AI-based oil well production prediction system and method

Abstract

This invention provides an AI-based oil well production prediction system and method, belonging to the field of artificial intelligence technology. The system includes: a production data collection module that collects real-time oil well production data and processes the real-time data to obtain processed data; a historical data collection module that collects historical oil well production data; a feature extraction module that processes the historical data to obtain feature data; and a prediction module that constructs a recurrent neural network model, optimizes the recurrent neural network model using an artificial lemming optimization algorithm to obtain an optimal model, and inputs the processed real-time data and the historical data into the optimal model to obtain the predicted production. This invention constructs a recurrent neural network model, optimizes the recurrent neural network model using an artificial lemming optimization algorithm, and inputs the processed data and the feature data into the optimal model to obtain the predicted production, thereby improving the accuracy and efficiency of oil well production prediction.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to an AI-based oil well production prediction system and method. Background Technology

[0002] In the early stages of oil well production prediction technology development, this field primarily relied on geostatistical methods and simplified physical simulation techniques. Geologists constructed reservoir geological models and, combined with flow mechanics theory, attempted to predict the production trends of oil wells at different development stages. However, this traditional approach has many limitations, the core issues being the simplification assumptions of the geological model and the constraints of computational resources. While geostatistical methods can establish statistical relationships using historical production data and geological information, their prediction accuracy is often limited by the degree of simplification of the geological model. For example, when constructing reservoir geological models, complex geological structures often need to be simplified, such as assuming heterogeneous reservoirs as homogeneous bodies, or simplifying complex structures like faults and fractures into regular geometric shapes. While this simplification reduces the complexity of model construction, it also leads to deviations between prediction results and actual conditions. Simultaneously, flow mechanics theory, in describing the flow patterns of fluids in reservoirs, also requires numerous assumptions. For example, assuming the fluid is incompressible and neglecting the effects of gravity. These assumptions may hold true under simple geological conditions, but under complex geological conditions, such as multiphase flow or non-Darcy flow, the prediction accuracy will be significantly reduced. As oil and gas field development progresses, a large amount of production data accumulates, including heterogeneous data from multiple sources such as geological parameters, engineering parameters, and production dynamics. However, traditional methods have failed to effectively integrate this data, resulting in limited prediction accuracy.

[0003] Traditional methods for predicting oil well production are inaccurate. Summary of the Invention

[0004] This invention provides an AI-based oil well production prediction system and method to address the shortcomings of inaccurate oil well production prediction in existing technologies.

[0005] On one hand, the present invention provides an AI-based oil well production prediction system, the system comprising:

[0006] The historical data collection module is used to collect historical data on oil well production.

[0007] The production data collection module is used to collect real-time data on oil well production and preprocess the real-time data to obtain processed data.

[0008] The feature extraction module is used to extract spatiotemporal features from the historical data to obtain multidimensional feature data;

[0009] The prediction module is used to construct a recurrent neural network model, optimize the recurrent neural network model using the artificial lemming optimization algorithm to obtain the optimal model, and input the processed data and the feature data into the optimal model to obtain the predicted output.

[0010] According to the AI-based oil well production prediction system provided by the present invention, the real-time data preprocessing steps include:

[0011] The dynamic threshold method was used to remove outliers from the production data, and the spatiotemporal kriging interpolation method was used to supplement the missing values ​​with the synchronous data from adjacent wells to obtain the initial data.

[0012] The initial data is normalized using the maximum value normalization method to obtain the processed data.

[0013] According to the AI-based oil well production prediction system provided by the present invention, the spatiotemporal feature extraction step for the historical data includes:

[0014] The gradient of each grid point in the permeability distribution map is calculated using the central difference method to obtain permeability gradient data.

[0015] The rate of change of the formation pressure time series data was calculated using the order difference method to obtain the pressure change rate;

[0016] The permeability gradient data is multiplied point by point with the pressure change rate to obtain the permeability-pressure coupling coefficient;

[0017] Wavelet decomposition is performed on the time series of moisture content to obtain multiple sub-bands. The energy of each sub-band is calculated to obtain the energy percentage.

[0018] The entropy value is calculated using the Shannon entropy formula in conjunction with the energy ratio mentioned above, thus obtaining the water content fluctuation entropy.

[0019] Based on the time series data of water injection volume and liquid production volume, the ratio of water injection volume to liquid production volume is calculated, and a logarithmic transformation is performed on the ratio to obtain the water injection efficiency index.

[0020] According to the AI-based oil well production prediction system provided by the present invention, the steps for constructing a recurrent neural network model include:

[0021] The network architecture includes an input layer, an RNN layer, a fully connected layer, and an output layer. The input layer receives the processed data and the feature data, and concatenates the processed data and the feature data to obtain a concatenated vector. The RNN layer captures the long-short-term dependencies in the concatenated vector. The fully connected layer fuses the temporal features output by the RNN layer with the static features in the input to obtain higher-order abstract features. The output layer maps the higher-order features output by the fully connected layer to the predicted output.

[0022] A Dropout layer is added between the RNN layer and the fully connected layer; the Dropout layer is used to prevent the model from overfitting.

[0023] A self-attention layer is set between the RNN layer and the Dropout layer, and the adjacency matrix is ​​superimposed on the self-attention score matrix as a bias term to generate spatiotemporal fusion features.

[0024] According to the AI-based oil well production prediction system provided by the present invention, the adjacency matrix acquisition step includes:

[0025] Obtain the coordinates of the target oil well and adjacent wells, and calculate the distance between the target oil well and adjacent wells using Euclidean distance. If the distance between the target oil well and adjacent wells is less than or equal to a preset threshold, calculate the matrix elements according to the inverse distance weight to obtain the initial matrix.

[0026] The initial matrix is ​​normalized to obtain the adjacency matrix.

[0027] According to the AI-based oil well production prediction system provided by the present invention, the step of superimposing the adjacency matrix with bias terms onto the self-attention score matrix includes:

[0028] The adjacency matrix is ​​scaled according to the spatial correlation weight between wells and then superimposed on the self-attention layer score matrix. The superposition intensity is dynamically adjusted by trainable parameters to obtain the corrected score matrix.

[0029] The corrected score matrix is ​​normalized to obtain the attention weights;

[0030] The spatiotemporal fusion features are obtained by weighting and aggregating the value matrix according to the attention weights.

[0031] According to the AI-based oil well production prediction system provided by the present invention, the optimization steps of the artificial lemming optimization algorithm include:

[0032] The mean squared error is used as the fitness function of the model;

[0033] Initialize the population and randomly generate lemming individuals. Each lemming individual represents a set of hyperparameters of a recurrent neural network.

[0034] An iterative mechanism is set up to randomly generate energy factors each time a new population is generated;

[0035] The initial population is updated according to the update strategy, and the fitness value of each lemming individual is calculated after the update. If the output fitness value is higher than the preset fitness threshold, the optimal hyperparameter is output. Otherwise, the iteration continues until the fitness value is higher than the preset fitness threshold or the maximum number of iterations is reached to obtain the optimal hyperparameter.

[0036] According to the AI-based oil well production prediction system provided by the present invention, the update strategy includes:

[0037] When the energy factor is greater than the preset threshold, the location of individuals in the population is updated by randomly selecting long-distance migration behavior or burrowing behavior.

[0038] When the energy factor is less than or equal to a preset threshold, the population's individual location is updated by randomly selecting either foraging behavior or predator avoidance behavior.

[0039] According to the AI-based oil well production prediction system provided by the present invention, the prediction module includes: a collection unit and an early warning unit; the collection unit is used to collect actual production data; the early warning unit is used to receive the actual production data, analyze the actual production and predicted production, and send the analysis results to the user.

[0040] On the other hand, the present invention also provides an AI-based oil well production prediction method, the method comprising:

[0041] Collect real-time and historical data on oil well production;

[0042] The production data and the historical data are preprocessed separately to obtain processed data and feature data;

[0043] A recurrent neural network model is constructed, and the artificial lemming optimization algorithm is used to optimize the recurrent neural network model to obtain the optimal model;

[0044] The processed data and the feature data are input into the optimal model to obtain the predicted output.

[0045] This invention provides an AI-based oil well production prediction system and method. The system collects real-time and historical oil well production data, preprocesses and extracts features from both data to obtain processed and feature data. A recurrent neural network model is constructed, and the artificial lemming optimization algorithm is used to optimize this model to obtain an optimal model. The processed and feature data are then input into the optimal model to obtain the predicted production. This invention solves the problem of inaccurate oil well production prediction using traditional techniques, improving the accuracy and efficiency of oil well production prediction. Outlier handling in real-time data improves data reliability, interpolation reduces the impact of missing values, and sliding window smoothing effectively suppresses short-term noise in historical data. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0047] Figure 1 This is a flowchart illustrating the AI-based oil well production prediction system provided in an embodiment of the present invention.

[0048] Figure 2 This is a schematic diagram of the installation process of the AI-based oil well production prediction method provided in an embodiment of the present invention;

[0049] Figure 3 This is a schematic diagram of the optimization steps of the artificial lemming optimization algorithm for an oil well operation status analysis system provided in an embodiment of the present invention. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0051] The following is combined with Figures 1-3 This invention describes an AI-based oil well production prediction system and method.

[0052] like Figures 1-3 As shown in the embodiment of the present invention, an AI-based oil well production prediction system is provided, which includes:

[0053] The production data collection module is used to collect real-time data on oil well production and preprocess the real-time data to obtain processed data.

[0054] In this embodiment, the actual data may include parameters such as formation pressure, oil reservoir permeability, and water cut.

[0055] The real-time data preprocessing steps include:

[0056] The dynamic threshold method was used to remove outliers from the production data, and the spatiotemporal kriging interpolation method was used to supplement the missing values ​​with the synchronous data from adjacent wells to obtain the initial data.

[0057] In this embodiment, efficient data structures, such as hash tables or tree structures, can be used to identify and delete duplicate data records to ensure data uniqueness and accuracy. Data is mapped to positions in the table, achieving an average time complexity of O(1) for searching, inserting, and deleting elements. When duplicate records need to be detected, each data item can be inserted into the hash table. During insertion, if a data item already exists in the hash table, it indicates a duplicate record, which can be marked for deletion or skipped directly. In this way, duplicate records can be quickly identified, and unique data can be effectively preserved. Data is mapped to positions in the table, achieving an average time complexity of O(1) for searching, inserting, and deleting elements. For outliers in the measured data, we can use the quartile method in statistics for identification. Furthermore, quartile calculation involves dividing the dataset into four parts to analyze the data distribution. Specifically, the first quartile (Q1) and the third quartile (Q3) represent the 25% and 75% positions of the data, respectively. Based on these two quartiles, we can calculate the interquartile range (IQR), which is the difference between Q3 and Q1, i.e., IQR = Q3 - Q1. Using IQR, outliers are defined as data points that are less than Q1 - 1.5 × IQR or greater than Q3 + 1.5 × IQR. The advantage of this method is that it is unaffected by extreme values ​​and can perform relatively robust outlier detection based on the central location of the data.

[0058] The initial data is normalized using the maximum value normalization method to obtain the processed data.

[0059] In this embodiment, the initial data includes parameters such as layer pressure, reservoir permeability, daily fluid production, and water cut. These data collectively constitute a complex multidimensional dataset, providing a basis for subsequent analysis and decision-making. Since the dimensions and numerical ranges of different parameters vary significantly, direct analysis may lead to the neglect or overemphasis of certain features, reducing the accuracy and reliability of the model. Therefore, preprocessing the initial data using the maximum value normalization method becomes a crucial step in improving data quality and model performance. Maximum value normalization, as a linear transformation technique, maps each feature value to the [0,1] interval using a formula. Specifically, for each feature dimension in the dataset, the maximum value of all sample points under that dimension is calculated, and then the original value of each sample point is divided by this maximum value. For example, when processing the feature "daily fluid production," the daily fluid production records of all oil wells are first traversed to find the maximum value. Then, the daily fluid production of each oil well is divided by this maximum value, ensuring that the normalized values ​​are all between 0 and 1. This not only eliminates the dimensional differences between different parameters but also preserves the relative relationships between the original data, providing a standardized data foundation for subsequent analysis. Define the names and corresponding values ​​of each parameter. Iterate through each parameter column and calculate the maximum value of all values ​​in that column. Next, apply a normalization formula to transform each data point. Finally, integrate the processed data into a new standardized dataset. In practice, special handling is needed for zero or negative values ​​to avoid errors such as negative values ​​or division by zero after normalization. After maximum value normalization, all features are unified to the same order of magnitude, significantly improving data consistency and comparability. This not only facilitates intuitive comparison between features during visualization but also provides high-quality input for subsequent machine learning model training. Normalized parameters such as penetration rate and pressure can participate more fairly in model weight calculations, avoiding model bias caused by numerical differences, thereby improving prediction accuracy and model generalization ability.

[0060] The historical data collection module is used to collect historical data on oil well production.

[0061] In this embodiment, historical data includes production parameters and corresponding production output. The historical data module is specifically designed to collect historical data generated during the oil well production process. Production parameters, as key indicators reflecting the oil well's production status, include wellhead pressure, temperature, production pressure differential, and water cut. These parameters are directly related to the oil well's production efficiency and reservoir characteristics, and are an important basis for the predictive model to understand the oil well's production dynamics.

[0062] The feature extraction module is used to extract spatiotemporal features from the historical data to obtain multidimensional feature data;

[0063] The steps for extracting spatiotemporal features from the historical data include:

[0064] The gradient of each grid point in the permeability distribution map is calculated using the central difference method to obtain permeability gradient data.

[0065] The rate of change of the formation pressure time series data was calculated using the order difference method to obtain the pressure change rate;

[0066] The permeability gradient data is multiplied point by point with the pressure change rate to obtain the permeability-pressure coupling coefficient;

[0067] Wavelet decomposition is performed on the time series of moisture content to obtain multiple sub-bands. The energy of each sub-band is calculated to obtain the energy percentage.

[0068] The entropy value is calculated using the Shannon entropy formula in conjunction with the energy ratio mentioned above, thus obtaining the water content fluctuation entropy.

[0069] Based on the time series data of water injection volume and liquid production volume, the ratio of water injection volume to liquid production volume is calculated, and a logarithmic transformation is performed on the ratio to obtain the water injection efficiency index.

[0070] The prediction module is used to construct a recurrent neural network model, optimize the recurrent neural network model using the artificial lemming optimization algorithm to obtain the optimal model, and input the processed data and the feature data into the optimal model to obtain the predicted output.

[0071] The steps to build a recurrent neural network model include:

[0072] The network architecture includes an input layer, an RNN layer, a fully connected layer, and an output layer. The input layer receives the processed data and the feature data, and concatenates the processed data and the feature data to obtain a concatenated vector. The RNN layer captures the long-short-term dependencies in the concatenated vector. The fully connected layer fuses the temporal features output by the RNN layer with the static features in the input to obtain higher-order abstract features. The output layer maps the higher-order features output by the fully connected layer to the predicted output.

[0073] A Dropout layer is added between the RNN layer and the fully connected layer. The Dropout layer is used to prevent the model from overfitting.

[0074] In this embodiment, adding a Dropout layer between the RNN layer and the fully connected layer is an effective regularization technique to prevent overfitting during training. Overfitting refers to a model performing well on the training set. The Dropout layer randomly ignores a portion of neurons during training, breaking the interdependencies between neurons and making the model more robust during learning. Specifically, a Dropout percentage can be set, such as 20% or 30%, meaning that each time the model parameters are updated, 20% to 30% of neurons are randomly selected not to participate in the computation. This improves the model's generalization ability, allowing it to better adapt to different datasets and real-world scenarios.

[0075] A self-attention layer is set between the RNN layer and the Dropout layer, and the adjacency matrix is ​​superimposed on the self-attention score matrix as a bias term to generate spatiotemporal fusion features.

[0076] In this embodiment, a self-attention layer is added between the RNN layer and the Dropout layer. The self-attention mechanism allows the model to dynamically focus on information from different parts of the input sequence when processing input data. For example, in oil well production data, changes in production at certain times may be influenced by data from previous times; the self-attention mechanism helps the model capture this relationship during learning. The self-attention layer enhances the model's learning ability by calculating the correlation between different time steps in the sequence, enabling the model to integrate information more effectively. Specifically, the self-attention layer dynamically allocates weights based on the importance of the input information, making it more biased towards highly relevant features during prediction. This method allows the model to flexibly respond to complex inputs, thereby improving prediction accuracy.

[0077] The adjacency matrix acquisition step includes:

[0078] Obtain the coordinates of the target oil well and adjacent wells, and calculate the distance between the target oil well and adjacent wells using Euclidean distance. If the distance between the target oil well and adjacent wells is less than or equal to a preset threshold, calculate the matrix elements according to the inverse distance weight to obtain the initial matrix.

[0079] The initial matrix is ​​normalized to obtain the adjacency matrix.

[0080] The steps of superimposing the adjacency matrix with bias terms into the self-attention score matrix include:

[0081] The adjacency matrix is ​​scaled according to the spatial correlation weight between wells and then superimposed on the self-attention layer score matrix. The superposition intensity is dynamically adjusted by trainable parameters to obtain the corrected score matrix.

[0082] The corrected score matrix is ​​normalized to obtain the attention weights;

[0083] The spatiotemporal fusion features are obtained by weighting and aggregating the value matrix according to the attention weights.

[0084] The optimization steps of the artificial lemming optimization algorithm include:

[0085] The mean squared error is used as the fitness function of the model.

[0086] In this embodiment, mean squared error is used as the fitness function. Mean squared error is a commonly used metric for evaluating regression prediction error. It calculates the average of the squared differences between predicted and actual values, reflecting the model's prediction accuracy. The formula is as follows:

[0087]

[0088] in, For the sum of squared errors, Based on actual observation data, These are predicted values.

[0089] Initialize the population and randomly generate lemming individuals. Each lemming individual represents a set of hyperparameters for a recurrent neural network.

[0090] In this embodiment, each lemming individual represents a set of hyperparameters of a recurrent neural network, such as the learning rate, batch size, and number of hidden layer units. A reasonable range of values ​​is set for each hyperparameter. For example, the learning rate can be randomly selected between [0.0001, 0.1], and the batch size can be randomly determined between [16, 128]. Generally, a population size of 50 to 200 lemming individuals ensures that a sufficient search space is covered, increasing the chance of finding the global optimum.

[0091] An iterative mechanism is set up to randomly generate energy factors each time a new population is generated.

[0092] The initial population is updated according to the update strategy, and the fitness value of each lemming individual is calculated after the update. If the output fitness value is higher than the preset fitness threshold, the optimal hyperparameter is output. Otherwise, the iteration continues until the fitness value is higher than the preset fitness threshold or the maximum number of iterations is reached to obtain the optimal hyperparameter.

[0093] The update strategy includes:

[0094] When the energy factor is greater than a preset threshold, the location of individuals in the population is updated by randomly selecting long-distance migration behavior or burrowing behavior.

[0095] In this embodiment, the formula for expressing long-distance migration behavior is:

[0096]

[0097]

[0098] in For the first The position of each individual after iteration This is the optimal position in the current population. As a directional marker, Let N be a Brownian motion vector, following a standard normal distribution N(0,1), to simulate an uncertain migration path. To randomly select the location of an individual, For the first The current location of each individual.

[0099] The formula for expressing the digging behavior is:

[0100]

[0101] in For the first The position of each individual after iteration For the first The current location of each individual. As a directional marker, A random number related to the current iteration number. This is the optimal position in the current population. The location of an individual is randomly selected.

[0102] When the energy factor is less than or equal to a preset threshold, the population's individual location is updated by randomly selecting either foraging behavior or predator avoidance behavior.

[0103] In this embodiment, the formula for expressing foraging behavior is:

[0104]

[0105] in, For the first The position of each individual after iteration This is the optimal position in the current population. As a directional marker, The spiral search radius is calculated using the formula P = radius × (sin(2π × rand) + cos(2π × rand)), simulating the randomness of the foraging path. radius represents the distance between the current individual and the optimal solution. It is a random number. For the first The current location of each individual.

[0106] The formula for expressing predator avoidance behavior is:

[0107]

[0108] in, For the first The position of each individual after iteration This is the optimal position in the current population. As a directional marker, This is the escape coefficient, which decreases with the number of iterations. For the Lévy flight function, generate random step sizes that follow a Lévy distribution to simulate deceptive escape maneuvers. For the first The current location of each individual.

[0109] The data processing module described in this embodiment may include a storage unit and an update unit. The storage unit stores historical data. The update unit updates the historical data stored in the storage unit according to preset time intervals. The storage unit is built using a high-performance distributed time-series database, specifically designed for time-series data, ensuring high throughput and low latency data access in massive data scenarios. Its distributed architecture supports horizontal scaling, dynamically adding nodes as data volume increases, effectively avoiding single-point performance bottlenecks. The storage unit uses a columnar storage engine, optimized for the characteristics of time-series data, offering high compression ratios and high query efficiency, making it particularly suitable for storing high-frequency, high-dimensional time-series data such as oil well production parameters. The update unit, as the core component of data processing, is configured with a daily timed trigger mechanism, specifically for execution during the early morning off-peak hours to avoid interfering with the production system. During the daily update process, the update unit first obtains the complete set of production parameters from the previous day (T-1 day) from the front-end data acquisition system, including key indicators such as formation pressure, reservoir permeability, water cut, wellhead temperature, and daily oil production. After undergoing integrity verification and logical consistency checks, this data is appended to the time-series database of the storage unit in batches, ensuring the atomicity and efficiency of data writes. Simultaneously, the update unit can implement intelligent data lifecycle management strategies. The system maintains a sliding time window, for example, retaining historical data from the last three years. During daily updates, it automatically calculates the furthest date currently in storage and removes data before that date from the storage unit. For example, when performing an update on April 8, 2025, the system will locate the furthest date of April 8, 2022, and safely delete data before that date from the distributed database, freeing up storage space and ensuring the timeliness of the dataset. Through the collaborative work of the storage unit and the update unit, the integrity and traceability of historical data are guaranteed, while the reasonable size of the dataset is maintained through automated update and eviction mechanisms.

[0110] The prediction module includes a collection unit and an early warning unit. The collection unit is used to collect actual production data. The early warning unit is used to receive the actual production data, analyze the actual production and predicted production, and send the analysis results to the user.

[0111] In this embodiment, the collection unit can employ an RTU for efficient acquisition of pulse signals from the wellhead flowmeter. It not only monitors flow rate in real time but also accurately post-processes the pulse signals to convert them into instantaneous production values. Specifically, after signal acquisition, signal amplification and filtering effectively improve signal quality, eliminate environmental noise and other interference factors, thereby ensuring more accurate and reliable data acquisition. To achieve dynamic data acquisition, the acquisition frequency can be set to 1 time / 5 seconds, ensuring effective monitoring of wellhead flow while reducing data redundancy and simplifying subsequent data processing. All collected real-time production data is uploaded to the edge computing node via the MQTT protocol. MQTT is a lightweight messaging protocol suitable for efficient transmission of small amounts of data in environments with limited bandwidth and unstable networks. This method ensures accurate, secure, and timely data upload to the edge node. The early warning unit calculates the relationship between actual production and the predicted range in real time: Normal range: actual value is within the predicted range; Yellow warning: actual value exceeds the predicted range but the deviation is less than 15%; Red warning: actual value deviation is greater than or equal to 15%. Real-time early warning and data updates ensure the long-term effective operation of the system.

[0112] Based on the same general inventive concept, this invention also protects an AI-based oil well production prediction method, which includes:

[0113] Collect real-time and historical data on oil well production.

[0114] The production data and the historical data are preprocessed separately to obtain processed data and feature data.

[0115] A recurrent neural network model is constructed, and the artificial lemming optimization algorithm is used to optimize the recurrent neural network model to obtain the optimal model.

[0116] The processed data and the feature data are input into the optimal model to obtain the predicted output.

[0117] Example 1: Real-time data collection: 10:00, formation pressure 3,200, reservoir permeability 150, water cut 45.2%, wellhead temperature 68.3℃. 10:15, formation pressure 3,195, reservoir permeability 148, water cut 45.5%, wellhead temperature 68.5℃. 10:30, reservoir permeability 152, water cut 45.8%, wellhead temperature 68.2℃. The outlier is formation pressure 3,195, calculated using the adjacent mean of 3,205. The missing formation pressure at 10:30 is calculated as 3,190 using linear interpolation. Normalization scales the parameters to the [0,1] interval. After smoothing the historical data, the smoothed data is input into the Bayesian network model to obtain feature data. For constructing the recurrent neural network model, the input layer nodes are set to four, the RNN layer has 64 units, the self-attention layer has 4 attention heads, the dropout layer has a dropout rate of 0.3, the output layer has 1 node, the current formation pressure is 3,190 psi, the permeability is 148 mD, the water cut is 45.8%, and the feature data are input into the optimized recurrent neural network model to obtain a predicted daily yield of 635 bbl / d.

[0118] This invention collects real-time and historical data on oil well production, performs feature extraction on both data to obtain processed data and feature data, constructs a recurrent neural network model, and optimizes the model using an artificial lemming optimization algorithm to obtain an optimal model. The processed data and feature data are then input into the optimal model to obtain predicted production, thus improving the accuracy and efficiency of oil well production prediction. Outlier handling in the implementation data improves data reliability, interpolation reduces the impact of missing values, and sliding window smoothing effectively suppresses short-term noise in historical data.

[0119] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0120] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0121] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An AI-based oil well production prediction system, characterized in that, The system includes: The historical data collection module is used to collect historical data on oil well production. The production data collection module is used to collect real-time data on oil well production and preprocess the real-time data to obtain processed data. The feature extraction module is used to extract spatiotemporal features from the historical data to obtain multidimensional feature data; The prediction module is used to construct a recurrent neural network model, optimize the recurrent neural network model using the artificial lemming optimization algorithm to obtain the optimal model, and input the processed data and the feature data into the optimal model to obtain the predicted output; The steps for extracting spatiotemporal features from the historical data include: The gradient of each grid point in the permeability distribution map is calculated using the central difference method to obtain permeability gradient data. The rate of change of formation pressure time series data was calculated using the step difference method to obtain the pressure change rate; The permeability gradient data is multiplied point by point with the pressure change rate to obtain the permeability-pressure coupling coefficient; Wavelet decomposition was performed on the time series of moisture content to obtain multiple sub-bands. The energy of each sub-band was calculated to obtain the energy percentage. The entropy value is calculated using the Shannon entropy formula in conjunction with the energy ratio mentioned above, thus obtaining the water content fluctuation entropy. Based on the time series data of water injection volume and liquid production volume, the ratio of water injection volume to liquid production volume is calculated, and the ratio is logarithmically transformed to obtain the water injection efficiency index. The steps to build a recurrent neural network model include: The network architecture includes an input layer, an RNN layer, a fully connected layer, and an output layer. The input layer receives the processed data and the feature data, and concatenates the processed data and the feature data to obtain a concatenated vector. The RNN layer captures the long-short-term dependencies in the concatenated vector. The fully connected layer fuses the temporal features output by the RNN layer with the static features in the input to obtain higher-order abstract features. The output layer maps the higher-order features output by the fully connected layer to the predicted output. A Dropout layer is added between the RNN layer and the fully connected layer; the Dropout layer is used to prevent the model from overfitting. A self-attention layer is set between the RNN layer and the Dropout layer, and the adjacency matrix is ​​superimposed on the self-attention score matrix as a bias term to generate spatiotemporal fusion features. The adjacency matrix acquisition includes: Obtain the coordinates of the target oil well and adjacent wells, and calculate the distance between the target oil well and adjacent wells using Euclidean distance. If the distance between the target oil well and adjacent wells is less than or equal to a preset threshold, calculate the matrix elements according to the inverse distance weight to obtain the initial matrix. The initial matrix is ​​normalized to obtain the adjacency matrix.

2. The AI-based oil well production prediction system according to claim 1, characterized in that, The real-time data preprocessing steps include: The dynamic threshold method was used to remove outliers from the production data, and the spatiotemporal kriging interpolation method was used to supplement the missing values ​​with the synchronous data from adjacent wells to obtain the initial data. The initial data is normalized using the maximum value normalization method to obtain the processed data.

3. The AI-based oil well production prediction system according to claim 1, characterized in that, The steps of superimposing the adjacency matrix with bias terms into the self-attention score matrix include: The adjacency matrix is ​​scaled according to the spatial correlation weight between wells and then superimposed on the self-attention layer score matrix. The superposition intensity is dynamically adjusted by trainable parameters to obtain the corrected score matrix. The corrected score matrix is ​​normalized to obtain the attention weights; The spatiotemporal fusion features are obtained by weighting and aggregating the value matrix based on the attention weights.

4. The AI-based oil well production prediction system according to claim 1, characterized in that, The optimization steps of the artificial lemming optimization algorithm include: The mean squared error is used as the fitness function of the model; Initialize the population and randomly generate lemming individuals. Each lemming individual represents a set of hyperparameters of a recurrent neural network. An iterative mechanism is set up to randomly generate energy factors each time a new population is generated; The initial population is updated according to the update strategy, and the fitness value of each lemming individual is calculated after the update. If the output fitness value is higher than the preset fitness threshold, the optimal hyperparameter is output. Otherwise, the iteration continues until the fitness value is higher than the preset fitness threshold or the maximum number of iterations is reached to obtain the optimal hyperparameter.

5. The AI-based oil well production prediction system according to claim 4, characterized in that, The update strategy includes: When the energy factor is greater than the preset threshold, the location of individuals in the population is updated by randomly selecting long-distance migration behavior or burrowing behavior. When the energy factor is less than or equal to a preset threshold, the population's individual location is updated by randomly selecting either foraging behavior or predator avoidance behavior.

6. The AI-based oil well production prediction system according to claim 1, characterized in that, The prediction module includes a collection unit and an early warning unit; the collection unit is used to collect actual production data; the early warning unit is used to receive the actual production data, analyze the actual production and predicted production, and send the analysis results to the user.

7. An AI-based oil well production prediction method, which employs the AI-based oil well production prediction system as described in any one of claims 1-6, characterized in that, The method includes: Collect real-time and historical data on oil well production; The production data and the historical data are preprocessed separately to obtain processed data and feature data; A recurrent neural network model is constructed, and the artificial lemming optimization algorithm is used to optimize the recurrent neural network model to obtain the optimal model; The processed data and the feature data are input into the optimal model to obtain the predicted output.

Images