An oil well production prediction method based on artificial intelligence

Abstract

The application relates to the technical field of digital monitoring of oil and gas field development, and discloses an oil well production prediction method based on artificial intelligence, which comprises the following steps: collecting multi-source real-time sensor data and discrete manual intervention event data in the production process of an oil well; constructing an event vector containing an intervention time decay operator by using the collected discrete manual intervention event data; calculating a data quality score based on the collected multi-source real-time sensor data and performing physical consistency repair on missing data by using the data quality score; and injecting the constructed event vector and the repaired multi-source real-time sensor data into a state space model. The state space technical scheme driven by the event vector is adopted, the technical effect of real-time capturing of data mutation caused by working condition switching is achieved, explicit representation of segmented production rules is realized, and the defects of existing models, such as prediction drift and logical misjudgment when distribution mutation occurs, are solved.

Description

Technical Field

[0001] This invention relates to the field of digital monitoring technology for oil and gas field development, specifically to an artificial intelligence-based method for predicting oil well production. Background Technology

[0002] In the context of production dynamic monitoring of unconventional oil wells, due to the high frequency of manual intervention and low frequency and discontinuous data measurement conditions involved in the development process, how to achieve high-precision dynamic production prediction with uncertainty measurement in complex variable point disturbance and data missing environment is a highly challenging and novel technical problem in the current intelligent development of oil and gas fields.

[0003] Currently, networks based on long short-term memory or... An artificial intelligence model is used to construct a production prediction scheme. This scheme uses deep learning networks to mine time-series features in historical production data in order to establish a nonlinear mapping relationship between oil well physical parameters and production, thereby achieving trend prediction of future production.

[0004] In existing technologies, models are prone to prediction drift when dealing with sudden changes in data distribution caused by changes in measures or operating conditions. Lacking the ability to discriminate changes in data distribution, models may misinterpret segmented production patterns under different operating conditions as random noise or trend reversals, leading to drastic fluctuations in short-term predictions and serious deviations in medium- and long-term predictions. Furthermore, existing solutions typically lack explicit mechanisms for handling changes in data distribution, preventing the system from effectively warning of model failure risks. To address this, this invention proposes an artificial intelligence-based oil well production prediction method. By constructing event vectors to explicitly represent human intervention and injecting them into a state-space model for trend correction, the method resets the production state at turning points, resolving the prediction drift and pattern misinterpretation problems caused by sudden changes in distribution. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an artificial intelligence-based method for predicting oil well production, thereby resolving the problems mentioned in the background section.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an artificial intelligence-based oil well production prediction method, comprising:

[0007] Collect multi-source real-time sensor data and discrete human intervention event data during the oil well production process;

[0008] An event vector containing an intervention time decay operator is constructed using the collected discrete artificial intervention event data;

[0009] The system calculates a data quality score based on the collected multi-source real-time sensor data and uses the data quality score to perform physical consistency repair on missing data.

[0010] The constructed event vector and the repaired multi-source real-time sensor data are injected into the state space model, and closed-loop filtering is performed in combination with the data quality score to obtain the production potential.

[0011] By using physical constraint channels and data-driven channels to perform nonlinear mapping on production potential, a preliminary predicted production output is generated.

[0012] By using a variable-point adaptive order-preserving prediction algorithm, the size of the verification window is dynamically adjusted according to the variable-point probability triggered by the event vector to obtain the prediction range of output.

[0013] The prediction residuals, change point probabilities, and data quality scores generated from the initial production forecast are coupled together to calculate and output the prediction reliability indicator.

[0014] The pre-trained task transfer model is used to adaptively fine-tune the initial predicted output and prediction range to obtain the final prediction result.

[0015] Preferably, the In this context, the multi-source real-time sensor data includes wellhead pressure, bottom hole pressure, pumping unit load, and instantaneous discharge rate.

[0016] The acquisition process of the multi-source real-time sensor data includes: high-frequency sampling of the wellhead pressure, bottom hole pressure, pumping unit load, and instantaneous displacement through edge computing nodes deployed at the well site; identification of instantaneous outliers in the sampled data using a sliding window statistical method, and correction of the data quality score based on the distribution density of the instantaneous outliers; before the sampled data is transmitted to the cloud database, illegal observations exceeding the rated dynamometer range of the pumping unit are removed using a threshold filtering algorithm based on physical boundaries, ensuring that the initial observation sequence entering the state space model conforms to the physical baseline logic of oil and gas production.

[0017] Preferably, the In this process, the logic for constructing event vectors is as follows: the discrete human intervention event data is vectorized and encoded, and a time decay operator is introduced in combination with the intervention time to explicitly characterize the intensity of the continuous impact of data distribution changes.

[0018] The logic for constructing the event vector includes: matching the corresponding preset influence half-life according to the type of discrete artificial intervention event, and calculating the amplitude of the time decay operator using the preset influence half-life; the event vector constructs a nonlinear activation mapping in the hidden layer of the state space model, and performs asymmetric enhancement or suppression on the gain coefficient of the state transition function within a preset time window after the intervention time, thereby achieving explicit capture of the segmented production pattern.

[0019] Preferably, the In this study, the criteria for calculating the data quality score include the data missing rate, sensor drift rate, and physical-logic consistency residuals.

[0020] The physical consistency repair process includes: identifying continuous missing intervals in multi-source real-time sensor data, extracting the production evolution trend before and after the continuous missing intervals, generating a physical constraint completion sequence by combining the oil well production decline model, and performing weighted fusion of the completion sequence based on the data quality score.

[0021] Preferably, the In this context, the evolution formula for the state-space model is:

[0022] ,

[0023] ,

[0024] in, For production potential, For event vectors, These are external control parameters. This is the state transition function. For process noise, To observe the noise, To observe yield;

[0025] During the closed-loop filtering process, the observation noise The covariance matrix is ​​dynamically adjusted in real time based on the data quality score; when the data quality score drops below a preset value, the observation noise is increased. The covariance weights make the production potential The estimation result is transferred to the state transition function. The predicted trajectory is tilted.

[0026] Preferably, the In this process, the physical constraint channel introduces an oil well production decline model to impose numerical boundary constraints on the nonlinear output of the data-driven channel.

[0027] Preferably, the In this context, the logic for dynamically adjusting the size of the verification window is as follows: when the probability of the change point exceeds a preset threshold, the rolling verification window of the order-preserving prediction algorithm is reduced or reset, and the width of the prediction interval is increased.

[0028] The nonlinear mapping process between the physical constraint channel and the data-driven channel adopts a dynamic gating fusion mechanism. By calculating the evolution rate of the prediction residual, the mixing ratio of the output trend of the physical constraint channel and the output characteristics of the data-driven channel is adjusted in real time.

[0029] Preferably, the In this context, the formula for calculating the predicted reliability indicator is:

[0030] ,

[0031] in, To predict reliability indicators, To predict residuals, For the probability of a change point, Score the data quality. To predict interval coverage, For multi-criteria logic gate functions;

[0032] The predicted reliability indicator is divided into three numerical levels: available, cautious, and unavailable.

[0033] The multi-criteria logic gate function The execution logic includes: real-time monitoring of the abrupt change intensity of the production potential; when the rate of change of the derivative of the production potential is abnormal and accompanied by a decrease in the data quality score, forcibly setting the prediction reliability flag to the unavailable level and triggering a model failure warning signal.

[0034] Preferably, the In this process, the pre-trained task transfer model is obtained by performing a self-supervised learning task on the historical sample set to reconstruct consistency constraints based on missing values;

[0035] The adaptive fine-tuning process includes: using the feature extraction operator obtained from the self-supervised learning task to extract deep features from the repaired multi-source real-time sensor data; dynamically concatenating the extracted deep features with the production potential as the input vector of the task transfer model; and during the fine-tuning process, by comparing the consistency between the preliminary predicted production and the final predicted result, reversibly correcting the fusion weights of the physical constraint channel and the data-driven channel to eliminate representation bias caused by data discontinuity.

[0036] Preferably, when adjusting the size of the rolling verification window, the order-preserving prediction algorithm performs a real-time calibration process for the inconsistency score: by comparing the deviation of the preliminary predicted output from the quantile of the historical prediction interval, a real-time uncertainty index is calculated; when the change point probability triggers a change point warning, the real-time uncertainty index is used to drive the confidence upper and lower limits of the prediction interval to expand asymmetrically until the fluctuation range of the prediction residual returns to the preset steady-state interval, and then the normal size of the rolling verification window is restored.

[0037] This invention provides an artificial intelligence-based method for predicting oil well production. It offers the following advantages:

[0038] 1. This invention adopts a state-space technology solution based on event vector driving to achieve the technical effect of real-time capture of data mutations caused by working condition switching, realize explicit representation of segmented production rules, and solve the shortcomings of existing models in predictive drift and logical misjudgment when distribution changes.

[0039] 2. This invention adopts a variable-point adaptive order-preserving prediction technology to achieve the technical effect of dynamically calibrating the prediction interval width, realizing scientific coverage of prediction uncertainty during production fluctuations, and solving the shortcomings of traditional solutions that cannot provide effective fluctuation boundaries in the early stage of operating condition switching.

[0040] 3. This invention adopts a multi-dimensional quality assessment and physical consistency repair technology to achieve the technical effect of maintaining the stability of model representation when sensor data is discontinuous, realize the continuous evolution of the prediction main line, and solve the shortcomings of existing technologies in prediction interruption or distortion under severe measurement lack environment.

[0041] 4. This invention adopts a reliability quantification technology scheme with coupled multi-criteria logic gates to achieve the technical effect of classifying and labeling the prediction results, realizing the automatic identification of model failure risk, and solving the shortcomings of existing schemes that are blindly confident in the prediction conclusions under abnormal working conditions.

[0042] 5. This invention employs a dual-channel fusion technology solution combining physical constraints and data-driven approaches to achieve forced correction. The technical effect of the model's nonlinear output is to achieve a high degree of consistency between the predicted trend and the objective laws of oil and gas development, and to solve the shortcomings of pure data-driven model prediction results that violate physical common sense. Attached Figure Description

[0043] Figure 1 This is a schematic diagram of the process of the present invention. Detailed Implementation

[0044] To enable those skilled in the art to understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. Other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort should fall within the scope of protection of the present invention.

[0045] The present invention will now be described in detail with reference to the accompanying drawings:

[0046] Please see the appendix Figure 1 This invention provides an artificial intelligence-based method for predicting oil well production, comprising:

[0047] It collects multi-source real-time sensor data and discrete human intervention event data during the oil well production process; the multi-source real-time sensor data includes wellhead pressure, bottom hole pressure, pumping unit load, and instantaneous discharge rate;

[0048] The process of acquiring multi-source real-time sensor data includes: high-frequency sampling of wellhead pressure, bottom hole pressure, pumping unit load, and instantaneous discharge rate through edge computing nodes deployed at the well site; identification of instantaneous outliers in the sampled data using the sliding window statistical method, and data quality scoring based on the distribution density of the instantaneous outliers; and removal of illegal observations exceeding the rated dynamometer range of the pumping unit using a threshold filtering algorithm based on physical boundaries before the sampled data is transmitted to the cloud database, ensuring that the initial observation sequence entering the state space model conforms to the physical baseline logic of oil and gas production.

[0049] The event vectors containing the intervention time decay operator are constructed using the collected discrete artificial intervention event data. The logic for constructing the event vectors is as follows: the discrete artificial intervention event data are vectorized and encoded, and the time decay operator is introduced in combination with the intervention time to explicitly characterize the intensity of the continuous impact of data distribution changes.

[0050] The logic for constructing event vectors includes: matching the corresponding preset influence half-life according to the type of discrete human intervention event, and calculating the amplitude of the time decay operator using the preset influence half-life; the event vector constructs a nonlinear activation mapping in the hidden layer of the state space model, and performs asymmetric enhancement or suppression on the gain coefficient of the state transition function within a preset time window after the intervention time, thereby achieving explicit capture of the segmented production pattern.

[0051] The system calculates a data quality score based on the collected multi-source real-time sensor data and uses the data quality score to perform physical consistency repair on missing data. The basis for calculating the data quality score includes the data missing rate, sensor drift rate, and physical-logical consistency residual.

[0052] The physical consistency restoration process includes: identifying continuous missing intervals in multi-source real-time sensor data, extracting the production evolution trend before and after the continuous missing intervals, generating a physical constraint completion sequence by combining the oil well production decline model, and performing weighted fusion of the completion sequence based on the data quality score;

[0053] The constructed event vector and the repaired multi-source real-time sensor data are injected into the state space model, and closed-loop filtering is performed in combination with the data quality score to obtain the production potential.

[0054] The evolution formula for the state-space model is:

[0055] ,

[0056] ,2

[0057] in, For production potential, For event vectors, These are external control parameters. This is the state transition function. For process noise, To observe the noise, To observe yield;

[0058] During closed-loop filtering, observation noise The covariance matrix is ​​dynamically adjusted in real time based on the data quality score; when the data quality score drops below a preset value, the observation noise is increased. The covariance weights make the production potential The estimation results are transformed into the state transition function. The predicted trajectory is tilted;

[0059] The production potential is nonlinearly mapped using physical constraint channels and data-driven channels to output a preliminary production forecast; the physical constraint channel introduces an oil well production decline model to impose numerical boundary constraints on the nonlinear output of the data-driven channel.

[0060] The order-preserving prediction algorithm with adaptive change point is used to dynamically adjust the size of the verification window based on the probability of change points triggered by the event vector, so as to obtain the prediction range of output. The logic of dynamically adjusting the size of the verification window is as follows: when the probability of change points exceeds the preset threshold, the rolling verification window of the order-preserving prediction algorithm is reduced or reset to increase the width of the prediction range.

[0061] The nonlinear mapping process between the physical constraint channel and the data-driven channel adopts a dynamic gating fusion mechanism. By calculating the evolution rate of the predicted residual, the mixing ratio of the output trend of the physical constraint channel and the output characteristics of the data-driven channel is adjusted in real time.

[0062] When adjusting the size of the rolling verification window, the order-preserving prediction algorithm performs a real-time calibration process for the inconsistency score: by comparing the deviation of the preliminary predicted output from the quantile of the historical prediction interval, a real-time uncertainty index is calculated; when the probability of a change point triggers a change point warning, the real-time uncertainty index is used to drive the confidence upper and lower limits of the prediction interval to expand asymmetrically until the fluctuation range of the prediction residual returns to the preset steady-state interval, and then the normal size of the rolling verification window is restored.

[0063] The prediction residuals, change point probabilities, and data quality scores generated from the initial output forecast are coupled together to calculate and output a prediction reliability indicator; the formula for calculating the prediction reliability indicator is: ,

[0064] in, To predict reliability indicators, To predict residuals, For the probability of a change point, Score the data quality. To predict interval coverage, For multi-criteria logic gate functions;

[0065] The predicted reliability indicator is divided into three numerical levels: available, cautious, and unavailable.

[0066] Multi-criteria logic gate functions The execution logic includes: real-time monitoring of the abrupt change intensity of the production potential; when the rate of change of the derivative of the production potential is abnormal and accompanied by a decrease in the data quality score, the prediction reliability flag is forcibly set to the unavailable level and a model failure warning signal is triggered.

[0067] The pre-trained task transfer model is used to adaptively fine-tune the initial predicted output and prediction range to obtain the final prediction result.

[0068] The pre-trained task transfer model is obtained by performing a self-supervised learning task on a historical sample set to reconstruct consistency constraints based on missing values;

[0069] The adaptive fine-tuning process includes: using the feature extraction operator obtained from the self-supervised learning task to extract deep features from the repaired multi-source real-time sensor data; dynamically concatenating the extracted deep features with the yield potential as the input vector of the task transfer model; and during the fine-tuning process, by comparing the consistency between the preliminary predicted yield and the final predicted result, the fusion weights of the physical constraint channel and the data-driven channel are reversed to eliminate the representation bias caused by data discontinuity.

[0070] Multi-source real-time sensor data refers to high-frequency observation sequences continuously collected by various types of monitoring equipment in the oil and gas field well site environment, reflecting the dynamics of downhole and surface production. This type of data constitutes the most basic and direct state feedback input in the physical information model, used to monitor production status in real time and provide temporally continuous physical signal support.

[0071] Discrete human intervention event data refers to the specific time points and operation types of non-continuous human operations recorded by the system. This type of data marks the sudden changes in the state of the production system caused by external intervention and is an important indicator of the causes of output fluctuations and changes in operating patterns.

[0072] The intervention time decay operator is a numerical weighted function based on time distance, where the weights decrease over time. It is used to mathematically simulate the effect of artificial intervention on oil well production as the effect gradually weakens over time.

[0073] Event vectors are digital representations of discrete human intervention events with fixed dimensions and numerical features, which are transformed into digital representations through encoding. They can be embedded as structured inputs into prediction models, and the logic of human intervention can be explicitly introduced into data-driven prediction architectures.

[0074] Data quality scoring is a quantitative indicator system for real-time evaluation of sensor data. This scoring is used to identify low-quality data segments, guide subsequent data repair strategies, and assign differentiated trust weights to different data sources.

[0075] Physical consistency repair refers to the technical process of correcting and reconstructing missing, anomaly, or logical conflicts in collected data based on physical laws. The aim is to ensure that the data input into the model conforms to physical constraints and improve the reliability of the underlying data.

[0076] State-space models are a class of mathematical frameworks used to describe the time-varying implicit states of dynamic systems, typically consisting of state equations and observation equations. These models can infer the true state of the system from noisy observation data that has not been directly measured, thus revealing the underlying patterns in production dynamics.

[0077] Closed-loop filtering is an algorithm that integrates the latest observation data and recursively updates and corrects the system state estimates. Through real-time data assimilation, this process enables the model output to closely track changes in actual production dynamics, maintaining a high degree of synchronization between predictions and measurements.

[0078] Production potential refers to the long-term trend of oil well production capacity after eliminating short-term random fluctuations and measurement noise. It is the core basis for the model to predict future production and represents the production capacity baseline of the oil well under ideal conditions.

[0079] Physical constraint channels refer to computable branches embedded in artificial intelligence models based on seepage mechanics, reservoir engineering principles, or production decline theories. This module ensures that the model output strictly follows known physical laws and engineering experience, avoiding predictions that defy common sense.

[0080] The data-driven channel automatically mines complex nonlinear relationships and high-dimensional feature patterns from historical data. This component can capture hidden patterns that are not easily expressed by traditional analytical formulas, enhancing the model's adaptability to complex production scenarios.

[0081] The adaptive order-preserving prediction algorithm is an intelligent algorithm that automatically detects abrupt changes in data distribution, dynamically adjusts the learning window, and maintains the monotonicity of the predicted values. It is suitable for rapid response during changes in operating conditions and can converge to a new stable prediction boundary in a timely manner after a change occurs.

[0082] A forecast interval refers to the range of possible values ​​given based on a point forecast, used to quantify the uncertainty of future output. This interval reflects the model's estimate of volatility and forecast error, providing decision support under risk perception.

[0083] A prediction reliability indicator is a metric or label that quantifies the credibility of current prediction results. It can promptly issue early warning signals to production managers when the prediction performance of the model deteriorates due to data anomalies, distribution shifts, or external interference.

[0084] By comprehensively and systematically collecting real-time monitoring data from various sensors and historical records of human intervention events, a complete and clearly structured underlying database is constructed, providing multi-dimensional and multi-layered observation perspectives for subsequent in-depth analysis and intelligent decision-making. This approach ensures that the predictive system always grasps the latest physical dynamics while keenly sensing external disturbances introduced by human operation, fundamentally breaking down information barriers and eliminating data silos, and significantly enhancing the system's perception depth and coverage in complex and ever-changing production environments.

[0085] Constructing event vectors with fusion decay operators can transform fragmented manual operation records into structured logical features with temporal correlation, explicitly marking the occurrence time of work condition switching and the trend of its evolution over time. This effectively avoids the model misidentifying sudden changes in output caused by human factors as system noise, and significantly improves the model's accuracy in judging segmented production patterns and its ability to identify scenarios.

[0086] By using a data quality scoring mechanism to guide the physical consistency repair process, the high reliability and continuity of the data sequence input into the model can be ensured. By introducing a physical law-based imputation method, data gaps and distortions caused by sensor failures or metering interruptions can be corrected, thus building a solid, reliable data foundation that conforms to the dynamic characteristics of oil and gas field development for subsequent high-precision prediction tasks.

[0087] By injecting event vectors as external input into the state-space model and implementing closed-loop filtering, a system architecture capable of dynamically adjusting based on data quality is constructed. This mechanism adaptively adjusts the trust weights of observations when data quality fluctuates and forces a state reset upon detecting a change point, thereby completely eliminating prediction trajectory drift caused by abrupt changes in data distribution and ensuring the stability and consistency of yield potential representation.

[0088] A dual-channel mapping mechanism combining physical constraints and data-driven approaches is employed to achieve complementary advantages between mechanistic models and artificial intelligence algorithms. Physical boundaries are used to strongly constrain the output of the data-driven model, preventing deep learning networks from outputting extreme predictions that violate oil and gas development principles when sample bias exists. This ensures that preliminary production predictions, while leveraging the advantages of nonlinear fitting, also possess rigorous engineering logic.

[0089] By employing a strategy combining a change-point adaptive mechanism and an order-preserving prediction algorithm, the system can dynamically adjust the width of the verification window based on the change-point detection probability, enabling rapid response when production conditions undergo drastic changes. By proactively expanding the prediction interval to accommodate the uncertainties in the early stages of a change point, the system provides a flexible and inclusive risk measurement solution.

[0090] By integrating residual analysis, change point probability, and data quality scoring, the system can output prediction reliability indicators, constructing an automated failure early warning system. This system can classify and label prediction risks, clearly indicating the model's confidence level when data quality is poor or operating conditions change abruptly. This achieves a qualitative change in prediction output from blind judgment to risk-controllable assessment, significantly improving the system's safety and reliability in practical applications.

[0091] By utilizing a pre-trained task transfer model combined with an adaptive fine-tuning strategy, shared features from a large historical sample are introduced to alleviate the problem of sparsity in target well data. Simultaneously, a self-supervised learning mechanism is employed to further eliminate representational biases caused by data discontinuity, achieving refined calibration of the initial prediction results and ensuring that the final output predictions reach optimal levels in terms of accuracy, stability, and generalization ability.

[0092] Example 1: Scenario of yield prediction for data distribution mutations caused by human intervention

[0093] This embodiment focuses on verifying the technical capability of the present invention in solving the problem of production forecast drift when dealing with sudden changes in data distribution caused by human operation.

[0094] The specific scenario is that after a tight oil well undergoes a high-pressure water injection diversion measure, its original production pattern changes significantly, and the system needs to monitor and adapt to this change in real time.

[0095] Execution steps and procedures:

[0096] The system collects dynamic data on wellhead pressure and instantaneous discharge in real time by deploying edge computing nodes at the well site, and simultaneously receives operating condition switching command signals from the production management system, thereby achieving high-frequency synchronization and fusion of multi-source heterogeneous data.

[0097] exist During the data preprocessing stage, the system performs unique thermal encoding on the specific event of high-pressure water replenishment shift, transforming it into an event feature vector recognizable by the model. Simultaneously, based on the well's historical production response feature library, the system automatically matches an exponentially decaying response operator that conforms to the law of underground physical energy release. This operator can effectively quantify the duration and intensity of the disturbance to the bottom seepage field caused by human intervention, ensuring that the model obtains maximum gain correction immediately upon intervention, and enhancing its sensitivity to abrupt changes.

[0098] exist During model inference, event feature vectors are dynamically injected into the hidden layer units of the state-space model. The core innovation of this invention lies in the construction of an adaptive nonlinear activation gating mechanism in the hidden layer. When an event vector is injected, this gating mechanism forces an asymmetric enhancement and partial reset of the weight coefficients in the state transition matrix. This means that the model acknowledges, at both the data and physical mechanism levels, that a phased break in the production pattern has occurred. By actively resetting the initial conditions of the state transition function, the production potential can quickly cross the change point, accurately capturing the stable production benchmark under the new operating conditions. This method completely solves the problems of traditional... The predictive model suffers from significant prediction lag at turning points due to its slow crawling based on historical information.

[0099] exist After a mutation occurs, the system calculates the probability value of the change point in real time. Once this probability instantaneously crosses a preset threshold, a window reset mechanism is triggered. At this point, the order-preserving prediction algorithm immediately stops referencing historical long-window data and instead executes a dynamic window reset strategy, focusing the modeling attention entirely on the newly generated production data sequence after the intervention. By calibrating the inconsistency scores in the sequence in real time, the prediction interval expands explosively at the change point to cover extreme uncertainties. Subsequently, due to the continuous accumulation of data under the new operating conditions, the prediction interval gradually converges to a reasonable range. This process ensures that, even when production patterns are not yet fully stable, the output probability distribution provided by the system can broadly cover possible extreme fluctuations, significantly improving the prediction robustness and decision reliability under mutation scenarios.

[0100] Example 2: In-depth verification of scenarios with severe data loss and discontinuity caused by sensor failure.

[0101] This embodiment focuses on verifying the comprehensive technical capability of the present invention to solve the problem of unreliable output under extreme data loss conditions.

[0102] The scenario is set in a remote oil and gas operation area where extreme weather causes a long-term interruption of sensor signals, resulting in a large-scale loss and discontinuity of critical production data.

[0103] Execution steps and procedures:

[0104] exist In this system, the input data undergoes multi-dimensional evaluation, and a real-time quality score is calculated based on the data missing rate and physical-logical consistency residuals. When the sensor data stream is interrupted, the system automatically initiates a physical repair mechanism, calling the relevant data for that well. Decreasing model historical parameters. The core innovation lies in the fact that the system does not use conventional linear interpolation methods, but instead uses the bottom hole pressure evolution trend at the last moment before the missing data. By inverting the actual production dynamic laws, it generates a complete sequence that highly matches the reservoir dynamic evolution characteristics in terms of physical mechanism. Furthermore, it uses real-time quality scores as confidence weights to adaptively smooth and fuse the reconstruction results, thereby significantly improving data continuity and reliability.

[0105] exist During the closed-loop filtering process, the system can dynamically adjust the observation noise based on the data quality. The covariance matrix. When data quality is extremely low or even completely missing, the system adaptively increases the weight of observation noise, mathematically achieving a soft masking of unreliable real-time observations, thus prompting the prediction engine to rely on the state transition function. The physical evolution trajectory described is calculated. This mechanism effectively ensures that even in environments without sensor data input for extended periods, the system's predicted main line can continue to evolve along a trajectory that conforms to physical laws, preventing the prediction process from failing due to data interruption.

[0106] exist In this system, multiple indicators are integrated, including the cumulative effect of prediction residuals, the data quality score at the current moment, and the coverage of the prediction interval. When the system detects that the prediction residuals exceed a preset warning threshold due to long periods of missing data, and the data quality score remains low, the built-in multi-criteria logic gate function... The system will activate a mandatory protection mechanism, automatically setting the output reliability flag to an unavailable state. This provides numerical-level judgment and embodies a proactive safety strategy that can effectively prevent downstream automated production systems from performing erroneous operations based on unreliable predictions, thus ensuring the overall stability and safety of the production system.

[0107] Example 3: Consistency Verification and Correction Scenario of Physical Constraints and Data-Driven Dual-Channel Integration

[0108] This embodiment focuses on examining how the present invention achieves forced correction of nonlinear output through mechanism fusion under complex real-world conditions, and how it promotes... The predicted technological effect of returning to basic physical principles can fundamentally alleviate the problems caused by... The problem of prediction bias and output distortion caused by the "black box" property of the model.

[0109] Execution steps and procedures:

[0110] exist In this system, subtle nonlinear features and correlation patterns are extracted from high-frequency load and pressure signals via a data-driven channel; simultaneously, a physical constraint channel applies the actual inflow dynamic curve of the oil well as boundary conditions in real time. The innovation lies in the introduction of a dynamic gating fusion mechanism, which can monitor the derivative change trend of the output value of the data-driven channel in real time. If the output shows a clear violation of the physical laws of oil production, the gating weight will automatically tilt towards the physical constraint channel, and the predicted output will be forcibly constrained within a reasonable range by means of physical boundary conditions, so as to ensure the engineering rationality of the output results.

[0111] exist In this system, a transfer model pre-trained on millions of historical samples is used. This model, through a self-supervised learning method that reconstructs consistency constraints based on missing values, has learned a general feature representation of multiphase flow in oil and gas. In the adaptive fine-tuning stage, the fine-tuning operator performs deep feature extraction on the repaired multi-source sensor data, and... The generated production potential forecasts are dynamically stitched and compared. By analyzing the consistency between the initial forecasts and the fine-tuned results, the system can reverse-optimize the fusion ratio between the physical model and the data-driven approach, effectively eliminating representational biases caused by data discontinuity and improving the reliability and robustness of the output.

[0112] The system's final output prediction results not only accurately capture minute fluctuations in sensor signals but also strictly adhere to the principles of material balance and reservoir engineering at the macroscopic dynamic evolution level. By applying this two-way constraint mechanism, the system ensures that the production prediction curve always operates within the safe corridor defined by reservoir physical laws under multiple disturbances such as complex liquid level fluctuations and pressure jumps, achieving a deep unity and synergy between data-driven high-precision nonlinear fitting and rigorous engineering logic.

[0113] Example 1: By introducing an intervention-time decay operator and a hidden layer nonlinear activation mapping mechanism, the prediction model shifts from passively fitting the data distribution to actively identifying and responding to breakpoints in the data. Through dynamically resetting the initial state space values ​​and combining this with an asymmetric expansion of the prediction interval, prediction drift is eliminated at points where physical laws fundamentally change. This mechanism significantly enhances the model's accuracy in capturing dynamic trajectories and greatly improves its risk coverage under extreme disturbances, ensuring stable and reliable prediction performance even when sudden changes occur in the production process.

[0114] Example 2: This scheme effectively fills data blind spots caused by observation interruptions by utilizing physical consistency repair technology. Simultaneously, it achieves smooth and intelligent migration of prediction weights under different information sources by adjusting the noise covariance parameter in the closed-loop filter in real time. Experiments demonstrate that even under the unfavorable condition of lacking real-time observation feedback, the system can still maintain the coherent evolution of prediction logic based on built-in physical mechanism priors, and dynamically output risk warning signals using reliability indicators. This fundamentally transforms prediction conclusions from blindly outputting point estimates to quantitative expressions with clearly measurable confidence intervals.

[0115] Example 3: By introducing a dynamic gating fusion mechanism and a self-supervised task transfer model, the powerful feature mining capabilities of artificial intelligence are successfully constrained within a safe corridor defined by physical laws. This architecture effectively suppresses the inverse physical trend prediction distortion that may occur in pure data-driven models when samples are biased, and significantly improves the accuracy of the model's state representation under sparse data conditions by leveraging the cross-well task transfer strategy.

[0116] In summary, this invention, based on the organic synergy and integrated operation of the three embodiments described above, constructs a complete prediction system that integrates physical mechanism perception, data defect self-healing, and dynamic quantitative assessment of uncertainty. This enables full-scenario, multi-dimensional intelligent monitoring of unconventional oil well production dynamics.

[0117] Example 1 focuses on solving the problem of whether the prediction is accurate under abrupt changes; Example 2 addresses the challenge of whether the system is still feasible when data is missing; Example 3 ensures the fundamental requirement that the prediction results conform to the common sense of physics.

[0118] This invention, by explicitly introducing event vectors and physical constraint channels, completely breaks down the "black box" barrier of traditional artificial intelligence models regarding the physical mechanisms of oil well production. In practical industrial applications, this method can significantly reduce prediction errors at points of change in operating conditions, and provides oilfield decision-makers with production prediction ranges accompanied by confidence assessments, thus providing solid and reliable technical support for intelligent management, precise production allocation, and risk-controlled development of oil and gas fields.

[0119] Embodiments of the present invention have been presented and described. It will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to the embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An artificial intelligence-based method for predicting oil well production, characterized in that, include: Collect multi-source real-time sensor data and discrete human intervention event data during the oil well production process; An event vector containing an intervention time decay operator is constructed using the collected discrete artificial intervention event data; The system calculates a data quality score based on the collected multi-source real-time sensor data and uses the data quality score to perform physical consistency repair on missing data. The constructed event vector and the repaired multi-source real-time sensor data are injected into the state space model, and closed-loop filtering is performed in combination with the data quality score to obtain the production potential. By using physical constraint channels and data-driven channels to perform nonlinear mapping on production potential, a preliminary predicted production output is generated. By using a variable-point adaptive order-preserving prediction algorithm, the size of the verification window is dynamically adjusted according to the variable-point probability triggered by the event vector to obtain the prediction range of output. The prediction residuals, change point probabilities, and data quality scores generated from the initial production forecast are coupled together to calculate and output the prediction reliability indicator. The pre-trained task transfer model is used to adaptively fine-tune the initial predicted output and prediction range to obtain the final prediction result.

2. The method for predicting oil well production based on artificial intelligence according to claim 1, characterized in that, The In this context, the multi-source real-time sensor data includes wellhead pressure, bottom hole pressure, pumping unit load, and instantaneous discharge rate. The process of acquiring multi-source real-time sensor data includes: high-frequency sampling of the wellhead pressure, the bottom hole pressure, the pumping unit load, and the instantaneous discharge rate through edge computing nodes deployed at the well site; The sliding window statistical method is used to identify transient outliers in the sampled data, and the data quality score is corrected based on the distribution density of the transient outliers. Before the sampled data is transmitted to the cloud database, an illegal observation value that exceeds the rated power diagram range of the pumping unit is removed using a threshold screening algorithm based on physical boundaries, so as to ensure that the initial observation sequence entering the state space model conforms to the physical benchmark logic of oil and gas production.

3. The method for predicting oil well production based on artificial intelligence according to claim 1, characterized in that, The In this process, the logic for constructing event vectors is as follows: the discrete human intervention event data is vectorized and encoded, and a time decay operator is introduced in combination with the intervention time to explicitly characterize the intensity of the continuous impact of data distribution changes. The logic for constructing the event vector includes: matching the corresponding preset influence half-life according to the type of discrete human intervention event, and calculating the amplitude of the time decay operator using the preset influence half-life; The event vector constructs a nonlinear activation mapping in the hidden layer of the state-space model. Within a preset time window after the intervention time, it asymmetrically enhances or suppresses the gain coefficient of the state transition function, thereby achieving explicit capture of the segmented production pattern.

4. The method for predicting oil well production based on artificial intelligence according to claim 1, characterized in that, The In this study, the criteria for calculating the data quality score include the data missing rate, sensor drift rate, and physical-logic consistency residuals. The physical consistency repair process includes: identifying continuous missing intervals in multi-source real-time sensor data, extracting the production evolution trend before and after the continuous missing intervals, generating a physical constraint completion sequence by combining the oil well production decline model, and performing weighted fusion of the completion sequence based on the data quality score.

5. The method for predicting oil well production based on artificial intelligence according to claim 1, characterized in that, The In this context, the evolution formula for the state-space model is: ， ，2 in, For production potential, For event vectors, These are external control parameters. This is the state transition function. For process noise, To observe the noise, To observe yield; During the closed-loop filtering process, the observed noise The covariance matrix is ​​dynamically adjusted in real time based on the data quality score; when the data quality score drops below a preset value, the observation noise is increased. The covariance weights make the production potential The estimation result is transferred to the state transition function. The predicted trajectory is tilted.

6. The method for predicting oil well production based on artificial intelligence according to claim 1, characterized in that, The In this process, the physical constraint channel introduces an oil well production decline model to impose numerical boundary constraints on the nonlinear output of the data-driven channel.

7. The method for predicting oil well production based on artificial intelligence according to claim 1, characterized in that, The In this context, the logic for dynamically adjusting the size of the verification window is as follows: when the probability of the change point exceeds a preset threshold, the rolling verification window of the order-preserving prediction algorithm is reduced or reset, and the width of the prediction interval is increased. The nonlinear mapping process between the physical constraint channel and the data-driven channel adopts a dynamic gating fusion mechanism. By calculating the evolution rate of the prediction residual, the mixing ratio of the output trend of the physical constraint channel and the output characteristics of the data-driven channel is adjusted in real time.

8. The method for predicting oil well production based on artificial intelligence according to claim 1, characterized in that, The In this context, the formula for calculating the predicted reliability indicator is: ， in, To predict reliability indicators, To predict residuals, For the probability of a change point, Score the data quality. To predict interval coverage, For multi-criteria logic gate functions; The predicted reliability indicator is divided into three numerical levels: available, cautious, and unavailable. The multi-criteria logic gate function The execution logic includes: real-time monitoring of the abrupt change intensity of the production potential; when the rate of change of the derivative of the production potential is abnormal and accompanied by a decrease in the data quality score, forcibly setting the prediction reliability flag to the unavailable level and triggering a model failure warning signal.

9. The method for predicting oil well production based on artificial intelligence according to claim 1, characterized in that, The In this process, the pre-trained task transfer model is obtained by performing a self-supervised learning task on the historical sample set to reconstruct consistency constraints based on missing values; The adaptive fine-tuning process includes: using the feature extraction operator obtained from the self-supervised learning task to extract deep features from the repaired multi-source real-time sensor data; dynamically concatenating the extracted deep features with the production potential as the input vector of the task transfer model; and during the fine-tuning process, by comparing the consistency between the preliminary predicted production and the final predicted result, reversibly correcting the fusion weights of the physical constraint channel and the data-driven channel to eliminate representation bias caused by data discontinuity.

10. The method for predicting oil well production based on artificial intelligence according to claim 7, characterized in that, When adjusting the size of the rolling verification window, the order-preserving prediction algorithm performs a real-time calibration process for the inconsistency score: by comparing the deviation of the preliminary predicted output from the quantile of the historical prediction interval, a real-time uncertainty index is calculated; when the change point probability triggers a change point warning, the real-time uncertainty index is used to drive the confidence upper and lower limits of the prediction interval to expand asymmetrically until the fluctuation range of the prediction residual returns to the preset steady-state interval, and then the normal size of the rolling verification window is restored.

Images