A shale gas horizontal well geosteering method

Abstract

The present application relates to the directional drilling technical field, specifically to a shale gas horizontal well geosteering method. It comprises: collecting drilling mechanics data and logging physical property data; processing the drilling mechanics data in a first recurrent neural network to generate a first hidden state vector; processing the second time series in a second recurrent neural network to generate a second hidden state vector; fusing the first and second hidden state vectors to create a fused state representation vector; providing the fused state representation vector as a state input to a trained deep reinforcement learning agent; selecting an action by the agent according to its learned optimal policy, which constitutes a steering instruction for a directional drilling tool. The present application realizes a closed-loop, self-optimizing steering system by combining a prediction model with a decision-making agent, improves the drilling success rate of high-quality reservoirs, and reduces non-productive drilling time.

Description

Technical Field

[0001] This invention relates to the field of directional drilling technology, specifically a geological steering method for shale gas horizontal wells. Background Technology

[0002] In the field of oil and gas drilling, geosteering technology is mainly used to guide the direction of the drilling trajectory, adjusting the wellbore trajectory to the optimal position of the oil and gas reservoir to achieve the best oil (gas) production or water injection effect. Early geosteering technologies utilized geological models established before drilling and simple single-point logging-while-drilling (LWD) measurements, the most commonly used of which was natural gamma ray (GR) measurement. These methods can only detect deviations in the wellbore trajectory after the drill bit has drilled out of the target section. For thin reservoirs, formations with lateral discontinuities (such as pinch-outs), or formations with developed faults, these methods are almost unable to achieve precise trajectory control.

[0003] With technological advancements, geosteering tools with greater depth of exploration have emerged, such as azimuth electromagnetic resistivity (LWD) tools. These tools can detect the presence of formation boundaries before the drill bit crosses them, providing valuable lead time for adjusting the wellbore trajectory. These advanced detection tools are often combined with high-precision trajectory control systems such as rotary steerable systems (RSS) to achieve finer wellbore trajectory control. However, the process of acquiring data from downhole tools and transmitting it to the surface involves significant time lag, which can last for minutes or even longer. Furthermore, the interpretation of this complex data remains highly dependent on the human judgment of geosteering engineers. Engineers must manually build geological models, analyze data, and make steerability decisions in real time while drilling. A single decision can take 20 minutes or more and is highly susceptible to human error, subjective bias, and inconsistencies, especially in operations that require the simultaneous management of multiple wells.

[0004] Currently, there are also approaches using machine learning to attempt to automate and improve geosteering processes. However, many existing methods focus only on single-modality data, either using only LWD data or only surface drilling mechanics data. This approach fails to leverage the potential synergistic effects between different data types. Furthermore, some methods attempting to integrate multi-source data often employ overly simplistic fusion methods, failing to address the fundamental mismatch between high-frequency, real-time but ambiguous drilling mechanics data and low-frequency, high-fidelity but significantly time-delayed LWD data.

[0005] Therefore, there is a need for a method and system that can effectively integrate real-time but fuzzy mechanical data with lagging but accurate logging data to achieve predictive, automated and efficient guided decision-making. Summary of the Invention

[0006] To address the problems in the prior art, this invention provides a geological steering method for shale gas horizontal wells.

[0007] A geological steering method for shale gas horizontal wells includes the following steps:

[0008] Real-time acquisition of drilling mechanics data in the first time series;

[0009] Collect logging-while-drilling property data of a second time series that has a time delay relative to the first time series;

[0010] The first time series is processed in a first recurrent neural network to generate a first hidden state vector;

[0011] The second time series is processed in a second recurrent neural network to generate a second hidden state vector;

[0012] The first and second hidden state vectors are fused using an attention mechanism to create a fused state representation vector;

[0013] The fused state representation vector is provided as state input to a trained deep reinforcement learning agent;

[0014] The agent selects an action based on its learned optimal strategy, which constitutes a guidance instruction for a directional drilling tool.

[0015] Preferably, the first and second recurrent neural networks are bidirectional long short-term memory networks.

[0016] Preferably, the drilling mechanics data includes mechanical drilling speed, torque, drilling pressure, and drill string speed.

[0017] Preferably, the logging-while-drilling physical property data includes natural gamma and resistivity.

[0018] Preferably, the step of processing the first time series in the first recurrent neural network to generate the first hidden state vector specifically includes:

[0019] The first time series is processed in forward temporal order in a feedforward long short-term memory network to generate a feedforward hidden state vector;

[0020] The first time series is processed in reverse chronological order in a backward long short-term memory network to generate a backward hidden state vector;

[0021] The forward hidden state vector and the backward hidden state vector are concatenated to generate the first hidden state vector.

[0022] Preferably, the step of using an attention mechanism to fuse the first and second hidden state vectors to create a fused state representation vector specifically includes:

[0023] The first hidden state vector and the second hidden state vector are input into an attention network;

[0024] Through the attention network, a first attention score is calculated for the first hidden state vector, and a second attention score is calculated for the second hidden state vector;

[0025] The first hidden state vector and the second hidden state vector are weighted and summed to generate the fused state representation vector, wherein the first attention score and the second attention score are used as their respective weights.

[0026] Preferably, the learning of the optimal strategy specifically includes:

[0027] Environment initialization: Define the current state, which is represented by the fused state representation vector, including the distance to the nearest stratigraphic boundary, the probability of intersection with the preceding fault, and the rock brittleness index;

[0028] Action space definition: Define the action space;

[0029] Reward function definition: Define a reward function to evaluate the effect of each action;

[0030] Strategy selection: Select a strategy for reinforcement learning algorithms, including Q-learning and deep Q-networks;

[0031] State transition: The agent performs a selected action, transitions from the current state to a new state, and receives a corresponding reward;

[0032] Update strategy: Update the strategy based on the result of the current action;

[0033] Exploration and utilization: In the learning process, try new strategies and apply existing optimal strategies;

[0034] Termination condition: Set the termination condition for learning, such as stopping the algorithm after a fixed number of time steps or when the optimal guidance strategy is found.

[0035] Preferably, the reward function is configured to provide a positive reward when the wellbore trajectory is within the predetermined target geological layer, and a negative reward when the wellbore trajectory deviates from the target geological layer.

[0036] Preferably, the action space of the deep reinforcement learning agent includes a set of discrete instructions for adjusting the inclination angle or azimuth angle of the directional drilling tool.

[0037] Preferably, the guiding command is transmitted to a rotary guiding system.

[0038] Compared with the prior art, the advantages of this invention are:

[0039] This invention utilizes a dual-loop neural network to fuse high-frequency, latency-free drilling mechanics data with lagging but accurate LWD (Low-Dip Dynamics) property data in real time. Its attention mechanism dynamically assesses and assigns higher weight to mechanics data exhibiting anomalies (indicating potential formation changes). This allows the system to proactively adjust before the wellbore trajectory deviates, thereby maximizing wellbore coverage within the target "sweet spot" area.

[0040] Furthermore, by utilizing drilling mechanics data, this invention reduces the reliance on LWD tools. By employing the optimal strategy learned by a deep reinforcement agent, it can still achieve high-precision guidance even with only basic logging data, thereby effectively controlling drilling costs while ensuring drilling success rate. Attached Figure Description

[0041] Figure 1 This is a flowchart of a geological steering method for shale gas horizontal wells proposed in this invention. Detailed Implementation

[0042] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0043] Example 1:

[0044] This embodiment illustrates an exemplary system architecture of the present invention. The system typically includes surface equipment and downhole equipment.

[0045] The surface equipment includes a conventional drilling platform, a mud pump system, and a processing unit. This processing unit can be a high-performance computer located on the surface, or in other embodiments, it can be integrated into the downhole tool. The processing unit includes one or more processors and a non-transitory computer-readable medium storing instructions for executing the predictive neural network model of the present invention.

[0046] The core of the downhole equipment is the bottom hole assembly (BHA). A typical BHA configured to implement this invention comprises the following components:

[0047] A drill bit.

[0048] A set of sensors for acquiring drilling mechanics data. These sensors may include accelerometers for measuring vibration and shock, and strain gauges for measuring downhole torque and drilling pressure.

[0049] A set of sensors for acquiring logging-while-drilling (LWD) physical property data. These sensors may include natural gamma detectors, resistivity measurement tools at different depths, and formation density and neutron porosity measurement tools.

[0050] A rotary steering system (RSS) capable of receiving and executing fine steering commands. RSS technology enables continuous and precise control of the wellbore trajectory by applying lateral forces to the drill bit or adjusting the drill bit's direction.

[0051] Example 2:

[0052] This embodiment illustrates the data acquisition and processing procedures of the invention.

[0053] The model input of this invention consists of two data streams that are different in nature but complementary to each other.

[0054] The first data stream is mechanical data, which is a high-frequency, real-time data stream. This data stream is typically acquired at frequencies of 1 to 10 Hz, with near-zero latency at the drill bit, but its geological meaning is ambiguous and its fidelity is low. Specific parameters may include mechanical rate of penetration (ROP), downhole or surface torque, weight on bit (WOB), standpipe pressure (SPP), drill string speed (RPM), and mechanical specific energy (MSE).

[0055] The second data stream is physical property data, a low-frequency stream that lags behind the first data stream but offers high fidelity. This stream is typically acquired at a frequency of 0.1 to 1 Hz, but due to the physical distance between the sensor and the drill bit, as well as data transmission limitations, its delay at the drill bit can range from 1 to 15 minutes. Specific parameters may include natural gamma, resistivity, formation density, and neutron porosity.

[0056] The raw data collected from the two data streams needs to undergo a series of preprocessing steps before it can be used as input to the neural network. These steps include:

[0057] Noise filtering: Use methods such as moving average or Kalman filtering to remove high-frequency noise from the data.

[0058] Normalization: Scaling parameters with different dimensions (such as ROP and torque) to a uniform range (e.g., 0 to 1) to avoid certain parameters dominating during model training.

[0059] Time series alignment: The continuous data stream is divided into fixed-length time series windows to form the input tensor. When processing the second data stream, its inherent time delay needs to be considered, and it needs to be aligned with the mechanical data at the corresponding depth.

[0060] Example 3:

[0061] This embodiment provides a detailed description of the dual-loop neural network model in the scheme.

[0062] The main body of the model uses two parallel processing streams, each of which is a bidirectional long short-term memory network.

[0063] Since drilling and logging data are essentially time (or depth) series data, recurrent neural networks (RNNs) are a natural choice for processing this type of data. Secondly, to address the gradient vanishing or exploding problems that standard RNNs often encounter when processing long sequences, this invention employs Long Short-Term Memory (LSTM) units. LSTMs, through their internal gating mechanisms (input gate, forget gate, output gate), can selectively remember or forget historical information, thereby effectively learning long-term dependencies in the sequence.

[0064] Furthermore, this invention employs a bidirectional LSTM (Bi-LSTM). A standard LSTM can only process sequences in the forward temporal direction, while a Bi-LSTM consists of a forward LSTM and a backward LSTM. It can extract features from both past and future information simultaneously, which is crucial for geologically guided tasks. For example, the changing trends of strata depend not only on previously encountered strata but also on strata to be encountered later. By combining information from both directions, Bi-LSTM can generate a more comprehensive feature representation of the current geological environment.

[0065] Existing simple splicing methods cannot resolve the fundamental differences in information quality and time scale between two data streams. This invention employs an attention-based fusion mechanism to address the core conflict between decision-making speed and data certainty, as described above. Real-time mechanical data provides speed, but its geological implications are ambiguous (e.g., a decrease in ROP could be due to formation hardening, drill bit wear, or improper hydraulic parameters). Lagging LWD data provides certainty (e.g., an increase in gamma value clearly indicates an increase in mudstone content), but its information is delayed. Through the attention-based fusion mechanism, the model can learn to dynamically determine which data source should be trusted more under different circumstances.

[0066] Its technical implementation is as follows:

[0067] After processing their input sequences, the two Bi-LSTM streams each generate a hidden state vector. and These two vectors encode the temporal information in the mechanical data and physical property data, respectively.

[0068] These two hidden state vectors are fed together into a small, jointly trainable feedforward neural network, which is called an attention network.

[0069] The attention network calculates a set of attention scores ( and These two scores are scalars between 0 and 1, and their sum is 1. They represent the model's quantitative assessment of the importance of the two data streams at the current moment.

[0070] Final fusion vector We obtain the following by weighted summation of the two hidden state vectors:

[0071] ;

[0072] The weighted summation process is dynamic. The attention score is not fixed, but calculated in real time based on the input data at each moment.

[0073] Its specific working principle can be illustrated through the following example:

[0074] Scenario 1: Smooth Drilling. When the wellbore trajectory is smoothly drilling in a homogeneous shale reservoir, real-time mechanical data (ROP, torque) may not change much and are relatively "informative." However, lagging LWD data shows that gamma values ​​and resistivity are within the target range. In this case, a trained attention mechanism will determine that the LWD data provides more reliable "on-target" confirmation information, and therefore will assign a higher weight to the LWD flow (e.g., ...). ), and assign a lower weight to the mechanical flow (e.g. ).

[0075] Scenario 2: Encountering a Hard Interlayer. Suppose the drill bit is about to encounter a hard carbonate rock interlayer not shown in the seismic model. The moment the drill bit contacts this interlayer, the mechanical data will respond immediately: ROP will drop sharply, and torque will increase sharply. Meanwhile, the LWD sensor located several meters behind the drill bit is still measuring the shale's physical properties, and its data has not yet changed. The attention mechanism will capture this high-variance, anomalous signal pattern in the mechanical data stream and determine that it is an important signal indicating a change in the formation. Therefore, it will dynamically shift its attention to the mechanical data stream, assigning it a very high weight (e.g., ...). ), while giving a very low weight to the lagging LWD data stream (e.g. ).

[0076] In this way, the system can use the real-time nature of mechanical data to make predictive responses to impending geological changes, rather than waiting for LWD data to confirm the change minutes later before making passive adjustments.

[0077] Example 4:

[0078] This embodiment describes the guidance strategy based on deep reinforcement learning in the scheme.

[0079] In this scheme, the geological guidance task is modeled as a Markov decision process, and its core elements are defined as follows:

[0080] State: At each decision point, the state of the environment is a fusion vector output by the aforementioned dual-loop neural network model. This vector comprehensively describes the geological environment of the current wellbore and the predictions ahead, such as the distance to the nearest formation boundary, the probability of intersection with a fault ahead, and the rock brittleness index.

[0081] Actions: The action space is defined as a discrete, physically executable set of steering commands. For example, actions may include a series of specific commands to the rotary steering system (RSS), such as "increase the wellbore inclination angle by 0.2 degrees", "decrease the wellbore inclination angle by 0.2 degrees", and "maintain the current angle".

[0082] Reward: The design of the reward function is crucial for guiding the agent's learning. In this embodiment, the reward function can be defined as:

[0083] For every meter drilled into the target "sweet spot" reservoir, the reward is +1.

[0084] Drilling into the target reservoir results in a large negative reward, such as -50.

[0085] If the curvature of the wellbore trajectory is too large, a negative reward proportional to the curvature is given to ensure the smoothness of the wellbore trajectory.

[0086] With this setup, the goal of a deep reinforcement learning agent (e.g., a deep Q-network) is to learn an optimal policy that selects an action for any given state that maximizes the long-term cumulative discount reward.

[0087] Example 5:

[0088] This embodiment provides a detailed description of the training of the model in this scheme.

[0089] The training of this system consists of two steps:

[0090] State Representation Network Training: First, the aforementioned dual-recurrent neural network model is trained using historical data from adjacent wells or pilot wells through supervised learning. The goal is for the network to learn to generate high-quality state representation vectors that accurately reflect the actual geological conditions from the input mechanical and physical property data. .

[0091] Reinforcement Learning Policy Training: Subsequently, a deep reinforcement learning agent is trained in a high-fidelity simulated drilling environment. In this environment, the pre-trained state representation network is used to generate states. The agent selects an action based on the current state, and the simulator updates the wellbore trajectory and geological environment based on this action, returning a new state and a reward. Through millions of iterations of trial and error, the agent continuously updates its internal Q-value network, ultimately learning the optimal guidance strategy under various complex geological scenarios.

[0092] Example 6:

[0093] This embodiment illustrates the method in this scheme through a typical real-time geological guidance operation, including:

[0094] They began drilling a new horizontal well.

[0095] The system continuously collects mechanical and physical property data through sensors on the BHA.

[0096] A pre-trained dual-stream neural network deployed on the processing unit receives and processes the two data streams in real time, generating a representation vector of the current state.

[0097] At each decision point, the state vector is input into a pre-trained deep reinforcement learning agent.

[0098] Deep reinforcement learning agents select the optimal action (i.e., guiding instruction) for the current state based on the optimal policy they have learned.

[0099] The instruction was automatically sent to the downhole RSS system.

[0100] The RSS makes minute and precise trajectory adjustments based on instructions, thereby actively and autonomously maintaining the wellbore trajectory within the "sweet spot" area.

[0101] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0102] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for geosteering a horizontal shale gas well, characterized in that, The method comprises the following steps: collecting a first time series of drilling mechanics data in real time; collecting a second time series of logging-while-drilling physical property data with a time delay relative to the first time series; processing the first time series in a first recurrent neural network to generate a first hidden state vector; processing the second time series in a second recurrent neural network to generate a second hidden state vector; fusing the first and second hidden state vectors using an attention mechanism to create a fused state representation vector; providing the fused state representation vector as state input to a trained deep reinforcement learning agent; selecting, by the agent, an action according to the optimal policy learned by it, the action constituting a steering instruction for a directional drilling tool; the step of fusing the first and second hidden state vectors using an attention mechanism to create a fused state representation vector specifically comprises: inputting the first hidden state vector and the second hidden state vector into an attention network; calculating, by the attention network, a first attention score for the first hidden state vector and a second attention score for the second hidden state vector; performing weighted summation on the first hidden state vector and the second hidden state vector to generate the fused state representation vector, wherein the first attention score and the second attention score are used as respective weights; wherein the process of weighted summation is dynamic, and the attention scores are not fixed but calculated in real time according to the input data at each time.

2. The method of claim 1, wherein, The first and second recurrent neural networks are bidirectional long short-term memory networks.

3. The method of claim 1, wherein, The drilling mechanics data includes rate of penetration, torque, weight on bit, and drill string speed.

4. The method of claim 1, wherein, The logging-while-drilling physical property data includes natural gamma and resistivity.

5. The method of claim 1, wherein, The step of processing the first time series in a first recurrent neural network to generate a first hidden state vector specifically comprises: processing the first time series in a forward long short-term memory network in a time forward order to generate a forward hidden state vector; processing the first time series in a backward long short-term memory network in a time backward order to generate a backward hidden state vector; splicing the forward hidden state vector and the backward hidden state vector to generate the first hidden state vector.

6. The method of claim 1, wherein, The learning of the optimal policy specifically comprises: environment initialization: defining a current state represented by the fused state representation vector, including the distance to the nearest formation boundary, the probability of intersecting a fault ahead, and the rock brittleness index; action space definition: defining an action space; reward function definition: setting a reward function for evaluating the effect of each action; policy selection: selecting a policy of a reinforcement learning algorithm, including Q-learning and deep Q-network; state transition: the agent performs the selected action to change from the current state to a new state and obtains a corresponding reward; updating the policy: updating the policy according to the result of the current action; exploration and utilization: trying new strategies and applying the existing optimal strategy during the learning process; Termination condition: set the termination condition of learning, such as after a fixed number of time steps or when the optimal steering strategy is found, the algorithm stops running.

7. The method of claim 6, wherein, The reward function is configured to provide a positive reward when the wellbore trajectory is within a predetermined target geological layer, and a negative reward when the wellbore trajectory deviates from the target geological layer.

8. The method of claim 6, wherein, The action space of the deep reinforcement learning agent includes a set of discrete instructions for adjusting the inclination angle or azimuth angle of the directional drilling tool.

9. The method of claim 1, wherein, The steering instructions are transmitted to a rotary steerable system.

Images