Shale gas horizontal well geosteering method

Abstract

The invention relates to the technical field of directional drilling, in particular to a shale gas horizontal well geosteering method. Comprising the steps that drilling mechanical data and logging while drilling physical property data are collected; processing the drilling mechanics data in a first recurrent neural network to generate a first hidden state vector; processing the second time sequence 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 state input to a trained deep reinforcement learning agent; and selecting an action by the intelligent agent according to the learned optimal strategy, wherein the action forms a guide instruction for the directional drilling tool. By combining the prediction model and the decision agent, a closed-loop and self-optimized guiding system is realized, the drilling rate of a high-quality reservoir is improved, and the non-productive drilling time is shortened.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of directional drilling, in particular to a shale gas horizontal well geosteering method. BACKGROUND

[0002] In the field of oil and gas drilling, geosteering technology is mainly used to guide the direction of the drilling trajectory, adjusting the well trajectory to the best position of the oil and gas reservoir to achieve the best oil (gas) production or water injection effect. Early geosteering technology uses the geological model established before drilling and simple single-point LWD measurements, among which the most commonly used is the natural gamma (GR) measurement. These methods can only detect the deviation of the well trajectory after the drill bit has drilled through the target interval. For thin reservoirs, laterally discontinuous (such as pinchout) or faulted formations, these methods are almost unable to achieve precise trajectory control.

[0003] With the development of technology, geosteering tools with deeper detection depth have emerged, such as azimuthal electromagnetic resistivity LWD tools. These tools can detect the presence of boundaries before the drill bit passes through them, providing valuable advance for adjusting the well trajectory. These advanced detection tools are usually combined with high-precision trajectory control systems such as RSS to achieve more precise well trajectory control. However, there is a significant time lag in the process of obtaining data from the downhole tool and transmitting it to the surface, which can be as long as several minutes or even longer, and the interpretation of these complex data still highly depends on the manual judgment of geosteering engineers. Engineers must manually establish a geological model, analyze data and make steering decisions in real time while drilling. The development of a decision can take 20 minutes or more, and is highly susceptible to human error, subjective bias and inconsistency, especially in operations that require simultaneous management of multiple wells.

[0004] There are also current solutions that apply machine learning to try to automate and improve the geosteering process. However, many existing methods only focus on single modal data, either using only LWD data or only using surface drilling mechanics data. This approach fails to take advantage of the potential synergies between different data types. While some methods attempt to integrate multi-source data, their fusion methods are often too simple and fail to address the fundamental mismatch between high-frequency, real-time but ambiguous drilling mechanics data and low-frequency, high-fidelity but significantly time-lagged LWD data.

[0005] Therefore, there is a need for a method and system that can effectively fuse real-time but ambiguous mechanical data with lagging but accurate logging data to achieve predictive, automated and efficient steering decisions. SUMMARY

[0006] To solve the problems in the prior art, the present application provides a shale gas horizontal well geosteering method.

[0007] A shale gas horizontal well geosteering method, comprising the following steps: collecting first time-series drilling mechanics data in real time; collecting second time-series logging physical property data with 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 an action by the agent according to the optimal policy it has learned, which constitutes a steering instruction for directional drilling tools.

[0008] Preferably, the first and second recurrent neural networks are bidirectional long short-term memory networks.

[0009] Preferably, the drilling mechanics data includes mechanical drilling speed, torque, drilling pressure and drill string drilling speed.

[0010] Preferably, the logging physical property data includes natural gamma and resistivity.

[0011] Preferably, the processing of 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 time forward order to generate a forward hidden state vector; processing the first time series in a backward long short-term memory network in time reverse 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.

[0012] Preferably, the fusing of 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 a first attention score for the first hidden state vector and a second attention score for the second hidden state vector; performing a weighted sum of 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.

[0013] Preferably, the learning of the optimal policy specifically comprises: environment initialization: defining a current state represented by a fused state representation vector, including distance to the nearest stratigraphic boundary, probability of forward fault intersection, 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, deep Q-network; state transition: the agent performs the selected action, transitions from the current state to a new state, and obtains the corresponding reward; updating the policy: updating the policy according to the result of the current action; exploration and exploitation: trying new strategies and applying existing optimal strategies during the learning process; termination condition: setting the termination condition of learning, such as stopping the algorithm when a fixed number of time steps have elapsed or when the optimal steering policy is found.

[0014] Preferably, the reward function is configured to provide a positive reward when the well trajectory is within a predetermined target geological layer, and a negative reward when the well trajectory deviates from the target geological layer.

[0015] 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.

[0016] Preferably, the steering instructions are transmitted to a rotary steering system.

[0017] Compared with the prior art, the advantages of the present application are: Through a double recurrent neural network, the present application can fuse high-frequency, non-delayed drilling mechanics data and lagging but accurate LWD physical property data in real time. Its attention mechanism can dynamically judge and give higher weight to mechanics data when it appears abnormal (indicating changes in the front formation). This enables the system to actively adjust before the well trajectory deviates, thereby maximizing the retention of the well trajectory within the target "sweet spot" area.

[0018] And the present application reduces the dependence on LWD tools by using drilling mechanics data, and by using the optimal strategy learned by the depth reinforcement intelligent agent, high-precision steering can be realized even in the case of only basic logging data, so as to effectively control the drilling cost while ensuring the drilling rate. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 A method flowchart of a shale gas horizontal well geosteering method is proposed in the present application. DETAILED DESCRIPTION

[0020] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0021] Embodiment 1: An exemplary system structure of the present application is described in this embodiment. The system generally includes surface equipment and downhole equipment.

[0022] The surface equipment includes a conventional drilling platform, a mud pump system, and a processing unit. The processing unit can be a high-performance computer located on the ground, or can be integrated into the downhole tool in other implementations. 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 application.

[0023] The core of the downhole equipment is the bottom hole assembly (BHA). A typical BHA is configured to implement the present application, and its components include: A drill bit.

[0024] A set of sensors for collecting drilling mechanics data. These sensors can include accelerometers for measuring vibration and impact, and strain gauges for measuring downhole torque and weight on bit.

[0025] A set of sensors for collecting logging-while-drilling (LWD) physical property data. These sensors can include a natural gamma detector, resistivity measurement tools of different detection depths, formation density and neutron porosity measurement tools.

[0026] A rotary steerable system (RSS) capable of receiving and executing fine steering instructions. RSS technology can achieve continuous and accurate control of the wellbore trajectory by applying lateral force or adjusting the pointing of the drill bit.

[0027] Embodiment 2: This embodiment illustrates the data acquisition and processing procedure in the present invention.

[0028] The model input of the present invention is two data streams with different properties but complementary to each other.

[0029] The first data stream is mechanical data, which is a high-frequency real-time data stream. This data stream is usually collected at a frequency of 1 to 10 Hz, with a delay close to zero at the drill bit, but its geological meaning is ambiguous and has low fidelity. Specific parameters can include rate of penetration (ROP), downhole or surface torque, weight on bit (WOB), standpipe pressure (SPP), drill string rotation rate (RPM), and mechanical specific energy (MSE).

[0030] The second data stream is physical property data, which is a low-frequency data stream, lagging behind the first data stream, but with high fidelity. This data stream is usually collected at a frequency of 0.1 to 1 Hz, but due to the physical distance between the sensor and the drill bit and the limitations of data transmission, its delay at the drill bit can be between 1 to 15 minutes. Specific parameters can include natural gamma, resistivity, formation density, and neutron porosity.

[0031] The raw data collected from the two data streams needs to go through a series of preprocessing steps before it can be used as input for the neural network. These steps include: Noise filtering: Use methods such as moving average or Kalman filtering to remove high-frequency noise in the data.

[0032] Normalization: Scale parameters of different dimensions (such as ROP and torque) to a unified range (e.g. 0 to 1) to avoid certain parameters dominating in model training.

[0033] Time series alignment: Cut the continuous data stream into fixed-length time series windows to form input tensors. When processing the second data stream, its inherent time delay needs to be considered and aligned with the mechanical data at the corresponding depth.

[0034] Embodiment 3: In this embodiment, the dual-cycle neural network model in the scheme is described in detail.

[0035] The main body of the model uses two parallel processing streams, each of which is a bidirectional long short-term memory network.

[0036] Since drilling and logging data are essentially time (or depth) series data, recurrent neural networks (RNN) are a natural choice for processing such data. Secondly, to address the problem of gradient vanishing or exploding that standard RNNs are prone to when dealing with long sequences, the present invention employs long short-term memory (LSTM) units. LSTM is able to selectively remember or forget history information through its internal gating mechanism (input gate, forget gate, output gate), thus effectively learning long-term dependencies in the sequence.

[0037] Furthermore, the present invention employs bidirectional LSTM (Bi-LSTM). A standard LSTM can only process a sequence in the forward direction in time, while a Bi-LSTM consists of a forward LSTM and a backward LSTM. It is able to extract features from both past and future information, which is crucial for the geosteering task. For example, the trend of a formation does not only depend on the formation that has been drilled, but also on the formation that will be drilled. By combining information from both directions, Bi-LSTM is able to generate a more comprehensive feature representation of the current geological environment.

[0038] The simple concatenation method in the prior art cannot solve the fundamental difference between the two data streams in terms of information quality and time scale. The present invention employs an attention-based fusion mechanism to address the core conflict between the decision speed and data certainty described in the foregoing. Real-time mechanical data provides speed, but its geological meaning is ambiguous (for example, a drop in ROP could be due to a harder formation, or it could be due to drill bit wear or improper hydraulic parameters). The lagging LWD data provides certainty (for example, a rise in gamma value clearly indicates an increase in shale content), but its information is delayed. Through the attention-based fusion mechanism, the model can learn to dynamically determine which data source to believe more in different situations.

[0039] The technical implementation is as follows: After the two Bi-LSTM streams process their input sequences respectively, each generates a hidden state vector (h and ), which encode the time series information in mechanical data and physical property data, respectively.

[0040] The two hidden state vectors are input together into a small, jointly trainable feedforward neural network, which is the attention network.

[0041] The attention network calculates a set of attention scores (a and ), which are scalars between 0 and 1, and their sum is 1. They represent the model's quantitative assessment of the importance of information from the two data streams at the current time.

[0042] The final fusion vector The two hidden state vectors are summed by weighting: ; Where the process of weighted summation is dynamic. The attention score is not fixed, but calculated in real time according to the input data at each moment.

[0043] The specific working principle can be illustrated by the following examples: Scenario one: smooth drilling. When the well trajectory is smoothly drilled in a homogeneous shale reservoir, the real-time mechanical data (ROP, torque) may not change much, and there is relatively "no information content". The lagged LWD data shows that the gamma value and the resistivity are within the target range. In this case, the trained attention mechanism will judge that the LWD data provides more reliable "on-target" confirmation information, so it will assign a higher weight to the LWD stream (for example ), and a lower weight to the mechanical stream (for example ).

[0044] Scenario two: suddenly encountering a hard interlayer. Suppose the drill bit is about to drill a hard carbonate interlayer that is not shown in the seismic model. At the moment when the drill bit contacts the interlayer, the mechanical data will immediately respond: the ROP will drop sharply, and the torque will rise sharply. At this time, the LWD sensor located several meters behind the drill bit is still measuring the physical properties of the shale, and its data has not changed. The attention mechanism will capture this high-variance, abnormal signal pattern in the mechanical data stream and judge that it is an important signal indicating a change in the formation. Therefore, it will dynamically shift its attention to the mechanical data stream and assign a very high weight (for example ), and give a very low weight to the lagged LWD data stream (for example ).

[0045] In this way, the system can use the real-time nature of the mechanical data to make a predictive response to the impending geological change, rather than waiting for the LWD data to confirm the change and then making a passive adjustment.

[0046] Embodiment 4: In this embodiment, the depth reinforcement learning-based steering strategy in the scheme is described.

[0047] In this scheme, the geosteering task is modeled as a Markov decision process, and its core elements are defined as follows: State: At each decision point, the state of the environment is represented by the fusion vector is represented by a vector. This vector comprehensively describes the current geologic environment the wellbore is in and the prediction of the front, such as the distance to the nearest geologic boundary, the probability of intersecting a fault in the front, the rock brittleness index, etc.

[0048] Action: The action space is defined as a discrete set of physically executable steering instructions. For example, the actions can include "increase the hole angle by 0.2 degrees", "decrease the hole angle by 0.2 degrees", "hold the current angle", etc. a series of specific instructions to the rotary steerable system (RSS).

[0049] Reward: The design of the reward function is the key to guide the learning of the agent. In this embodiment, the reward function can be defined as: +1 for every meter drilled into the target "sweet spot" reservoir.

[0050] A large negative reward, e.g. -50, is given for drilling out of the target reservoir.

[0051] A negative reward proportional to the curvature size is given for a too large curvature of the wellbore trajectory, to ensure the smoothness of the wellbore trajectory.

[0052] With this setup, the goal of the deep reinforcement learning agent (e.g. a deep Q network) is to learn an optimal policy that can choose an action for any given state that maximizes the future long-term cumulative discounted reward.

[0053] Embodiment 5: The training of the model in this scheme is described in detail in this embodiment.

[0054] The training of the system is divided into two steps: State representation network training: First, the aforementioned bi-lstm network model is trained using historical data from neighboring wells or pilot wells through supervised learning. The goal is to let the network learn to generate high-quality state representation vectors that accurately reflect the true geologic conditions from the input mechanical and physical data .

[0055] Reinforcement learning strategy training: Subsequently, a deep reinforcement learning agent is trained in a highly realistic simulated drilling environment. In this environment, the state representation network trained above is used to generate the state. The agent selects an action according to the current state, and the simulator updates the wellbore trajectory and geologic environment according to the action and returns a new state and a reward. Through millions of "trial and error" iterations, the agent constantly updates its internal Q value network, and finally learns the optimal steering strategy in various complex geologic scenarios.

[0056] Embodiment 6: The method in the scheme is illustrated by a typical real-time geosteering operation in this embodiment, including: Start drilling a new horizontal well.

[0057] The system continuously collects mechanical and physical data through sensors on the BHA.

[0058] The pre-trained dual-stream neural network deployed on the processing unit receives and processes the two data streams in real time to generate a representation vector of the current state.

[0059] At each decision point, the state vector is input into the pre-trained deep reinforcement learning agent.

[0060] The deep reinforcement learning agent selects the optimal action (i.e., steering instruction) in the current state according to the optimal policy it has learned.

[0061] The instruction is automatically sent to the RSS system downhole.

[0062] The RSS makes small and precise trajectory adjustments according to the instruction, thereby actively and autonomously maintaining the wellbore trajectory within the "sweet spot" area.

[0063] In the description of the present specification, the description of the terms "one embodiment", "example", "specific example", etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In the present specification, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0064] The preferred embodiments of the application disclosed above are only used to help explain the application. The preferred embodiments do not describe all the details, nor limit the application to the specific embodiments described. Obviously, many modifications and changes can be made according to the content of the present specification. The present specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the application, so that those skilled in the art can well understand and utilize the application. The application is limited only by the claims and their full scope and equivalents.

Claims

1. A shale gas horizontal well geosteering method, characterized in that: The following steps are involved: Real-time collection of first time series drilling mechanics data; Acquiring a second time series of logging-while-drilling physical property data having 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 a state input to a trained deep reinforcement learning agent; The agent selects an action according to the optimal strategy it has learned, and the action constitutes a steering instruction for the directional drilling tool.

2. A shale gas horizontal well geosteering method according to claim 1, characterized in that: The first and second recurrent neural networks are bidirectional long short-term memory networks.

3. The shale gas horizontal well geosteering method according to claim 1, characterized in that: The drilling mechanics data include mechanical penetration rate, torque, bit weight and drill string penetration rate.

4. The shale gas horizontal well geosteering method according to claim 1, characterized in that: The LWD physical property data include natural gamma and resistivity.

5. The shale gas horizontal well geosteering method according to claim 1, characterized in that: The processing of the first time series in the first recurrent neural network to generate a first hidden state vector specifically includes: Processing the first time series in a forward long short-term memory network in a forward chronological order to generate a forward hidden state vector; processing the first time series in a backward long short-term memory network in reverse time order to generate a backward hidden state vector; The forward hidden state vector and the backward hidden state vector are concatenated to generate the first hidden state vector.

6. The shale gas horizontal well geosteering method according to claim 1, characterized in that: The method of fusing the first and second hidden state vectors using an attention mechanism to create a fused state representation vector 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; A weighted sum is performed 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.

7. The shale gas horizontal well geosteering method according to claim 1, characterized in that: The learning of the optimal strategy specifically includes: Environment initialization: define the current state, which is represented by a fused state representation vector, including the distance to the nearest stratum boundary, the probability of intersection of the forward fault, and the rock brittleness index; Action space definition: define the action space; Reward function definition: Set the reward function to evaluate the effect of each action; Strategy selection: Select the strategy for reinforcement learning algorithms, including Q-learning and deep Q-networks; State transition: The agent performs the selected action, transitions from the current state to the new state, and obtains the corresponding reward; Update strategy: Update the strategy based on the result of the current action; Explore and exploit: During the learning process, try new strategies and apply existing optimal strategies; Termination condition: Set the termination condition for learning, such as when a fixed number of time steps have passed or the optimal guidance strategy is found, the algorithm stops running.

8. The shale gas horizontal well geosteering method according to claim 7, characterized in that: The reward function is configured to provide a positive reward when the wellbore trajectory is within a predetermined target geological formation and to provide a negative reward when the wellbore trajectory deviates from the target geological formation.

9. The shale gas horizontal well geosteering method according to claim 7, characterized in that: The action space of the deep reinforcement learning agent includes a set of discrete instructions for adjusting the well inclination or azimuth of the directional drilling tool.

10. The shale gas horizontal well geosteering method according to claim 1, characterized in that: The steering commands are transmitted to a rotary steerable system.

Images