A well-ground multi-domain data-based injection-production linkage control method

Abstract

The application discloses a kind of injection-production linkage control methods based on well-ground multi-domain data, belong to injection-production monitoring and control technical field, can solve the problem that present technology cannot realize the fusion perception of underground-ground multi-source information, second-level abnormal early warning and the linkage intelligent control of injection-production parameter simultaneously.The method comprises the following steps: S1, according to the state feature vector of the well to be measured in the corresponding sampling period is determined according to the collected downhole sensing data and ground sensing data;S2, according to the state feature vector of the well to be measured in different sampling periods to determine the risk index of the well to be measured in the current sampling period;S3, according to the state feature vector and risk index of the well to be measured in the current sampling period, the return function of the well to be measured in the current sampling period is constructed;S4, the return function of the well to be measured in the current sampling period is solved, and the injection-production equipment of the well to be measured is controlled according to the optimal control strategy obtained by solving in the next sampling period injection-production action.The application is used for injection-production linkage control.

Description

Technical Field

[0001] This invention relates to a method for injection-production linkage control based on well-to-ground multi-domain data, belonging to the field of injection-production monitoring and control technology. Background Technology

[0002] Currently, the monitoring and control of oil well or geothermal well injection and production processes mostly rely on traditional manual or semi-automatic methods, lacking second-level real-time early warning and linkage control capabilities. The main shortcomings are as follows:

[0003] (1) Single-source data monitoring and delayed response: Existing injection and production monitoring usually relies on single-point data such as wellhead pressure and temperature, and alarms are triggered by setting simple thresholds. When anomalies occur downhole or on the surface, traditional SCADA (Supervisory Control and Data Acquisition) systems often only issue alarms after parameters exceed limits, and rely on manual analysis and decision-making, resulting in delayed responses and difficulty in timely prevention of emergencies. For example, remote oil wells usually rely on manual inspections, which cannot achieve continuous real-time monitoring, so anomalies are often only discovered after they have caused an impact. Because only single-source data is used, it is impossible to fully depict the linkage effect of multiple physical scenarios downhole and on the surface, and there is a risk of missed and false alarms.

[0004] (2) Lack of multi-source data fusion and intelligent analysis: Information provided by downhole sensors (such as bottom hole pressure gauges) and surface sensors (such as wellhead pressure, injection / production flow meters, etc.) is scattered across different domains. Existing technologies lack effective data fusion methods and cannot comprehensively judge reservoir dynamic changes in real time. For example, traditional reservoir analysis relies more on offline numerical simulations and engineers' experience, which are time-consuming and difficult to guide injection and production adjustments in a timely manner. Some studies have begun to apply machine learning to water injection optimization and production prediction, such as using random forest models to predict the trend of water cut increase or LSTM neural networks to predict changes in oil well parameters. However, these methods usually require a large amount of historical data for training and are computationally complex, making it impossible to run them directly at high frequency at the field edge and difficult to provide timely warnings and handle anomalies. In addition, deep learning-based solutions are subject to hardware limitations in field deployment, the models are not transparent, and they cannot flexibly adapt to changing working conditions.

[0005] (3) Lack of coordinated optimization in injection and production control: Traditional water injection schemes are mostly pre-planned setpoints or simple feedback control, while the production parameters of oil wells (such as pumping unit frequency and nozzle opening) are adjusted by independent strategies. The injection and production stages operate independently and lack coordination. In practice, some oilfield trials have used reinforcement learning to optimize water injection well allocation schemes and dynamically adjust wellhead pressure, achieving an increase in production of about 8%. This indicates that intelligent algorithms have potential in optimizing injection and production parameters. However, such technologies focus on macro-level optimization to improve recovery rates, and strategies are usually distributed after offline calculations, making it difficult to handle abnormal operating conditions that occur instantaneously during production. More importantly, most existing schemes run in the cloud or on central servers, which are highly dependent on the network and cannot respond within seconds in emergency situations on-site, making it difficult to meet the safety requirements of high-risk scenarios.

[0006] (4) Lack of edge intelligent early warning and automatic intervention: At present, the digital construction of oil and gas fields has begun to introduce "edge computing + AI" to improve real-time monitoring capabilities. For example, edge controllers are deployed at the well site to analyze single-well dynamometer cards, electrical parameters, etc. in real time, thereby identifying abnormal operating conditions and implementing protective well shutdown and alarm. There are also industrial intelligent platforms that use edge data mining and trend prediction to achieve early warning of abnormal equipment status. However, these applications mostly focus on the safety monitoring of single equipment or single wells, without integrating data from multiple wells and multiple sensors, and without combining them with injection and production control strategies. When complex anomalies occur across well domains (such as a sudden change in production of adjacent wells due to an abnormality in a water injection well), there is a lack of global linkage control strategies. In addition, some solutions still rely on specific AI models (such as CNN, LSTM, etc.) for pattern recognition, which has high requirements for computing resources and is not conducive to long-term stable operation on resource-constrained edge nodes. In summary, existing technologies cannot simultaneously achieve the fusion perception of multi-source information from downhole and surface, second-level anomaly early warning, and intelligent linkage control of injection and production parameters. They have drawbacks such as untimely response, lack of global optimization, and reliance on dedicated algorithm models, and cannot meet the needs of efficient oil and gas field development and inherent safety. Summary of the Invention

[0007] This invention provides an injection-production linkage control method based on well-to-surface multi-domain data, which can solve the problems of existing technologies that cannot simultaneously achieve fusion perception of downhole-to-surface multi-source information, second-level anomaly early warning, and linkage intelligent control of injection and production parameters, and have problems such as untimely response, lack of global optimization, and reliance on dedicated algorithm models.

[0008] This invention provides a method for injection-production linkage control based on well-to-ground multi-domain data, characterized in that the method includes:

[0009] S1. Collect downhole sensing data and surface sensing data of the well to be logged in each sampling period, and determine the state feature vector of the well to be logged in the corresponding sampling period based on the downhole sensing data and the surface sensing data.

[0010] S2. Determine the risk index of the well to be logged in the current sampling period based on the state feature vector of the well to be logged in different sampling periods;

[0011] S3. Construct the return function of the well to be logged in the current sampling period based on the state feature vector and risk index of the well to be logged in the current sampling period;

[0012] S4. Solve for the return function of the well to be logged in the current sampling period to obtain the optimal control strategy of the well to be logged in the next sampling period, and control the injection and production operation of the injection and production equipment of the well to be logged in the next sampling period according to the optimal control strategy.

[0013] Optionally, determining the state feature vector of the well to be logged in the corresponding sampling period based on the downhole sensing data and the surface sensing data in S1 specifically includes:

[0014] The downhole sensing data and the surface sensing data are preprocessed to obtain downhole data vectors and surface data vectors, respectively.

[0015] The downhole data vector and the surface data vector are spliced ​​together to form the state feature vector of the well to be measured in the corresponding sampling period.

[0016] Optionally, the downhole data vector and the surface data vector are concatenated to form a state feature vector of the well to be logged in the corresponding sampling period, specifically as follows:

[0017] Determine the derived vector based on the downhole data vector and the surface data vector;

[0018] The downhole data vector, the surface data vector, and the derived vector are sequentially concatenated to form the state feature vector of the well to be measured in the corresponding sampling period.

[0019] Optionally, S2 specifically includes:

[0020] A reference state vector is determined based on the state feature vector of the well to be logged in the historical sampling period;

[0021] The abnormal indicators of the well to be logged in the current sampling period are determined based on the reference state vector and the state feature vector of the well to be logged in the current sampling period.

[0022] The risk index of the well to be logged in the current sampling period is determined based on the abnormal indicators of the well to be logged in the current sampling period.

[0023] Optionally, a reference state vector is determined based on the state feature vector of the well to be logged during the historical sampling period, specifically as follows:

[0024] The average vector of the state feature vectors of the well to be measured in the first M sampling periods is used as the reference state vector; where M is an integer greater than or equal to 2.

[0025] Optionally, determining the abnormal indicators of the well to be logged in the current sampling period based on the reference state vector and the state feature vector of the well to be logged in the current sampling period specifically includes:

[0026] Calculate the residual vector between the state feature vector of the well to be measured in the current sampling period and the reference state vector;

[0027] The abnormal indicators of the well to be measured in the current sampling period are determined based on the residual vector.

[0028] Optionally, S3 specifically includes:

[0029] The revenue function and cost function of the well to be logged are constructed based on the state feature vector of the well to be logged in the current sampling period, and the risk penalty function of the well to be logged is constructed based on the risk index of the well to be logged in the current period.

[0030] The return function of the well to be logged in the current sampling period is constructed based on the return function, the cost function, and the risk penalty function.

[0031] Optionally, a return function for the well to be logged in the current sampling period is constructed based on the return function, the cost function, and the risk penalty function, specifically as follows:

[0032] Calculate the sum of the cost function and the risk penalty function, and use the difference between the return function and the sum as the return function of the well to be tested in the current sampling period.

[0033] Optionally, the step S4, controlling the injection and production operation of the well to be logged in the next sampling cycle according to the optimal control strategy, specifically includes:

[0034] Based on the optimal control strategy, action control commands are generated for the injection and production equipment of the well to be logged.

[0035] The injection and production equipment of the well to be tested is controlled according to the action control command in the next sampling cycle.

[0036] Optionally, after S2, the method further includes:

[0037] Execute corresponding early warning strategies based on the risk index.

[0038] The beneficial effects that this invention can produce include:

[0039] The injection-production linkage control method based on well-to-surface multi-domain data provided by this invention determines the risk index by using a state feature vector that integrates downhole sensing data and surface sensing data, and then constructs a reward function to solve the optimal control strategy of the well to be monitored in the next sampling period. This state-action-reward model realizes true closed-loop autonomous optimization control.

[0040] This invention provides a well-to-surface multi-domain data-based injection-production linkage control method that integrates in-well and surface sensors to construct a multi-source data fusion model, enabling joint monitoring of downhole operating conditions and surface equipment status. By synchronously acquiring and normalizing heterogeneous data such as pressure, temperature, and flow rate, this invention forms a feature vector that comprehensively characterizes the well-to-surface system status, achieving holistic perception of reservoir dynamics and surface response. This multi-domain fusion overcomes the limitations of traditional single-point monitoring and improves the accuracy of anomaly identification.

[0041] The injection-production linkage control method based on well-to-ground multi-domain data provided by this invention deploys the early warning algorithm on field edge computing nodes. Using general-purpose computing devices, the data stream can be processed within milliseconds to seconds, enabling real-time risk detection and early warning. The early warning algorithm does not rely on a specific deep learning network structure but employs general data processing and anomaly identification methods, such as sliding window statistical analysis and multivariate anomaly index calculation. This significantly reduces the algorithm's requirements for computing resources and data scale. Edge deployment eliminates cloud communication latency, enabling autonomous monitoring and early warning even under poor network conditions, ensuring field safety.

[0042] This invention provides an injection-production linkage control method based on well-to-ground multi-domain data. It incorporates an intelligent identification mechanism for oil and gas well disturbances, promptly identifying abnormal conditions such as sudden changes in wellbore pressure drop or abnormal increases in injection pressure by comparing deviations between multi-source real-time data and normal patterns. A disturbance risk function is established to quantitatively assess detected anomalies, outputting a risk index to indicate the degree and trend of the anomaly. This risk assessment comprehensively considers the deviations of multiple sensors, making early warning more sensitive and reliable, and providing a basis for subsequent control decisions. This method differs from existing methods that rely solely on single-parameter over-limit alarms, enabling the identification of complex correlated disturbances and quantitative description of risk magnitude.

[0043] This invention provides an injection-production linkage control method based on well-to-surface multi-domain data. For detected risk states, it employs a reinforcement learning (RL) strategy to generate real-time control decisions for injection-production linkage. By constructing a state-action-reward model, with the reservoir-wellbore-surface system state as the environmental state and actions such as water injection adjustment and well production parameter adjustment, a comprehensive reward function incorporating both production and risk is designed, enabling the agent to continuously learn and obtain the optimal control strategy. This strategy can automatically adjust injection and production parameters (such as reducing water injection and optimizing production pressure differential) to mitigate anomalies when risks occur, and optimize the injection-production ratio to improve recovery during stable phases. Compared to traditional fixed control schemes, the reinforcement learning strategy possesses adaptability and predictive capabilities, enabling optimal responses under unknown conditions and achieving true closed-loop autonomous optimization control. The RL control strategy of this invention does not rely on a specific hardware platform and can run inference or be updated online on general-purpose edge computing nodes, exhibiting strong deployment flexibility. Attached Figure Description

[0044] Figure 1 A flowchart of the injection-production linkage control method based on well-to-ground multi-domain data provided in this embodiment of the invention;

[0045] Figure 2 This is a schematic diagram of the overall structure of injection-production linkage control based on well-to-ground multi-domain data, provided for an embodiment of the present invention. Detailed Implementation

[0046] The present invention will now be described in detail with reference to the embodiments, but the present invention is not limited to these embodiments.

[0047] This invention provides a method for injection-production linkage control based on well-to-ground multi-domain data, such as... Figure 1 and Figure 2 As shown, the method includes:

[0048] S1. Collect downhole sensing data and surface sensing data of the well to be logged in each sampling period, and determine the state feature vector of the well to be logged in the corresponding sampling period based on the downhole sensing data and surface sensing data.

[0049] The system of this invention includes downhole sensing devices, surface sensing devices, edge computing nodes, and injection-production control execution units. Downhole sensing devices refer to various sensors installed downhole, such as pressure gauges, thermometers, and multiphase flow meters, which can collect downhole fluid pressure data in real time. ,temperature Parameter data; ground sensing equipment includes wellhead pressure sensors, water injection pump flow meters, oil well production metering devices, and environmental monitoring sensors, used to acquire ground pipeline pressure. Injected traffic Extracted flow rate In addition to data on vibrations in the surrounding environment, these sensors are connected to edge computing nodes at the well site via wired or wireless means. The edge computing nodes have built-in data acquisition modules and are configured with a uniform sampling period. (For example, 1 second), synchronously read data from all sensors, and perform time alignment and buffering on data of different frequencies. After processing by the acquisition module, at each sampling moment... Forming downhole sensing data and ground sensing data ,For example:

[0050] ,in For a moment The bottom hole pressure, The value is the bottom hole temperature, and "..." indicates other downhole sensing parameters. ,in For wellhead pressure, To inject traffic, For the sampled flow rate, “…” indicates other ground-based sensor parameters.

[0051] S1 determines the state feature vector of the well to be logged in the corresponding sampling period based on downhole sensing data and surface sensing data, specifically including:

[0052] The downhole sensing data and the surface sensing data are preprocessed separately to obtain downhole data vectors and surface data vectors;

[0053] The downhole data vector and the surface data vector are spliced ​​together to form the state feature vector of the well to be measured in the corresponding sampling period.

[0054] Furthermore, the downhole data vector and the surface data vector are concatenated to form the state feature vector of the well to be logged in the corresponding sampling period, specifically:

[0055] Determine derived vectors based on downhole data vectors and surface data vectors;

[0056] The downhole data vector, surface data vector, and derived vector are sequentially concatenated to form the state feature vector of the well to be logged in the corresponding sampling period.

[0057] The above and Together, they constitute the multi-domain sensing data of this invention. In each sampling period, the data preprocessing module of the edge computing node performs denoising and standardization processing on the collected raw data. For example, a sliding window mean filter can be used to smooth noise: assuming the original bottom-hole pressure is... The pressure after smoothing Can be used according to the formula Calculate (where the window length is) For data with different dimensions, a normalization transformation is used to convert each parameter into a dimensionless form for fusion analysis. For example, a linear normalization method is used: for any sensor data... Its normalized value ,in and This represents the empirical minimum / maximum range of this parameter. After cleaning and normalization, the downhole data vector and the surface data vector are obtained.

[0058] After data preprocessing, the data fusion module of the edge computing node performs data fusion on the downhole... With the ground The data is fused to construct a comprehensive feature vector, i.e., a state feature vector, that can characterize the current operating state of the well-to-surface system. Specifically, the data vectors can be concatenated along their dimensions to form:

[0059] ;

[0060] In the above The first half of the vector data is downhole data (such as bottomhole pressure and temperature), and the second half is surface data (such as wellhead pressure and injection / production flow rates). At any given time... of It fully reflects the coupling state between the downhole well conditions and the operating conditions of the surface equipment. For example, when excessive water is injected, It may appear in the middle rise and Abnormal fluctuations are characterized by a combination of features; for example, when excessively rapid heat extraction leads to conical advance, it may manifest as... First rise then fall Features such as pressure drop can be extracted. To further extract these cross-domain correlation features, derived vectors can be calculated in the fusion module. For example, the pressure drop difference can be calculated. To characterize wellbore pressure loss, or to calculate the rate of change in water cut. And so on. These derived vectors serve as... The extended components help characterize subtle changes in well-to-surface dynamics. Finally, the fusion module outputs the current system's state feature vector. This is for use in subsequent algorithms. It should be noted that... The dimensions and content can be flexibly adjusted according to specific applications, and this invention is not limited to the parameters listed above.

[0061] S2. Determine the risk index of the well to be logged in the current sampling period based on the state feature vector of the well in different sampling periods.

[0062] Specifically, it includes:

[0063] (1) Determine the reference state vector based on the state feature vector of the well to be measured in the historical sampling period.

[0064] Specifically, this means: placing the well to be measured first... The average vector of the state feature vectors over several sampling periods is used as the reference state vector; where... It is an integer greater than or equal to 2.

[0065] Obtain in each sampling period Subsequently, the edge computing nodes immediately perform anomaly detection to determine whether the well-to-ground system has entered a disturbed state. The anomaly detection module pre-establishes a baseline model of the system state under normal operating conditions. When the system is running smoothly, the state feature vector... It should meet certain patterns or rules. One implementation method is to calculate the reference state vector using data within a sliding time window. As a moment The expected normal state. For example, it can be made For the past One cycle Moving average of values: .Should This reflects the recent normal trend.

[0066] (2) Determine the abnormal indicators of the well to be measured in the current sampling period based on the reference state vector and the state feature vector of the well to be measured in the current sampling period.

[0067] Specifically, this includes: first, calculating the residual vector between the state feature vector of the well to be logged in the current sampling period and the reference state vector; and then determining the abnormal indicators of the well to be logged in the current sampling period based on the residual vector.

[0068] The currently observed state feature vector With reference state vector Perform the difference operation to obtain the residual vector. :

[0069] ;

[0070] in express The Middle The deviation of each characteristic component from the reference value for The dimension of . If there are no significant anomalies, the deviation of each component . It should be close to 0. Once any component deviates abruptly from the normal range, it indicates that the system has entered an abnormal disturbance state. To measure the overall degree of deviation, this invention introduces a multivariate anomaly index. For example, using the weighted sum of squares form:

[0071] ;

[0072] in For the first The standard deviation of each feature under normal operating conditions can be obtained through offline statistical analysis of historical data. The weights for this feature are used to reflect the relative importance of different parameters to the overall risk. Those skilled in the art can customize the specific weight values ​​for each feature. This is how it is defined. In fact, it is the deviation The weighted L2 norm, calculated after normalization based on the standard deviations of each dimension, can be understood as the normalized "distance" of the current state from the normal state. The larger, the more it means The more significant the overall deviation from the normal pattern, the more severe the anomaly. To facilitate judgment, an anomaly detection threshold can be set in the implementation. :when When an abnormal disturbance occurs, it is considered to have occurred. The anomaly detection method of this invention is not limited to the norm method described above; other methods can also be used, such as principal component analysis to extract the main deviation components, or probabilistic statistical calculations. Anomalies are detected using likelihood values. Some implementations may also incorporate expert rules (such as classifying anomalies as occurring when a specific parameter exceeds a certain limit) to improve detection reliability. The anomaly detection module outputs a Boolean variable. Indicates whether an anomaly was detected: express An anomaly exists if the threshold is exceeded. This indicates that the operating condition is normal.

[0073] (3) Determine the risk index of the well to be logged in the current sampling period based on the abnormal indicators of the well to be logged in the current sampling period.

[0074] When detected When an anomaly occurs, the system enters the early warning decision-making process. First, the severity of the anomaly is quantitatively assessed, and a risk index is calculated. The risk index is composed of abnormal indicators. The standardized score is obtained through mapping. For example, a 0-100 score scale can be used to assess risk. Mapping to a finite interval using a sigmoid function:

[0075] ;

[0076] in and Parameters set based on experience: express The baseline level (when) From time to time ), Control the steepness of the curve. Thus, when When I was very young, Approaching 0; when Increase more than hour, The rapid increase nearing 100 highlights the severity of the anomaly. Alternatively, a linear scaling mapping could be used, such as setting... ,in This is the predefined maximum reasonable deviation value.

[0077] Following S2, the method further includes: implementing a corresponding early warning strategy based on the risk index.

[0078] Calculated risk index Then, the system executes corresponding early warning strategies based on the risk level: if If the score exceeds the high-risk threshold (e.g., 80 points), a red alert will be immediately issued, notifying on-site personnel of a potential serious anomaly via audible and visual alarms or SMS messages; if If the score is at a moderate level (e.g., 50-80 points), a yellow alert will be issued, indicating that attention is needed; if... If the score is only slightly higher than normal (e.g., 20-50 points), it is issued as a blue alert or recorded for analysis. The entire anomaly detection and risk assessment process is completed locally on the edge computing node, and the delay from sensor sampling to issuing an alert can be controlled within 1 second, truly achieving "second-level early warning." This enables on-site personnel or subsequent automatic control to take measures in the early stages of an anomaly, preventing the accident from escalating.

[0079] The early warning system of this invention does not rely on complex models such as convolutional neural networks or recurrent neural networks, but instead uses the aforementioned general algorithm for rapid calculation, thus enabling efficient operation on edge devices. Furthermore, to reduce false positives and false negatives, further judgments can be made based on the duration and frequency of the anomaly: for example, requiring... Continuously exceeding the threshold Anomalies are only confirmed after a certain number of cycles, thus filtering out transient interference. In summary, the early warning module outputs a risk index. And early warning level signals. In particularly urgent high-risk situations, the system can also automatically trigger emergency well shutdown and other protective measures (such as disconnecting the water injection pump or closing the wellhead) to ensure safety.

[0080] It should be noted that this invention employs an anomaly identification method based on multivariate bias analysis, but is not limited to this. Alternative solutions can use machine learning or deep learning models for anomaly detection, such as training a one-class support vector machine (SVM) model to learn normal operating condition characteristics before detecting anomalies, or using the isolated forest algorithm to score anomalies in multidimensional data. Temporal prediction models (such as ARIMA or autoregressive networks) can also be introduced to predict the state at the next time step and calculate prediction errors for early warning. These different anomaly detection methods can be re-embedded according to the characteristics of the field data. Regardless of the algorithm used, the core objective is to identify abnormal disturbances in the well-to-ground system and output risk indicators. It is important to note that if a deep learning model is used, it should be lightweighted for deployment in edge computing environments.

[0081] S3. Construct the return function of the well to be logged in the current sampling period based on the state feature vector and risk index of the well to be logged in the current sampling period.

[0082] Specifically, it includes:

[0083] The revenue and cost functions of the well to be logged are constructed based on the state feature vector of the well to be logged in the current sampling period, and the risk penalty function of the well to be logged is constructed based on the risk index of the well to be logged in the current period.

[0084] Construct the return function of the well to be logged in the current sampling period based on the return function, cost function, and risk penalty function.

[0085] Specifically: calculate the sum of the cost function and the risk penalty function, and use the difference between the return function and the sum as the return function of the well to be logged in the current sampling period.

[0086] In addition to early warning prompts and human intervention, this invention introduces a reinforcement learning agent to achieve automatic linkage control and decision-making for the injection and production process. First, the well-to-surface system is abstracted as a reinforcement learning environment: state... This is the state feature vector obtained in the previous section, which reflects the current system dynamics; actions Defined as a set of control commands for injection and production equipment, including regulation of both the water injection and production sides. For example, for a water injection well, the action can be defined as adjusting the water injection pump frequency. Or adjust the opening of the water injection valve This corresponds to the increase or decrease in the actual injection volume; for production wells, the action can be defined as the adjustment of the production throttle valve opening. Or adjust the frequency of the manual lifting system Etc. To simplify the explanation, actions can be represented using continuous vectors: ,in This indicates an adjustment to the injection volume (positive values ​​increase the injection volume, negative values ​​decrease it). This indicates an adjustment to production (positive values ​​increase production, negative values ​​decrease it, for example, by opening / closing the wellhead throttle valve). Environmental conditions will change under the influence of these actions: the reservoir-wellbore-surface is a dynamic system. When injection or production parameters are adjusted, pressure, production, etc., will change over time, thus affecting the state at the next moment. This invention does not require explicitly listing the physical equations of the system, but rather uses a reinforcement learning algorithm to gradually approximate the optimal control strategy through interaction with the environment.

[0087] In reinforcement learning frameworks, it is necessary to design reward functions. This guides the intelligent agent to optimize control objectives. The control objectives of this invention include both increasing oil production and reducing anomaly risks. Therefore, the reward function can be used... Defined as the revenue from output minus the risk cost:

[0088] ;

[0089] in, and These are the current output flow and the injected flow (which are states). (one of the components) It is a (production) benefit function that reflects the positive benefits of increasing oil production from oil wells; It is a (water injection) cost function, representing the resource consumption or energy consumption cost brought about by the injected water volume; This is a risk penalty function that applies an additional negative penalty to cases with high risk indices. For illustration, a linear approximation can be used: , , ,in Let be the weighting factor. Then we have:

[0090] ;

[0091] In the above reward function, This can be understood as the economic value per unit of output. The cost or adverse effect factor per unit of water injection (e.g., excessive water injection may lead to increased energy consumption or damage to the reservoir structure). The penalty for the risk unit index should be much greater than To ensure safety is prioritized, this design ensures that the agent's reward for each step both encourages it to increase oil production and prompts it to avoid high-risk operations. Specifically, when the system is in an abnormal state ( When it is relatively high, Some actions will significantly reduce returns, and in order to obtain higher cumulative returns, agents will tend to take actions that reduce risk (such as reducing water injection or slowing down production rates to mitigate anomalies). Conversely, in a safe state ( When the output is close to zero, the agent will focus more on the output and revenue, and strive to improve production efficiency.

[0092] S4. Solve for the return function of the well to be logged in the current sampling period to obtain the optimal control strategy for the well to be logged in the next sampling period, and control the injection and production actions of the injection and production equipment of the well to be logged in the next sampling period according to the optimal control strategy.

[0093] Specifically, controlling the injection and production actions of the injection and production equipment of the well to be logged in the next sampling period according to the optimal control strategy includes: first, generating action control commands for the injection and production equipment of the well to be logged in the next sampling period according to the optimal control strategy; and then controlling the injection and production actions of the injection and production equipment of the well to be logged in the next sampling period according to the action control commands.

[0094] Once the state, action, and reward functions are determined, a reinforcement learning algorithm can be selected to solve for the optimal control policy. The algorithms in this invention are not limited to a specific type; value iteration algorithms (such as Q-learning, Deep Q-Network (DQN), policy gradient algorithms (such as REINFORCE, DDPG), actor-critic algorithms, etc., can be used as needed. Describes the state in any given state The next step is to determine the probability distribution or deterministic mapping of which action to choose. The learning objective is to maximize the cumulative discounted reward, i.e., to maximize the expected goal. ,in This is a discount factor used to balance the importance of current and future returns. Reinforcement learning updates the policy through repeated trial and error, gradually approaching the optimal level. The training process can be conducted offline using a reservoir numerical simulation environment: a simulation environment including an injection-production physical model and anomaly mechanisms is built, allowing the agent to interact and learn. Once training converges, the resulting policy is... The strategy is permanently deployed in edge computing nodes to enable real-time decision-making. The strategy can be represented as a parameterized function, such as a neural network or a table. During real-time execution, the edge node adjusts its strategy based on the current state in each sampling period. By strategy Calculate the action For example, a deterministic strategy can be represented as For strategies involving exploration, probability can be used. Actions are obtained through sampling. Since the computational load of the trained policy inference process is very small (only one forward computation or table lookup is required), it can be completed in milliseconds on a general-purpose CPU, making it suitable for field deployment.

[0095] Actions output by edge computing nodes based on policies Specific action control commands are generated and sent to the injection-production control execution unit. The execution unit includes a water injection control module and an oil production control module, which can be integrated into the oilfield automation control system. For example, the water injection control module receives... After receiving the command, the actual injection volume is adjusted to the new set value by controlling the speed of the water injection pump or adjusting the opening of the water injection valve through frequency conversion drive. Oil production control module receives Following the command, the wellhead valves or pumping unit frequency of the oil well are controlled accordingly to change the production flow rate. To ensure coordinated operation, the execution unit can simultaneously issue water injection and oil production adjustment commands, or adjust them sequentially according to predetermined priorities. During the adjustment process, the actual state of the system gradually changes, and these changes are fed back to the next cycle through sensors. In this process, closed-loop control is achieved. If an abnormal deterioration occurs (e.g.) If the risk is still rising, the strategy will further adjust the intensity of actions to mitigate the risk; if the risk is mitigated and production is insufficient, the strategy may increase injection and production to increase output. The entire control operates autonomously at the edge node without human intervention, thus forming a "perception-decision-execution" closed loop within seconds. For example, in one application scenario, the system detects a sudden increase in wellhead pressure in the injection well accompanied by an increase in water cut in the production well (suspected risk of sudden crossflow). The edge intelligence immediately triggers a yellow warning and automatically reduces the injection volume by 10%, while slightly closing the production well valve to reduce the pressure difference. After several cycles of adjustment, the risk of pressure drop is eliminated, and the system then slowly restores injection to the optimized level. In this process, the time from the occurrence of the anomaly to the control response is only a few seconds, which is far superior to the delay of several minutes or more for manual operation, fully demonstrating the advantages of the second-level early warning and linkage control of this invention.

[0096] Although the system of this invention can operate autonomously at the edge, it still provides a human-machine interface for operators to monitor and intervene. Edge computing nodes can process key data (sensor readings, risk indices, etc.). The system uploads data (such as current control actions) to a host computer or cloud platform via industrial Ethernet or IoT modules for remote monitoring. Early warning information can also be simultaneously sent to the control room. In emergencies, operators can start / stop the water pump or switch to manual mode with a single click through the interface. If the network is good, the edge system can also collaborate with the cloud-based intelligent decision-making system, for example, periodically receiving optimization strategy parameters from the cloud or uploading historical data for cloud model training. However, it is important to emphasize that even during network outages or isolated operation, the edge system can still independently complete early warning and control tasks, ensuring the continuity of safe production on-site.

[0097] In summary, based on the above steps, this invention achieves the fusion processing of multi-source downhole and surface data, second-level intelligent early warning of abnormal disturbances, and reinforcement learning-driven linkage control of injection and production parameters. This method fully utilizes the low latency and high reliability of edge computing, bringing AI algorithms to the field front-end to achieve a real-time closed loop from perception to control, providing a novel technical solution for the digitalization of oil and gas fields.

[0098] It should be noted that the types of sensors and control devices mentioned in this invention are not limited. In specific implementations, sensing and control methods can be added or removed according to site conditions. For example, microseismic monitoring instruments and fiber optic distributed temperature / acoustic sensors can be added downhole to obtain information on reservoir fracture dynamics and wellbore integrity, further enriching the data source. On the surface, wellhead torque sensors and pump power monitoring devices can be integrated to enhance the monitoring of equipment operating conditions. In terms of control execution, if the oil well adopts intermittent pumping or gas lift processes, corresponding parameters (such as gas lift gas volume adjustment) can be added to the action set. Even if sensors and actuators are replaced or added, similar early warning and control effects can be achieved as long as the data fusion and intelligent decision-making approach of this invention is followed.

[0099] While this invention recommends using reinforcement learning for injection-production linkage control, other intelligent optimization strategies can be considered in some applications. For example, Model Predictive Control (MPC) algorithms can be used: based on the established reservoir-pipeline dynamic model, the injection-production trajectory for a future period is continuously optimized in real time, achieving an optimization control effect similar to reinforcement learning. Furthermore, for scenarios with simple well conditions, fuzzy control or expert rule systems can be used instead of RL strategies, adjusting water injection and production parameters according to preset rules. These methods differ in implementation, but if they achieve linkage control and optimization, they are also within the scope of this invention. Moreover, even when using reinforcement learning, the specific algorithm and network structure can vary. For example, using deep neural network approximation strategies (DQN, DDPG, etc.) or directly based on tabular Q-learning without using neural networks are all equivalent implementation methods in this invention.

[0100] This description primarily focuses on the coordinated control of a single water injection well and a single production well, but the ideas of this invention can be extended to the collaborative optimization of multi-well systems. In complex reservoirs where multiple water injection and production wells influence each other, this invention can introduce multi-agent reinforcement learning, with each well corresponding to one agent, sharing partial states (such as regional pressure) and achieving overall optimization through cooperative strategies. Alternatively, a single agent can be constructed to control multi-dimensional actions (vectors). This includes adjustments across multiple wells, enabling coordinated control at the injection-production network level. This extension requires state feature vectors. Includes more well parameters, actions It also employs higher-dimensionality methods, but the principle remains the same as for single-well applications. Therefore, this invention is also applicable to the optimization of injection and production parameters and the joint prevention and control of anomalies within oilfields.

[0101] While this invention emphasizes autonomous edge computing, a "device + edge + cloud" collaborative architecture can be adopted as a supplementary solution in some scenarios. For example, while edge nodes complete second-level early warning and control, they upload data to the cloud, utilizing the cloud's powerful computing resources for deeper analysis (such as fine reservoir modeling or long-cycle production optimization), and then feed the optimization suggestions back to the edge agent to adjust strategy parameters. This forms a hierarchical and progressive intelligent system. Alternatively, for platforms with good data transmission conditions, some computing can be deployed in small data centers near the well sites (considered a type of edge computing), allowing multiple well sites to share computing power and reduce the cost of equipment per well. These changes in deployment methods do not affect the core essence of the invention but provide flexible options.

[0102] Although this invention primarily focuses on the water injection-oil production process in oilfields, its technical concepts are universal and can be extended to other fields requiring multi-source monitoring and coordinated control. For example, in geothermal energy development, where water injection wells and production wells operate in tandem, this system can be used to achieve early warning of downhole temperature / pressure anomalies and optimized control of injection and production flow rates. In the gas injection and production processes of underground gas storage facilities, and in reservoir dam seepage monitoring and drainage control, the edge AI early warning and reinforcement learning control mechanisms of this invention can be applied to achieve safe and efficient management of complex coupled systems. By adapting to specific industry parameters, the principles and methods of this invention will produce similar beneficial effects in the aforementioned fields.

[0103] To verify the practicality and innovation of this invention, a specific implementation case is given here, taking a geothermal demonstration well (numbered DHW-01) as an example.

[0104] 1. Background and Deployment.

[0105] Reservoir type: shallow geothermal sandstone layer with hot water, initial formation temperature 120 °C, porosity 0.18.

[0106] Production well and water injection well: A water injection well (numbered INJ-01) is arranged next to the DHW-01 production well, with a distance of 800m between the two wells.

[0107] Equipment deployment:

[0108] (1) Downhole sensing: Install pressure sensors (measurement range 0 MPa~50 MPa, accuracy ±0.05 MPa) and temperature sensors (range 0°C~200°C, accuracy ±0.1 °C) at the bottom of the production well.

[0109] (2) Ground sensing: water injection pump flow meter (0 m³ / h~300 m³ / h, accuracy ±0.5%), production well output flow meter (same specification), wellhead pressure sensor (0 MPa~60 MPa, accuracy ±0.05 MPa).

[0110] (3) Edge computing node: Industrial control computer (Intel i5-10400, 8 GB RAM), running Linux, deploying the algorithm of this invention.

[0111] Sampling period: =1s, sliding window length set =5, moving average period =10.

[0112] 2. Implementation process.

[0113] Step 1: Data acquisition and preprocessing.

[0114] Synchronous reading of downhole data every second and ground And perform exponential smoothing: .

[0115] Step 2: State construction.

[0116] Constructing state feature vectors:

[0117] ;

[0118] in,

[0119] ;

[0120] .

[0121] Step 3: Anomaly detection.

[0122] Predicting production wellhead pressure using a two-order AR model:

[0123] ;

[0124] Among them, the coefficients 0.6 and 0.3 are empirical values, which can be obtained by fitting historical data in actual application.

[0125] Calculate residuals and standardized outlier indicators:

[0126] ;

[0127] ;

[0128] ;

[0129] like If the value is greater than 1, the pressure is considered abnormal.

[0130] Step 4: Risk assessment.

[0131] Summarize the abnormal indicators of production well pressure and injection flow rate, and assign a weight to production well pressure. The weight of the water injection flow rate is , + =1; then the risk index is 1. As shown below:

[0132] ;

[0133] in .

[0134] In the experiment, the risk index This can be represented as an abnormal indicator. Mapping function:

[0135] ;

[0136] Generally, a sigmoid function mapping can be used, that is:

[0137] ;

[0138] In practical applications, it is generally taken as .

[0139] when When the value exceeds 0.8, the system immediately issues a red alert, with an average response time of 0.9 seconds.

[0140] Step 5: Strengthen learning control.

[0141] state: ;

[0142] action: ;

[0143] in ;

[0144] Reward function:

[0145] ;

[0146] Algorithm: Uses tabular Q-learning, learning rate Discount factor .

[0147] Training: After 10,000 offline training sessions in a simulation environment, the training converges, and the policy is obtained. .

[0148] 3. Implementation results.

[0149] Early warning performance: average detection and alarm delay of 0.9 s (95th percentile <1.2 s), false alarm rate of 1.2%, and false alarm rate of 3.5%.

[0150] Production optimization: Two weeks after deployment, the average daily output of the production well increased from 480m³ to 517m³, an increase of approximately 7.7%; the energy consumption of the water injection pump decreased by approximately 4.2%.

[0151] Safety: The number of well kicks caused by abnormal operations during the period decreased from an average of 2 per month to 0.

[0152] This implementation case fully demonstrates the advantages of the present invention in terms of second-level early warning, linkage control, and production optimization.

[0153] In summary, this invention achieves significant beneficial effects based on existing technologies:

[0154] (1) Achieving second-level anomaly early warning to ensure production safety: By deploying intelligent algorithms at the edge of the well site, this invention can complete detection and alarm within 1 second after the appearance of abnormal signs, which is at least an order of magnitude faster than traditional manual monitoring. In the event of emergencies such as sudden changes in downhole pressure or water flooding, the system issues early warnings in a timely manner and automatically takes control measures to prevent the accident from escalating. This second-level response capability greatly improves the inherent safety level of oil and gas production and avoids the risks of blowouts and overflows caused by response delays in existing technologies.

[0155] (2) Downhole-Surface Multi-Source Fusion for High Global Sensing Accuracy: This invention integrates data from multiple downhole and surface sensors to construct a complete state feature vector, enabling the capture of subtle changes and cross-domain coupling effects in the injection and production system. For example, when excessive water injection threatens the safety of the production layer, it manifests as both abnormal wellhead pressure and changes in production parameters. The system can detect such correlated anomalies early through multi-source data fusion. Compared to existing single-point monitoring schemes, this invention improves the accuracy and robustness of anomaly detection and reduces false alarms and missed alarms. Multi-domain fusion makes early warning criteria more comprehensive and reliable, contributing to the long-term stability of the reservoir.

[0156] (3) Reinforcement Learning-Based Coordinated Control for Optimized Production and Risk Balance: This invention introduces a reinforcement learning strategy, overcoming the limitations of traditional water injection and oil production separation control, and achieving coordinated optimization and control of injection and production parameters. The reinforcement learning agent can autonomously decide to increase or decrease water injection and adjust oil production intensity based on reservoir dynamics and risk conditions, thereby maximizing recovery rate while ensuring safety. For example, in a test in an oilfield, the application of this invention's strategy increased the production of the test well group by approximately 8% while maintaining stable water cut. Compared with manual or pre-programmed control, the reinforcement learning strategy has adaptability and learning capabilities, continuously optimizing the injection and production scheme to cope with changes in reservoir characteristics at different stages. This coordinated control improves the effectiveness of water drive development, extends the stable production cycle of oil wells, and provides an effective means for tapping the potential of old oilfields.

[0157] (4) Edge-based intelligent autonomous operation, reducing latency and dependence: This system completes all data processing and decision control at the field edge nodes, and its normal operation does not rely on remote communication and cloud computing resources. On the one hand, it eliminates the risk of network latency and interruption in data transmission, enabling the system to operate reliably even in poor network environments; on the other hand, it reduces dependence on the central computing platform, saving bandwidth and cloud computing costs. This distributed autonomous architecture is particularly suitable for oilfield field environments, enabling local intelligent management of hundreds or thousands of wells. The edge controller of each well can optimize a single well specifically, and can also coordinate with surrounding wells through the network, enabling distributed intelligent control of the entire oilfield. In contrast, traditional cloud-based centralized monitoring is difficult to process massive amounts of data in a timely manner under large-scale well networks and has the potential for single-point failures. This invention effectively solves this problem.

[0158] (5) The algorithm is highly versatile and flexible in deployment and maintenance: The anomaly detection and reinforcement learning algorithm used in this invention does not rely on specific deep learning structures and dedicated hardware. All steps can be executed on general-purpose CPUs or industrial control computers. This means that the system can be easily deployed to existing oilfield automation equipment with good compatibility. In addition, since the algorithm model is relatively lightweight, on-site updates and parameter tuning are also relatively easy, without the need for a large amount of data and computing power as in training deep neural networks. The modular design of the system decouples the functional units, allowing maintenance personnel to adjust specific functions (such as early warning thresholds or strategy parameters) without requiring a complete shutdown or redesign of the entire system. This greatly reduces maintenance costs and improves the reliability and sustainable operation capability of the system.

[0159] (6) Innovative Integration of Multiple Technologies to Enhance Oilfield Digitalization: This invention organically combines advanced technologies such as edge computing, reinforcement learning, adaptive control, and multi-source data fusion to form a complete intelligent control system, representing an innovative achievement in the digital transformation of oil and gas production. Its creativity lies in: constructing a well-to-ground integrated multi-domain intelligent sensing system; proposing a novel disturbance risk quantification function; and designing a reinforcement learning control framework for injection-production linkage. These technical features are not disclosed in existing technologies, giving this invention a significant advantage in overall performance. By applying this invention, the automation and intelligence levels of oilfield production management can be significantly improved, reducing labor intensity and achieving the dual goals of quality improvement, efficiency enhancement, and inherent safety.

[0160] In summary, this invention is superior to existing technologies in terms of safety, production efficiency, real-time performance, and applicability, and has great practical application value and prospects for promotion.

[0161] The above description is merely a few embodiments of this application and is not intended to limit this application in any way. Although this application discloses preferred embodiments as described above, it is not intended to limit this application. Any changes or modifications made by those skilled in the art without departing from the scope of the technical solution of this application using the disclosed technical content are equivalent to equivalent implementation cases and fall within the scope of the technical solution.

Claims

1. A method for injection-production linkage control based on well-to-ground multi-domain data, characterized in that, The method includes: S1. Collect downhole sensing data and surface sensing data of the well to be logged in each sampling period, and determine the state feature vector of the well to be logged in the corresponding sampling period based on the downhole sensing data and the surface sensing data. S2. Determine the risk index of the well to be logged in the current sampling period based on the state feature vector of the well to be logged in different sampling periods; S3. Construct the return function of the well to be logged in the current sampling period based on the state feature vector and risk index of the well to be logged in the current sampling period; S4. Solve for the return function of the well to be logged in the current sampling period to obtain the optimal control strategy of the well to be logged in the next sampling period, and control the injection and production operation of the injection and production equipment of the well to be logged in the next sampling period according to the optimal control strategy.

2. The method according to claim 1, characterized in that, The step S1, determining the state feature vector of the well to be logged in the corresponding sampling period based on the downhole sensing data and the surface sensing data, specifically includes: The downhole sensing data and the surface sensing data are preprocessed to obtain downhole data vectors and surface data vectors, respectively. The downhole data vector and the surface data vector are spliced ​​together to form the state feature vector of the well to be measured in the corresponding sampling period.

3. The method according to claim 2, characterized in that, The downhole data vector and the surface data vector are concatenated to form the state feature vector of the well to be logged in the corresponding sampling period, specifically: Determine the derived vector based on the downhole data vector and the surface data vector; The downhole data vector, the surface data vector, and the derived vector are sequentially concatenated to form the state feature vector of the well to be measured in the corresponding sampling period.

4. The method according to claim 1, characterized in that, S2 specifically includes: A reference state vector is determined based on the state feature vector of the well to be logged in the historical sampling period; The abnormal indicators of the well to be logged in the current sampling period are determined based on the reference state vector and the state feature vector of the well to be logged in the current sampling period. The risk index of the well to be logged in the current sampling period is determined based on the abnormal indicators of the well to be logged in the current sampling period.

5. The method according to claim 4, characterized in that, The reference state vector is determined based on the state feature vector of the well to be logged during the historical sampling period, specifically as follows: The average vector of the state feature vectors of the well to be measured in the first M sampling periods is used as the reference state vector; where M is an integer greater than or equal to 2.

6. The method according to claim 4, characterized in that, The step of determining the abnormal indicators of the well to be logged in the current sampling period based on the reference state vector and the state feature vector of the well to be logged in the current sampling period specifically includes: Calculate the residual vector between the state feature vector of the well to be measured in the current sampling period and the reference state vector; The abnormal indicators of the well to be measured in the current sampling period are determined based on the residual vector.

7. The method according to claim 1, characterized in that, S3 specifically includes: The revenue function and cost function of the well to be logged are constructed based on the state feature vector of the well to be logged in the current sampling period, and the risk penalty function of the well to be logged is constructed based on the risk index of the well to be logged in the current period. The return function of the well to be logged in the current sampling period is constructed based on the return function, the cost function, and the risk penalty function.

8. The method according to claim 7, characterized in that, The return function of the well to be logged in the current sampling period is constructed based on the profit function, the cost function, and the risk penalty function, specifically as follows: Calculate the sum of the cost function and the risk penalty function, and use the difference between the return function and the sum as the return function of the well to be tested in the current sampling period.

9. The method according to claim 1, characterized in that, S4, which controls the injection and production operations of the well to be logged in the next sampling cycle according to the optimal control strategy, specifically includes: Based on the optimal control strategy, action control commands are generated for the injection and production equipment of the well to be logged. The injection and production equipment of the well to be tested is controlled according to the action control command in the next sampling cycle.

10. The method according to claim 1, characterized in that, Following S2, the method further includes: Execute corresponding early warning strategies based on the risk index.

Images