification of the packer sealing status.

[0044] Furthermore, a risk trend smoothing suppression term is introduced to reduce false alarms caused by short-term fluctuations.

[0045] The beneficial technical effects of the present invention are at least as follows:

[0046] In order to overcome the above-mentioned technical defects, the present invention proposes a systematic technical solution for the cement leakage risk prediction method based on the semi-process cementing technology, which can effectively solve the problems commonly existing in the prior art, such as cementing sleeve backflow, abnormal pressure transmission of the seat seal, difficulty in multiple positioning of downhole tools, and unpredictable leakage risk. The present invention establishes a set of intelligent closed-loop cementing control system by integrating regular monitoring, intelligent risk prediction and dynamic optimization control in cementing operations: on the one hand, the system can make regular dynamic predictions of cement leakage risks based on regular downhole data and cementing physical models, solving the pain point of the inability to accurately predict cement backflow and leakage risks in the existing cementing process; on the other hand, through the intelligent adaptive optimization and control of cementing parameters, the reliability of key links such as sleeve opening and closing, plug placement, seat seal and backflow prevention is effectively improved, and the leakage risk caused by human factors and uncertainty in working conditions is reduced. Furthermore, the present invention addresses complex downhole environments, such as high solids content and severe corrosion, by designing a multifunctional composite downhole tool assembly. This system integrates a cementing sleeve, a backflow prevention device, and an emergency release switch claw, significantly improving the flexibility and safety of downhole operations and comprehensively resolving issues such as post-cement cement leakage, plug misalignment, seal failure, and tool string coordination difficulties. This invention enables intelligent management of the entire process of mid-process cementing operations, from risk perception and risk prediction to dynamic control and optimization. This significantly improves the success rate of cementing operations, reduces construction costs, and meets the demand for efficient cementing in multiple scenarios, including deep and complex wells. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] The present invention is further described with reference to the accompanying drawings. However, the embodiments in the accompanying drawings do not constitute any limitation to the present invention. A person skilled in the art can obtain other drawings based on the following drawings without creative effort.

[0048] Figure 1 This is a flow chart of the cement leakage risk prediction method based on the semi-process cementing technology of the present invention. DETAILED DESCRIPTION

[0049] The following describes embodiments of the present invention in detail. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended only to explain the present invention and are not to be construed as limiting the present invention.

[0050] In one embodiment, Figure 1 As shown, a cement leakage risk prediction method based on the semi-process cementing technology is provided; the method includes the following steps:

[0051] S1. Use several groups of sensors to collect downhole periodic data at the same time, and build a physical model based on the periodic data to simulate the flow state of cement slurry to obtain pressure field, pressure field and temperature field; among them, the several groups of sensors include pressure sensors, temperature sensors, mud flow sensors and displacement sensors.

[0052] Specifically, multimodal sensors are installed at the cementing site to regularly collect key data from the downhole environment:

[0053] Pressure sensor: monitors pressure changes at the wellhead, annulus, and casing to help determine whether cement is experiencing backflow or blockage risk.

[0054] Temperature sensor: measures temperature changes underground. Temperature is an important factor affecting the rheology and curing time of cement slurry, and has a decisive impact on the operation results.

[0055] Mud flow sensor: Regularly measures the flow state of cement slurry. Flow data can reveal changes in flow resistance and thus indicate the risk of backflow or blockage.

[0056] Displacement sensor: Regularly obtain the specific position of cementing tools (such as sliding sleeves, packers, etc.) to ensure accurate positioning.

[0057] Data collected by all sensors is regularly transmitted to a data processing center via wireless communication, ensuring that every moment of the operation is captured for subsequent analysis and prediction. This data provides regular monitoring information during cementing operations, becoming an essential input for subsequent risk prediction and operation optimization.

[0058] Based on sensor data, combined with fluid dynamics models and stress analysis models used in cementing operations, environmental modeling and simulation are performed to simulate factors such as the flow, pressure distribution, and temperature field of cement slurry in the tubing and annulus. This helps us gain a deeper understanding of phenomena such as cement backflow and blockage that may occur during cementing operations. For example, the following fluid dynamics equation is used to describe cement slurry flow:

[0059]

[0060] Among them, P is the pressure field, v is the velocity field, ρ is the fluid density, μ is the fluid viscosity, and g is the acceleration due to gravity.

[0061] The model describes the flow state and pressure changes of cement slurry in the tubing string and annulus, helping to simulate possible backflow risks and blockage problems during downhole operations.

[0062] Regularly collected sensor data and the results of the physical model are combined to form a dynamic environmental model. This model not only regularly describes changes in the downhole operating environment but also constrains subsequent machine learning models based on physical laws, ensuring that predictions conform to actual physical phenomena. This allows subsequent predictions to rely not solely on statistical analysis of historical data but rather on regular updates of the physical environment and dynamic sensor inputs, providing a high-quality foundation for predictions. The fusion of physical models and sensor data provides a scientific foundation for subsequent steps, such as risk prediction and cementing parameter optimization. In actual operations, the likelihood of cement backflow is closely related to the temperature, pressure, and flow rate fields, information that is regularly predicted through the fusion of physical modeling and sensor data. The fused output includes risk values ​​(such as the probability of cement backflow) and environmental conditions (such as temperature and pressure), serving as input data for subsequent steps.

[0063] S2. Input the sensor data and physical model into the machine learning model for training to obtain the cement leakage risk prediction value.

[0064] Specifically, we construct a dual-branch input physical information constrained deep neural network (PINN-DNN) architecture:

[0065] The first branch directly inputs x t (regular sensor data), enters the standard fully connected layer network (MLP structure) for feature extraction and gradual dimensionality reduction.

[0066] The second branch input P phys 、v phys and T phys, processed by the Gaussian Kernel Convolution module to extract spatial distribution features and emphasize the spatial continuity and local similarity of the physical field. The Gaussian kernel uses a normalized radius of σ = 1.5 for smooth modeling of the local environment field.

[0067] The two branches perform feature splicing in the fusion layer and then input into the subsequent multi-layer perceptron network after fusion.

[0068] Furthermore, a physical perception weight mechanism is introduced in the feature fusion layer, and a weighted splicing method is adopted. The fusion ratio of physical branch features and periodic data branch features is adjusted by the adaptive parameter w phys and w data Perform dynamic balance:

[0069] F fusion =w phys ·F phys +w data ·F data ,w phys +w data =1

[0070] Among them, F phys is the physical field characteristic, F data is the periodic sensor characteristic, w phys During high-risk cementing stages (e.g., large fluctuations in cement slurry velocity), the value will be automatically adjusted upwards to give the physical branch a greater weight and enhance the model's physical field perception capabilities.

[0071] Furthermore, the main body of the network is divided into 5 layers of MLP, which are 256, 128, 64, 32 and 1 neurons respectively, and the ReLU activation function is used to output the cement leakage risk prediction value R t ∈[0,1]. The physical field characteristics are also retained within the model To assist in calculating physical consistency loss.

[0072] Furthermore, a composite loss function with multiple physical field constraints is designed:

[0073]

[0074] is the mean square error between the crosstalk risk prediction value and the historical true risk label;

[0075] To calculate the difference between the predicted pressure field and the physical model pressure field;

[0076] is the loss of physical consistency of velocity field;

[0077] It is a vortex balance constraint term that ensures that the velocity field predicted by the model has the physical rationality of the fluid rotation in the annulus and suppresses the numerical noise in the physical field;

[0078] It is a risk smoothing item that suppresses the “high-low risk mutation” phenomenon in the cementing environment and avoids misleading the subsequent cementing control system.

[0079] Model training uses historical job data (such as known and corresponding x t , physical field) is combined with the data collected regularly in step 1, and the Adam optimizer is used with an initial learning rate of 1e-4. The physical perception weight w is dynamically adjusted during the training process. phys , so that the model can reasonably balance the feature fusion of data branch and physical branch at different cementing stages.

[0080] It is understandable that after training is completed, it is deployed to the cementing control system and the model reads x regularly. t With P phys ,v phys ,T phys , and output R t , when R t >0.7 and If the physical tolerance is exceeded, a "high risk + highly abnormal flow field" signal will be automatically sent to the subsequent cementing optimization system, prompting adjustments to the operating pressure, flow rate, etc. to ensure the safety of the cementing operation.

[0081] S3. Based on the cement leakage risk prediction value and sensor data, a reinforcement learning model is designed, and a composite reward function is designed with the goal of minimizing and ensuring operation smoothness to generate optimal adjustment values ​​for cementing operation parameters; wherein the optimal adjustment values ​​for cementing operation parameters include adjustment amounts for downhole pressure, slurry flow rate, and cement slurry density.

[0082] Specifically, receive the cement leakage risk prediction value R output from the previous link (step 2) t , and regular environmental data of cementing operations S t =[P t ,T t ,Q t ], which constitutes the state input of the reinforcement learning optimization model. t is the current downhole pressure, T t is the current downhole temperature, Q t is the current slurry flow rate, R t is the cement leakage risk value predicted by the machine learning model in step 2. Construct a reinforcement learning optimizer based on deep deterministic policy gradient (DDPG) to generate three-dimensional continuous action A t =ΔP i ,ΔQ i ,Δρi ], corresponding to the water injection pressure, water injection flow rate and cement slurry density adjustment to be optimized during the cementing process.

[0083] Furthermore, the reinforcement learning framework is designed:

[0084] State space S t :By [P t ,T t ,Q t ,R t ], including the dynamic environment information of the current working conditions and the cementing risk prediction value;

[0085] Action space A t : Output continuous ΔP i ,ΔQ i ,Δρ i , used to directly adjust the current job parameters;

[0086] Physical boundary limitation: Through action projection, A t Limited to the range allowed by actual process:

[0087]

[0088] Among them A min ,A max Taken from cementing site parameter specifications, such as ΔP i ∈[-1MPa,1MPa].

[0089] Furthermore, the reward function is designed:

[0090] Combining the dual goals of risk minimization and operational stability, we design compound rewards:

[0091]

[0092] in It is adopted After that, re-use the risk value predicted by the model in step 2, γ is the penalty factor, is the recommended value for the current stage. This reward mechanism aims to:

[0093] Strengthen and reduce R t strategy; suppress pressure fluctuations or flow instability caused by large-scale parameter adjustments to the cementing system.

[0094] Furthermore, the network structure:

[0095] Actor network: 3-layer fully connected network, using ReLU activation, output action A t , the dimension is 3 (corresponding to ΔP i ,ΔQ i ,Δρi );

[0096] Critic network: LSTM+FC structure, input [S t ,A t ], to capture the environmental temporal dependency during cementing and output the state-action value Q(S t ,A t ), assisting in Actor strategy optimization.

[0097] Furthermore, the physical field dynamic weighting mechanism:

[0098] The system has a built-in temperature sensitive weight mechanism. When T t >T crit (underground high temperature area), the system automatically increases P i Adjusted penalty weight γ T , suppressing sudden pressure changes in high temperature environments and preventing the risk of premature solidification or backflow of cement slurry due to temperature rise:

[0099] γ T =γ0+η·max(0,T t -T crit )

[0100] Among them, η is the sensitivity factor, T crit It is the temperature warning value (such as 90℃).

[0101] Algorithm flow:

[0102] System receives S t With R t ;

[0103] Actor Output A t , clipped by the physical boundary

[0104] application Update current [P i ,Q i ,ρ i ], the updated result is:

[0105]

[0106] Apply new parameters to field operations;

[0107] Get and the new state S t+1 , and updates the experience replay buffer to complete a complete "perception-optimization-execution-feedback" closed loop.

[0108] Further, training details:

[0109] Using cementing history data [S t ,A t ,r t ,S t+1 ]Pre-trained optimizer;

[0110] Adopt DDPG+ experience replay mechanism, use Adam optimizer (learning rate 1e -4 );

[0111] During training, a delayed update mechanism is enabled (e.g., after each update of the Critic, the Actor is updated every other step to prevent the strategy from converging prematurely);

[0112] Use soft target network update method to improve stability during training.

[0113] S4. Map the optimal adjustment values ​​of cementing operation parameters to the corresponding physical control equipment, regularly adjust the downhole pressure, slurry flow rate and cement slurry density, and regularly monitor the physical conditions of the downhole through multimodal sensors.

[0114] Specifically, we built an execution and feedback system for the cementing operation process (GZ-Exec-Feedback). The system consists of three parts:

[0115] Execution layer: optimize the results Mapped to corresponding physical control devices;

[0116] Monitoring layer: Regularly monitor the physical status of the well through the deployment of multi-modal sensors underground and on the ground t+1 =P t+1 ,T t+1 ,Q t+1 ];

[0117] Feedback layer: The monitored environmental status and equipment response data are packaged and transmitted back to the optimization system to form a complete closed-loop feedback.

[0118] Execution layer actions:

[0119] Water injection pressure control: The system transmits the information to the water injection pump group through the on-site PLC system to control the pump speed and the on / off status of the pump group, and adjust the water injection pressure regularly;

[0120] Slurry flow control: Transmit to the flow control unit (valve / pump speed linkage system);

[0121] Cement slurry density control: The data is transferred to the automatic slurry mixing unit to adjust the solid-liquid ratio and realize slurry density control.

[0122] Monitoring layer actions:

[0123] Deploy pressure, temperature, flow and other sensors around cementing tools (sleeves, packers);

[0124] Add monitoring of the sleeve opening and closing status, packer setting status, and annular slurry return flow rate;

[0125] The sensor gateway collects S t+1 =[P t+1 ,T t+1 ,Q t+1 ];

[0126] Monitor whether key anomalies occur during cementing operations, such as abnormal fluctuations in return slurry volume and abnormal pressure increases.

[0127] Feedback layer action:

[0128] Data aggregation and preprocessing: The collected S t+1 Transmit to the edge control unit for preliminary data denoising and outlier detection;

[0129] Dynamic reflux mechanism: S t+1 and with The execution response data is packaged and sent back to the optimizer in step 3 as new input to trigger the next round of cementing parameter optimization.

[0130] To ensure the regularity of control instructions, a dual-loop architecture of high-frequency control and low-frequency monitoring is established:

[0131] Control instructions are refreshed every 5 seconds, and monitoring data is sampled every 1 second, forming a 1:5 execution / monitoring ratio, which improves system response speed;

[0132] A high-frequency execution-monitoring strategy is used in the early stages of cementing operations (e.g., initial grouting), and a low-frequency steady-state strategy is used in the later stages (e.g., after the packer is set) to reduce equipment load.

[0133] It is understandable that the data closed loop in the solution is: execution system -> on-site physical environment (equipment action + environmental response) -> monitoring system -> data preprocessing -> reflux optimizer; through regular equipment driving + synchronous environmental data monitoring + optimization system data reflux, the organic connection of "prediction-optimization-execution-feedback" in the cementing operation process is realized, ensuring continuous, stable and low-risk operation in the dynamic environment of cementing operations.

[0134] S5. Receive the latest environmental data and on-site equipment operating status, build a multi-dimensional joint early warning model, and dynamically monitor key risk points in cementing over the entire time domain;

[0135] Specifically, this step receives the latest environmental data S returned by step 4 "execution and feedback system" t+1 =P t+1 ,Tt+1 ,Q t+1 ] and on-site equipment operating status (including actual Execution status), and combined with the regular cement leakage risk prediction value R output in step 2 t+1 , forming a complete dynamic monitoring data stream, which serves as the input for the dynamic monitoring and early warning system in this step. A multi-channel dynamic monitoring and intelligent early warning system (GZ-DMP) dedicated to cementing was constructed. For high-risk scenarios such as cement backflow, cross-leakage, packer failure, and downhole pressure loss during the cementing process, a composite monitoring mechanism was designed to achieve full-time dynamic monitoring of key cementing risk points.

[0136] Furthermore, a physical risk field perception feature set is constructed:

[0137] Define the “cementing rheological stress-pressure-risk synergistic characteristics” Φ t , comprehensively measure the abnormal trend of the current cementing conditions:

[0138]

[0139] Where λ is the risk perception enhancement factor, R t+1 is the risk prediction value in step 2. This feature combines the synergistic effect of cementing flow, pressure and predicted risk. t+1 When the pressure tends to increase, the system automatically amplifies its sensitivity to the pressure-flow ratio, and is particularly proactive in detecting precursors of leakage (such as high pressure-low flow, abnormal slurry return).

[0140] Furthermore, a dynamic adaptive threshold mechanism is designed: combining the downhole environment model in step 1, the system automatically calculates the current dynamic warning threshold in the well, making the warning judgment process environmentally adaptive:

[0141]

[0142] Where α and β are weight factors of temperature and pressure respectively, T crit and P max As the downhole temperature or pressure approaches the limit, the system automatically lowers the alarm threshold θ t , enhance early warning sensitivity in high-risk environments.

[0143] Furthermore, a multi-dimensional joint early warning model is constructed based on Φ t and θ t Determine the risk level:

[0144] When Φ t <0.7θ t , the system determines it as "low risk";

[0145] When 0.7θ t ≤Φt <θ t , the system determines it as "medium risk" and triggers a local alarm;

[0146] When Φ t ≥θ t , the system determines it as "high risk", triggers a station-wide emergency warning, and generates a dynamic risk message, which is regularly sent back to the optimizer in step 3 as an important input for the next round of optimization, prompting the system to adjust the operation parameters for the abnormal points first.

[0147] Furthermore, a high-order risk confirmation mechanism based on time series anomaly detection is designed. The system is in a continuous K cycle, when Φ t Multiple crossings of θ t , and R t+1 If the synchronization is higher than 0.8, it is judged as a "suspected cement leakage event" and the main control system is automatically advised to lower the P i With Q i The optimization goal is to recommend a secondary verification of the packer seating status to form a software-hardware collaborative early warning.

[0148] Introducing risk trend smoothing suppression items to reduce short-term false alarms and ensure the stability of the early warning system:

[0149]

[0150] when When it is lower than the set threshold ζ, even if Φ t Short-term excess θ t , the system will maintain a medium risk level to avoid false alarms triggered by instantaneous pressure or flow fluctuations during cementing operations.

[0151] Furthermore, the early warning feedback closed loop mechanism:

[0152] Early warning information (risk level, abnormal operating conditions, key characteristic values) is synchronously transmitted back to the optimizer in step 3 and passed to the execution system in step 4 to support automatic adjustment of operation parameters;

[0153] For emergency control in high-risk situations, the system can directly interact with the execution system through an interface to temporarily apply a "safety mode", such as locking the upper pressure limit, reducing the water injection rate, etc., to assist in optimizing the system and alleviate cementing risks.

[0154] Unless otherwise specifically stated, the relative steps, numerical expressions and values ​​of the components and steps set forth in these embodiments do not limit the scope of the present application.

[0155] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a non-volatile computer-readable storage medium that is executable by a processor. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the system described in each embodiment of the present application. The aforementioned storage medium includes various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0156] In the description of this application, it should be noted that the terms "upper" and "lower" etc. indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, or the orientations or positional relationships in which the invented product is usually placed when in use. These are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation. Therefore, they should not be understood as limitations on this application.

[0157] It should also be noted that, in the description of this application, unless otherwise expressly specified or limited, the terms "disposed," "installed," "connected," and "connected" should be understood in a broad sense. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections or electrical connections; they can refer to direct connections or indirect connections through an intermediate medium; and they can refer to internal connections between two components. Those skilled in the art will understand the specific meanings of the above terms in this application based on the specific circumstances.

[0158] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some or all of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A cement leakage risk prediction method based on mid-process cementing technology is characterized by: The method comprises: Several sets of sensors are used to collect downhole periodic data simultaneously and regularly. Based on the periodic data, a physical model is constructed to simulate the flow state of cement slurry to obtain the pressure field, pressure field and temperature field. Among them, the several sets of sensors include pressure sensors, temperature sensors, mud flow sensors and displacement sensors. The sensor data and physical model are input into the machine learning model for training to obtain the cement leakage risk prediction value; Based on the cement leakage risk prediction value and sensor data, a reinforcement learning model is designed. A composite reward function is designed with minimization and operation smoothness as the goal to generate optimal adjustment values ​​for cementing operation parameters. The optimal adjustment values ​​for cementing operation parameters include adjustment amounts for downhole pressure, slurry flow rate, and cement slurry density. Mapping the optimal adjustment values ​​of cementing operation parameters to corresponding physical control devices, regularly adjusting downhole pressure, slurry flow rate, and cement slurry density, and regularly monitoring the physical conditions downhole through multimodal sensors; Receive the latest environmental data and on-site equipment operating status, build a multi-dimensional joint early warning model, and dynamically monitor key risk points in cementing over the entire time domain; The machine learning model is a physical information constrained deep neural network with dual-branch input, where: The first branch inputs periodic data and enters the standard fully connected layer network for feature extraction and gradual dimensionality reduction; The second branch inputs the pressure field, velocity field, and temperature field, and processes them through the Gaussian kernel convolution module to extract spatial distribution features; wherein the pressure field, velocity field, and temperature field are physical branch features; The features of the first branch and the second branch are spliced ​​in the fusion layer, and the fusion is input into the multi-layer perceptron network to output the cement leakage risk prediction value.

2. The cement leakage risk prediction method based on the semi-process cementing technology according to claim 1 is characterized in that: The physical model is: combining a fluid mechanics model and a stress analysis model in cementing operations to simulate downhole periodic data and simulate the flow state of cement slurry.

3. The cement leakage risk prediction method based on the semi-process cementing technology according to claim 1 is characterized in that: The first branch and the second branch are subjected to feature splicing in the fusion layer, and the fusion is input into the multi-layer perceptron network to output the cement leakage risk prediction value, specifically: The physical perception weight mechanism is introduced in the fusion layer, and a weighted splicing method is adopted. The fusion ratio of physical branch features and periodic data branch features is calculated by the physical perception weight w phys and periodic data weight w data Perform dynamic balance; where w phys and w data The sum of is 1; The multilayer perceptron network is a 5-layer MLP with 256, 128, 64, 32 and 1 neurons, respectively. It uses the ReLU activation function, combines the vortex balance constraint and the risk smoothing term, and outputs the cement leakage risk prediction value R t ∈[0,1]; The training of the machine learning model uses historical operation data and periodic data, combined with the Adam optimizer, with an initial learning rate of 1e-4, and dynamically adjusts the physical perception weight w during the training process. phys , so that the machine learning model can reasonably balance the feature fusion of data branch and physical branch at different cementing stages.

4. The method for predicting cement leakage risk based on the semi-process cementing technology according to claim 1 is characterized in that: According to the cement leakage risk prediction value and sensor data, a reinforcement learning model is designed, and a composite reward function is designed with the goal of minimizing and ensuring operation stability to generate optimal adjustment values ​​of cementing operation parameters, specifically: Taking the periodic data as a state space; The adjustment amount of downhole pressure, slurry flow rate and cement slurry density is designed as the action space. Physical boundary constraints are designed through action projection to limit the action space to the range allowed by the actual process. Combining the dual goals of risk minimization and operational stability, a compound reward system is designed. t : in, Is the physical boundary After that, the risk value predicted by the machine learning model is reused, and γ is the penalty factor. is the recommended value for the current stage; ΔP i is the downhole pressure adjustment, ΔQ i is the slurry flow adjustment amount.

5. The method for predicting cement leakage risk based on the mid-process cementing technology according to claim 4 is characterized in that: The reinforcement learning model structure is: Actor network: A 3-layer fully connected network using ReLU activation, outputting an action space consisting of three-dimensional continuous actions corresponding to adjustments to downhole pressure, slurry flow rate, and cement slurry density. Critic network: LSTM+FC structure, input state space and action space to capture the environmental temporal dependencies of the cementing process, output state-action value, and assist in Actor network strategy optimization; Among them, the algorithm flow of the reinforcement learning model is: Receive state space and cement leakage risk prediction value; The Actor network outputs the action space, which is adjusted by the physical boundaries. Apply the adjusted action space to update the current downhole pressure, slurry flow rate, and cement slurry density to generate new operating parameters; Apply new operating parameters to field operations; Obtain a new state space and a new cement leakage risk prediction value, and update the experience replay buffer.

6. The method for predicting cement leakage risk based on the semi-process cementing technology according to claim 1, characterized in that: The mapping of the optimal adjustment value of the cementing operation parameter to the corresponding physical control device is specifically as follows: Build a feedback system for the cementing operation process, including: execution layer, monitoring layer and feedback layer; Among them, the execution layer is used to control the current downhole pressure, slurry flow rate and cement slurry density; the monitoring layer is used to regularly collect current data through sensors; and the feedback layer is used to transmit the collected current data to the edge device for preliminary data denoising and outlier detection, and feed back to the reinforcement learning model for optimization.

7. The method for predicting cement leakage risk based on the mid-process cementing technology according to claim 5, characterized in that: The physical boundary also includes a temperature-sensitive weight mechanism: when the current temperature is greater than the temperature warning value, the current downhole pressure is automatically increased, and the penalty weight is adjusted to suppress pressure mutations in high-temperature environments, preventing the cement slurry from solidifying prematurely or flowing back due to temperature increases.

8. The method for predicting cement leakage risk based on the semi-process cementing technology according to claim 1, characterized in that: The receiving of the latest environmental data and on-site equipment operating status, the construction of a multi-dimensional joint early warning model, and the dynamic monitoring of key cementing risk points in the full time domain are specifically as follows: A physical risk field perception feature set is constructed based on the cement leakage risk prediction value and the predicted slurry flow rate to measure the abnormal trend of the current cementing conditions; Dynamic warning thresholds are designed based on downhole pressure and temperature thresholds to make the warning judgment process environmentally adaptive; Constructing a multi-dimensional joint warning model based on the physical risk field perception feature set and dynamic warning thresholds: low risk, medium risk and high risk; If the physical risk field perception feature set crosses the dynamic warning threshold multiple times within K consecutive cycles, and the cement leakage risk prediction value is simultaneously higher than 0.8, it is determined to be a suspected cement leakage event, and the main control system is automatically advised to lower the optimization targets of slurry flow and cement slurry density, or to perform a secondary verification of the packer sealing status.

9. The method for predicting cement leakage risk based on the mid-process cementing technology according to claim 8, characterized in that: A risk trend smoothing suppression term is also introduced to reduce false alarms caused by short-term fluctuations.

Images