A Multi-Reservoir Production Prediction Method Based on CatBoost and Feature Alignment

Abstract

This invention relates to the interdisciplinary field of oil and gas field development and artificial intelligence. Addressing the problem of low prediction accuracy caused by inconsistent feature dimensions across multiple reservoirs and scarcity of target reservoir data, this invention discloses a high-precision multi-reservoir oil production prediction method based on common feature alignment and CatBoost ensemble learning. Historical production data from different reservoirs are collected, preprocessed, and then automatically aligned using set intersection operations. A CatBoost time-series prediction model is constructed, establishing a dual-reservoir fusion training framework of "source reservoir pre-training + target reservoir fine-tuning," extracting source domain knowledge and transferring it to the target model for adaptation. The model is then optimized based on evaluation metrics such as RMSE and MAPE. This method solves the problems of cross-reservoir feature mismatch and data scarcity, achieves knowledge transfer and generalization, significantly improves prediction accuracy, and is applicable to engineering predictions of complex reservoirs.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of oil and gas field development engineering and artificial intelligence. Specifically, it relates to a high-precision oil production prediction method based on common feature alignment and CatBoost ensemble learning in multi-reservoir scenarios. It is applicable to engineering prediction scenarios where different reservoirs have inconsistent feature dimensions, heterogeneous data distribution, and scarce target reservoir data. Background Technology

[0002] As a crucial energy source and chemical raw material, oil production fluctuations directly impact oilfield development efficiency and energy supply-demand balance. Traditional oil production forecasting relies primarily on reservoir engineering empirical formulas, numerical simulations, or linear regression. These methods typically assume relatively stable geological conditions and production processes, making it difficult to address complex factors such as geological heterogeneity among multiple reservoirs, sensor noise, and variations in oil production technologies. This results in insufficient prediction accuracy and poor model generalization. With the development of information technology and intelligent algorithms, using machine learning methods for oil production forecasting has become a research hotspot. Among these, ensemble learning models can automatically learn nonlinear relationships from large amounts of historical data, significantly improving prediction accuracy. However, existing technologies still have core shortcomings: the number of monitoring fields and data collection dimensions differ across reservoirs, making direct modeling prone to feature dimension mismatch errors; deep learning models are susceptible to problems such as prediction flattening and overfitting in small-sample, heterogeneous data scenarios; traditional cross-reservoir modeling lacks an automatic feature alignment mechanism, requiring manual feature selection, which is inefficient and prone to missing key information. Summary of the Invention

[0003] To overcome the shortcomings of the existing technology, the present invention aims to provide a high-precision oil production prediction method for multiple reservoirs based on common feature alignment and CatBoost ensemble learning. This method solves the problem of incompatibility between features across reservoirs, alleviates the problem of data scarcity, and utilizes the CatBoost ensemble learning model to achieve high-precision oil production prediction. To achieve the above objective, the present invention adopts the following technical solution:

[0004] 1. A high-precision prediction method for oil production in multiple reservoirs based on common feature alignment and CatBoost ensemble learning, characterized by the following steps:

[0005] S1. Collect historical oil production data from different oil fields, and automatically align and preprocess the collected data based on common features to construct training and test sets for each oil field.

[0006] S2. Construct a time-series oil production prediction model based on the CatBoost algorithm as the core tool for multi-reservoir production prediction;

[0007] S3. Build a dual-reservoir data fusion training framework, determine the knowledge transfer path between the source reservoir and the target reservoir, and initialize the model parameter transfer strategy and adaptation rules.

[0008] S4. Based on the CatBoost model constructed in step S2, pre-train it on the source reservoir training set and extract the domain knowledge features in the model.

[0009] S5. Based on the fusion training framework built in step S3, transfer the domain knowledge of the source reservoir pre-trained model to the target reservoir.

[0010] S5. Based on the fusion training framework built in step S3, the domain knowledge of the source reservoir pre-trained model is transferred to the target reservoir model, and the model is adapted and trained on the target reservoir training set; the adapted target reservoir model is used to predict the production of the test set and obtain the model evaluation index results.

[0011] S6. Based on the evaluation index results of the model in step S5, analyze and compare from multiple perspectives such as error control, trend fitting, and feature adaptability to determine the final optimal oil production prediction model.

[0012] 2. The method according to claim 1, wherein the data preprocessing in step S1 includes:

[0013] S1.1 Perform preliminary cleaning on the collected historical oil production and related data of multiple oil reservoirs, retain numerical feature columns, and remove string fields, invalid and redundant columns and completely missing fields;

[0014] S1.2. The three-standard-deviation method is used to identify and remove outlier data. Linear interpolation is used to process missing values, and cubic Lagrange interpolation is used to correct abnormal negative values ​​of daily oil production.

[0015] S1.3. Extract year-day and month features based on the date field, and use StandardScaler to standardize the features and target variables. Fit the standardized parameters to the source reservoir data, and reuse the parameters for the target reservoir.

[0016] 3. The method according to claim 1, wherein the automatic alignment process of common features in step S1 is as follows:

[0017] S1.1. Through set intersection operation, the numerical features common to the two reservoirs are automatically extracted to form a common feature set, ensuring the consistency of feature dimensions in cross-reservoir modeling;

[0018] S1.2. Daily oil production is uniformly used as the target prediction variable. If the field names of the target variables differ among different oil reservoirs, they are uniformly named through a field mapping table.

[0019] 4. The method according to claim 1, characterized in that, the process of constructing the time-series prediction model for oil production in step S2 includes:

[0020] S2.1. The CatBoost regression algorithm is used as the core framework to construct an oil production prediction model for multiple oil reservoirs. The model consists of an input feature layer, a tree node splitting layer, a feature importance calculation layer, and an output prediction layer.

[0021] S2.2, Set model parameters: including number of iterations, learning rate, decision tree depth, and regularization coefficient;

[0022] S2.3 The model's optimizer uses the built-in CatBoost optimizer for iterative parameter updates, and the training loss function is the mean squared error.

[0023] 5. The method according to claim 1, characterized in that, in step S3, establishing a dual-reservoir data fusion training framework, determining the knowledge transfer path between the source reservoir and the target reservoir, and initializing the model parameter transfer strategy and adaptation rules, further includes:

[0024] S3.1 Based on the scale and development maturity of reservoir data, divide the reservoirs into source reservoirs and target reservoirs, establish pre-training units for source reservoirs, and establish knowledge adaptation units for target reservoirs;

[0025] S3.2 Source Reservoir Knowledge Extraction: After the source reservoir pre-training unit completes the initial training based on the training data, it extracts core domain knowledge, including feature importance weights, model hyperparameters, etc.

[0026] S3.3 Knowledge Adaptation: The "pre-training-fine-tuning" strategy is adopted to transfer knowledge from the source reservoir domain to the target reservoir model. The model is then adapted and optimized using the target reservoir training set to ensure that the model is adapted to the production patterns of the target reservoir.

[0027] S3.4 Model Update: During the target reservoir model adaptation training process, the validation set loss is calculated periodically. If there is no decrease for several consecutive rounds, training is stopped and the optimal model weights are saved.

[0028] 6. The method according to claim 1, characterized in that, in step S5, knowledge transfer and model adaptation training are completed based on the dual-reservoir fusion training framework, a pre-trained model is trained in the source reservoir and the model is adapted in the target reservoir, the trained target reservoir adaptation model is used to predict the production of the test set, and model evaluation calculation is performed to obtain the model evaluation index calculation results, including the following sub-steps:

[0029] S5.1. Complete source model pre-training and knowledge transfer based on the fusion training framework: Train the CatBoost pre-trained model on the source reservoir training set until the model loss converges, determine the end of training and record the number of training sessions; Based on the transfer strategy in step S3, transfer the knowledge of the source reservoir domain to the target reservoir initial model, perform adaptation training on the target reservoir training set, and match the number of adaptation training sessions with the number of source model training sessions.

[0030] S5.2 Use the trained target reservoir adaptation model to predict oil production on the test set and output the production prediction results;

[0031] S5.3 Obtain the model's predicted output values ​​and the actual output values ​​of the test set, using three complementary evaluation metrics: root mean square error (RMSE), mean percentage error (MAPE), and coefficient of determination (R²). 2 The prediction results are evaluated using the following formulas:

[0032]

[0033] Among them, y i This represents the actual oil production value. This represents the oil production value predicted by the model. denoted as the average actual oil production, and n represents the sample size.

[0034] 7. The method according to claim 1, characterized in that, in step S6, the calculation results of the evaluation index of the target reservoir adaptation model are analyzed, and the model prediction performance is evaluated from multiple perspectives, including error control, trend fitting, and feature adaptability, and finally the optimal prediction model is comprehensively selected. Specifically, this includes:

[0035] S6.1 Error control dimension: Select models with lower root mean square error and average percentage error;

[0036] S6.2 Trend Fitting Dimension: Judged by the degree of overlap between the predicted curve and the actual curve;

[0037] S6.3 Feature Adaptability Dimension: Ensure that the importance weight of core production features in the target reservoir model is ranked high.

[0038] 8. The method according to any one of claims 1-7, characterized in that the method has the following advantages:

[0039] S8.1, Automatic alignment mechanism for common features: Automatically extracts common features from different reservoirs through set operations, solving the modeling error problem caused by inconsistent feature quantities across reservoirs, without manual intervention, thus improving modeling efficiency;

[0040] S8.2 Dual reservoir data fusion training strategy: The fusion training mode of "source reservoir pre-training + target reservoir fine-tuning" is adopted to fully reuse the general production knowledge of the source reservoir, alleviate the problem of target reservoir data scarcity, and improve the model generalization ability.

[0041] S8.3, Robust Model Structure and Efficient Training: The CatBoost ensemble learning model is used as the core predictor, which has the advantages of anti-overfitting, stable prediction, and less prone to flattening predictions, making it suitable for heterogeneous data scenarios with multiple reservoirs.

[0042] In summary, this invention proposes a high-precision oil production prediction method for multiple reservoirs based on common feature alignment and CatBoost ensemble learning. Through common feature alignment, standardized preprocessing, dual-reservoir data fusion, and an ensemble learning model, high-precision oil production prediction is achieved. This method maintains high model training efficiency while fully considering temporal characteristics, trend information, and domain differences between reservoirs, enabling knowledge transfer and generalization of the model from single-reservoir to multi-reservoir scenarios. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the embodiments will be briefly introduced below.

[0044] Figure 1 This is an overall flowchart of the method of the present invention;

[0045] Figure 2 A diagram illustrating the model training framework;

[0046] Figure 3 For model performance comparison;

[0047] Figure 4 A scatter plot of predicted vs. actual values;

[0048] Figure 5 This is a distribution chart of prediction errors;

[0049] Figure 6 This is a feature importance graph. Detailed Implementation

[0050] Example 1:

[0051] See Figure 1 The present invention provides a high-precision prediction method for multi-reservoir oil production based on common feature alignment and CatBoost ensemble learning, comprising the following steps:

[0052] In this embodiment, step S1 specifically includes the following:

[0053] S1.1 Data Acquisition: Historical oil production data for the F15_1_C and data reservoirs were collected through the oilfield field production monitoring system. Data parameters included date, average bottom hole pressure, average bottom hole temperature, average tubing differential pressure, average annular pressure, average choke size percentage, average wellhead pressure, average wellhead temperature, choke size differential pressure, and daily oil production. Specifically, the F15_1_C reservoir depth range corresponds to a time span from July 2013 to March 2015, with a total of 696 data points; the data reservoir time span is from January 2016 to August 2022, with a total of 3375 data points. The data acquisition interval for both was one day, meeting the requirements for continuity of time-series data.

[0054] S1.2 Outlier Removal: The collected data undergoes preliminary cleaning, and outlier data is removed using the 3σ criterion. The mean μ and standard deviation σ of each parameter are calculated, and data exceeding the range of [μ-3σ, μ+3σ] are identified as outliers. These outliers are often caused by sensor malfunctions or transmission errors. For example, in the daily oil production data of the F15_1_C reservoir, three outlier samples exceeded the mean range, and 12 outlier samples in the data reservoir were removed, increasing the data pass rate to 99.2%.

[0055] Table 1. Partial Sample Data of the Source Domain Training Set

[0056] 2013 / 7 / 24 1.92627 0.3522 26.12304 17.85 2013 / 7 / 25 163.59426 60.31574 61.47308 725.0 2013 / 7 / 31 273.94693 105.55137 179.38234 190.35

[0057] S1.3 Missing Value Imputation: After initial data cleaning, missing data is imputed using linear interpolation. The linear interpolation formula is as follows:

[0058] y=y1+(x-x1)×(y2-y1) / (x2-x1)

[0059] Where x represents the date of the missing point, x1 and x2 are the adjacent valid dates of the missing point, and y1 and y2 are the parameter values ​​for the corresponding dates. This method ensures the continuity of the dataset, achieving a 100% data integrity rate after imputation.

[0060] S1.4, Negative Value Correction: Five data points were corrected for daily oil production values ​​that were not positive, with the smallest being -3.2m. 3 / d is corrected using cubic Lagrange interpolation. The Lagrange interpolation polynomial for six adjacent valid points is constructed as follows:

[0061] L(x)=∑(y i ×l i (x))

[0062] l i (x)=Π((xx j ) / (x i -x j))

[0063] Where x is the negative point date, x i y i For adjacent valid points, the date and output value, l i (x) is the Lagrange basis function. This method fits the data trend and avoids errors caused by simple substitution.

[0064] S1.5 Feature Engineering: Extract day of year and month features based on the date field to capture seasonal and periodic patterns. The formula for calculating day of year is:

[0065] DOY=(date-date.replace(month=1, day=1)).days+1

[0066] The month features are encoded using one-hot encoding, resulting in a total of 12 binary features.

[0067] S1.6 Standardization Processing: StandardScaler is used to standardize the features and target variables. The formula is as follows:

[0068] x_normalized=(x-μ) / σ

[0069] Where μ is the characteristic mean and σ is the characteristic standard deviation. Based on the F15_1_C reservoir data fitting normalizer parameters, the data reservoir data reuses these parameters to ensure the consistency of data distribution across reservoirs.

[0070] S1.7 Automatic Alignment of Common Features: Through set intersection operations, eight common features shared by the two reservoirs are automatically identified and extracted: average bottomhole pressure, average bottomhole temperature, average tubing differential pressure, average annular pressure, average choke size percentage, average wellhead pressure, average wellhead temperature, and choke size differential pressure, ensuring consistency in feature dimensions. The alignment formula is:

[0071] Feature_common=Feature_F15∩Feature_data

[0072] In this embodiment, step S2 constructs a time-series oil production prediction model based on CatBoost ensemble learning, aiming to capture the nonlinear laws governing oil production changes over time and with multiple parameters. Specifically, this includes the following:

[0073] S2.1 Model Architecture Design: The CatBoost regression algorithm is used as the core framework to construct an oil production prediction model for multiple oil reservoirs. The model consists of an input feature layer, a tree node splitting layer, a feature importance calculation layer, and an output prediction layer. The input layer receives a 56-dimensional time-series feature vector; the tree node splitting layer selects the optimal splitting features and thresholds using a greedy algorithm; the feature importance calculation layer adopts a feature importance evaluation method based on the number of splits; the output layer outputs the predicted daily oil production value for day 8.

[0074] S2.2 Parameter Configuration: The model hyperparameters were set as follows: number of iterations n_estimators = 1500, learning rate = 0.02, decision tree depth max_depth = 7, L2 regularization coefficient 12_leaf_reg = 1, random seed random_state = 42, early stopping rounds early_stopping_rounds = 50. These parameters were optimized through grid search and 5-fold cross-validation, achieving optimal performance on the F15_1_C reservoir validation set.

[0075] S2.3 Loss Function and Optimization: The loss function used for model training is the mean squared error (MSE), as shown in the following formula:

[0076]

[0077] Among them, y i This represents the actual oil production value. Let n be the predicted oil production value, and n be the number of samples. The optimizer uses CatBoost's built-in gradient boosting optimizer, which adjusts the parameter update step size through adaptive learning rate adjustment to ensure model convergence.

[0078] In this embodiment, step S3, establishing a dual-reservoir fusion training framework, specifically includes the following:

[0079] S3.1 Model Construction: Based on the scale and development maturity of the reservoir data, the F15_1_C reservoir is designated as the source reservoir, and the data reservoir as the target reservoir. A pre-training unit is established for the source reservoir, and a knowledge-adapted unit is established for the target reservoir. Both share the feature input layer and CatBoost tree structure framework; only the model parameters are adapted based on the features of the source / target domain data.

[0080] S3.2 Parameter Transfer Strategy: The core parameters of the source domain pre-trained model are transferred as initial values ​​to the target domain model. Transferred parameters include: feature importance weights (e.g., average bottom hole pressure weight = 0.23, average bottom hole temperature weight = 0.18, average tubing pressure differential weight = 0.21, etc.), decision tree splitting thresholds, leaf node output values, and tree structure parameters. Through parameter transfer, an optimized initial state is provided for the target domain model, reducing the difficulty of training with small samples.

[0081] S3.3 Knowledge Adaptation Strategy: A "pre-training-fine-tuning" strategy is adopted to transfer knowledge from the source reservoir domain to the target reservoir model. Specifically, the tree structure parameters of the source domain model are frozen, and only the output values ​​of the leaf nodes are fine-tuned; the model is adapted and optimized using the target reservoir training set to ensure that the model adapts to the production patterns of the target reservoir. The fine-tuning learning rate is set to 0.01, lower than the pre-training learning rate, to avoid catastrophic forgetting.

[0082] S3.4 Model Update Mechanism: During the target reservoir model adaptation training, the validation set loss is calculated every 50 iterations. If the validation set loss does not decrease for 50 consecutive iterations (i.e., loss_t - loss_{t-50} > -0.001), training is stopped, and the optimal model weights are saved. This early stopping mechanism effectively prevents overfitting and improves the model's generalization ability.

[0083] In this embodiment, step S4 completes the CatBoost model pre-training in the source reservoir, specifically including the following:

[0084] S4.1 Data partitioning: The 689 samples in the F15_1_C reservoir training set were divided into a training subset of 551 samples and a validation subset of 138 samples in an 8:2 ratio.

[0085] S4.2 Training Process: Input the training subset into the CatBoost model with hyperparameters: n_estimators = 1500, learning_rate = 0.02, max_depth = 7. Iterate through the training process, aiming to minimize the MSE, until the validation loss converges. Monitor performance on the validation subset after each training round. Stop training when the validation set RMSE shows no decrease for 50 consecutive rounds.

[0086] S4.3 Training Results: The source domain model demonstrated loss convergence after 1200 iterations. The final training results were: MSE = 892.5, RMSE = 29.87m. 3 / d, MAPE = 6.8%, the feature importance ranking is: average tubing differential pressure (0.23) > average bottom hole pressure (0.21) > average choke size percentage (0.19) > average annular pressure (0.15) > average wellhead temperature (0.12), model fit goodness R 2=0.89, save the model as a pre-trained model file "source_model.pkl".

[0087] In this embodiment, step S5, based on the transfer framework, performs target reservoir adaptation training and prediction, specifically including the following:

[0088] S5.1 Model initialization: Load the source domain pre-trained model "source_model.pkl", initialize the target domain model, and reuse the tree structure and feature weights of the source domain model;

[0089] S5.2 Data partitioning: The reservoir data is divided into a training subset of 2375 data points and a test subset of 1000 data points in a 7:3 ratio;

[0090] S5.3 Fine-tuning process: Using training rules consistent with the source domain, perform 300 rounds of adaptation training on the target reservoir training set, with a batch size of 64; only update the output values ​​of leaf nodes, and freeze the tree structure parameters; when the RMSE of the test set is stable at 30-33m for 50 consecutive rounds... 3 Stop fine-tuning when the range is / d;

[0091] S5.4 Fine-tuning results: The target domain model parameters converged, and the leaf node weights were updated to adapt to the target reservoir data characteristics (e.g., the average pressure difference weight of the tubing was fine-tuned to 0.26), and the model's fit to the target reservoir data was significantly improved.

[0092] S5.5 Model Prediction: Use the trained target reservoir adaptation model to predict oil production for 1000 samples in the test set. The output production prediction results are stored as "prediction_results.csv", which includes the GAMMA_PREDICTED and GAMMA_ACTUAL fields.

[0093] S5.6 Evaluation Metric Calculation: Obtain the model's predicted output values ​​and the actual output values ​​on the test set, and evaluate them using three evaluation metrics:

[0094] - Root mean square error (RMSE) = 31.5m 3 / d

[0095] -Mean percentage error (MAPE) = 4.1%

[0096] - Coefficient of determination R 2 =0.972

[0097] The formulas for calculating the evaluation indicators are as follows:

[0098]

[0099] in, is the average actual oil production, and n is the total number of samples (1000).

[0100] In this embodiment, step S6 determines the optimal model through multi-dimensional comparison, specifically including the following:

[0101] S6.1 Comparison of error magnitude: The RMSE and MAPE of the target domain fine-tuning model are reduced by 53.9% and 67.2% respectively compared with the traditional single reservoir CatBoost model, and the residual variance is reduced from 4664.89 to 992.25, and the prediction error is significantly reduced;

[0102] S6.2 Trend Fitting Accuracy: The target domain fine-tuning model can accurately capture the trend of reservoir production change, and the overlap between the predicted curve and the actual curve is ≥95%. In the high fluctuation range, the prediction error is significantly smaller than that of the traditional model.

[0103] S6.3 Performance Comparison Analysis: To verify the effectiveness of the method of this invention, it is compared with the traditional single reservoir CatBoost model and LSTM deep learning model. The results are shown in Table 2:

[0104] Table 2 Performance Comparison of Different Models

[0105]

[0106] As shown in Table 2, compared with the traditional model, the present invention achieves significant improvements in all evaluation indicators, among which R... 2 The average value increased by approximately 14.7%-18.9%, RMSE decreased by 48.6%-58.5%, and MAPE decreased by 67.2%-70.4%.

[0107] S6.4 Model Optimization Decision: Based on the model evaluation index results in step S5, an analysis and comparison are conducted from multiple perspectives, including error control, trend fitting, and feature adaptability. In this embodiment, the RMSE is 31.5m. 3 / d (less than 35m) 3 / d threshold), MAPE is 4.1% (below the 5% threshold), R 2 The value is 0.972, the predicted curve has a high degree of overlap with the actual curve, and the importance weight of the core production characteristics (average tubing pressure difference, average bottom hole pressure, and average choke size percentage) ranks in the top 3, meeting the criteria for the optimal model. Therefore, this model is determined to be the final optimal oil production prediction model.

[0108] Figure 3 This is a performance comparison chart between the model of this invention and the traditional model. Figure 4 A scatter plot of predicted and actual values. Figure 5 This is a distribution map of prediction errors. Figure 6The figure shows the feature importance. As can be seen from the figure, the dual reservoir fusion training model of the present invention can more accurately capture the trend of reservoir production changes, and the prediction error is significantly smaller than that of traditional models in the high fluctuation range.

[0109] In summary, the high-precision oil production prediction method for multiple reservoirs based on common feature alignment and CatBoost ensemble learning proposed in this invention achieves high-precision oil production prediction through common feature alignment, standardized preprocessing, dual-reservoir data fusion, and an ensemble learning model. This method demonstrates superior prediction accuracy and stability in multi-reservoir collaborative prediction tasks, validating the effectiveness and engineering applicability of the proposed method.

Claims

1. A method for multi-reservoir production prediction based on CatBoost and feature alignment, characterized in that, Includes the following steps: S1. Collect historical oil production data from different oil fields, and automatically align and preprocess the collected data based on common features to construct training and test sets for each oil field. S2. Construct a time-series oil production prediction model based on the CatBoost algorithm as the core tool for multi-reservoir production prediction; S3. Build a dual-reservoir data fusion training framework, determine the knowledge transfer path between the source reservoir and the target reservoir, and initialize the model parameter transfer strategy and adaptation rules. S4. Based on the CatBoost model constructed in step S2, pre-train it on the source reservoir training set and extract the domain knowledge features in the model. S5. Based on the fusion training framework built in step S3, the domain knowledge of the source reservoir pre-trained model is transferred to the target reservoir model, and the model is adapted and trained on the target reservoir training set; the adapted target reservoir model is used to predict the production of the test set and obtain the model evaluation index results. S6. Based on the evaluation index results of the model in step S5, analyze and compare from multiple perspectives such as error control, trend fitting, and feature adaptability to determine the final optimal oil production prediction model.

2. The method of claim 1, wherein, The data preprocessing in step S1 includes: S1.1 Perform preliminary cleaning on the collected historical oil production and related data of multiple oil reservoirs, retain numerical feature columns, and remove string fields, invalid and redundant columns and completely missing fields; S1.

2. The three-standard-deviation method is used to identify and remove outlier data. Linear interpolation is used to process missing values, and cubic Lagrange interpolation is used to correct abnormal negative values ​​of daily oil production. S1.

3. Extract year-day and month features based on the date field, and use StandardScaler to standardize the features and target variables. Fit the standardized parameters to the source reservoir data, and reuse the parameters for the target reservoir.

3. The method of claim 1, wherein, The automatic alignment process for common features in step S1 is as follows: S1.

1. Through set intersection operation, the numerical features common to the two reservoirs are automatically extracted to form a common feature set, ensuring the consistency of feature dimensions in cross-reservoir modeling; S1.

2. Daily oil production is uniformly used as the target prediction variable. If the field names of the target variables differ among different oil reservoirs, they are uniformly named through a field mapping table.

4. The method according to claim 1, characterized in that, The process of constructing a time-series prediction model for oil production in step S2 includes: S2.

1. The CatBoost regression algorithm is used as the core framework to construct an oil production prediction model for multiple oil reservoirs. The model consists of an input feature layer, a tree node splitting layer, a feature importance calculation layer, and an output prediction layer. S2.2, Set model parameters: including number of iterations, learning rate, decision tree depth, and regularization coefficient; S2.3 The model optimizer uses the CatBoost built-in optimizer for iterative parameter updates, and the training loss function is the average... S2.3 The model's optimizer uses the built-in CatBoost optimizer for iterative parameter updates, and the training loss function is the mean squared error.

5. The method according to claim 1, characterized in that, In step S3, a dual-reservoir data fusion training framework is established, the knowledge transfer paths between the source and target reservoirs are determined, and the model parameter transfer strategy and adaptation rules are initialized. This further includes: S3.1 Based on the scale and development maturity of reservoir data, divide the reservoirs into source reservoirs and target reservoirs, establish pre-training units for source reservoirs, and establish knowledge adaptation units for target reservoirs; S3.2 Source Reservoir Knowledge Extraction: After the source reservoir pre-training unit completes the initial training based on the training data, it extracts core domain knowledge, including feature importance weights, model hyperparameters, etc. S3.3 Knowledge Adaptation: The "pre-training-fine-tuning" strategy is adopted to transfer knowledge from the source reservoir domain to the target reservoir model. The model is then adapted and optimized using the target reservoir training set to ensure that the model is adapted to the production patterns of the target reservoir. S3.4 Model Update: During the target reservoir model adaptation training process, the validation set loss is calculated periodically. If there is no decrease for several consecutive rounds, training is stopped and the optimal model weights are saved.

6. The method according to claim 1, characterized in that, In step S5, knowledge transfer and model adaptation training are completed based on the dual-reservoir fusion training framework. A pre-trained model is trained in the source reservoir and adapted to the target reservoir. The trained target reservoir adaptation model is used to predict production on the test set, and model evaluation calculations are performed to obtain the model evaluation index calculation results. This includes the following sub-steps: S5.

1. Based on the fusion training framework, complete the source model pre-training and knowledge transfer: train the CatBoost pre-trained model on the source reservoir training set until the model loss converges, determine the end of training and record the number of training times. Based on the transfer strategy in step S3, the knowledge of the source reservoir domain is transferred to the initial model of the target reservoir, and adaptation training is performed on the training set of the target reservoir. The number of adaptation trainings matches the number of trainings of the source model. S5.2 Use the trained target reservoir adaptation model to predict oil production on the test set and output the production prediction results; S5.3, Obtain the yield prediction value of the model and the true yield value of the test set, and use three complementary evaluation indexes: root mean square error RMSE, mean absolute percentage error MAPE and determination coefficient R 2 The prediction results are evaluated, and the formulas are as follows: Among them, y i This represents the actual oil production value. This represents the oil production value predicted by the model. denoted as the average actual oil production, and n represents the sample size.

7. The method according to claim 1, characterized in that, In step S6, the calculation results of the evaluation indicators of the target reservoir adaptation model are analyzed. The predictive performance of the model is evaluated from multiple perspectives, including error control, trend fitting, and feature adaptability. Finally, the optimal prediction model is comprehensively selected. Specifically, this includes: S6.1 Error control dimension: Select models with lower root mean square error and average percentage error; S6.2 Trend Fitting Dimension: Judged by the degree of overlap between the predicted curve and the actual curve; S6.3 Feature Adaptability Dimension: Ensure that the importance weight of core production features in the target reservoir model is ranked high.

8. The method according to any one of claims 1-7, characterized in that, The method has the following advantages: S8.1, Automatic alignment mechanism for common features: Automatically extracts common features from different reservoirs through set operations, solving the modeling error problem caused by inconsistent feature quantities across reservoirs, without manual intervention, thus improving modeling efficiency; S8.2 Dual reservoir data fusion training strategy: The fusion training mode of "source reservoir pre-training + target reservoir fine-tuning" is adopted to fully reuse the general production knowledge of the source reservoir, alleviate the problem of target reservoir data scarcity, and improve the model generalization ability. S8.3, Robust Model Structure and Efficient Training: The CatBoost ensemble learning model is used as the core predictor, which has the advantages of anti-overfitting, stable prediction, and less prone to flattening predictions, making it suitable for heterogeneous data scenarios with multiple reservoirs. In summary, this invention proposes a high-precision oil production prediction method for multiple reservoirs based on common feature alignment and CatBoost ensemble learning. Through common feature alignment, standardized preprocessing, dual-reservoir data fusion, and an ensemble learning model, high-precision oil production prediction is achieved. This method maintains high model training efficiency while fully considering temporal characteristics, trend information, and domain differences between reservoirs, enabling knowledge transfer and generalization of the model from single-reservoir to multi-reservoir scenarios.

Images