Petroleum yield prediction model construction of domain knowledge migration under multi-reservoir scene

Abstract

The invention discloses construction of a petroleum yield prediction model for domain knowledge migration in a multi-reservoir scene, and belongs to the technical field of petroleum engineering and intelligent prediction. According to the method, a domain knowledge migration mechanism is introduced, so that the model can share depth feature information among multiple oil reservoirs, the prediction precision and model robustness of small sample oil reservoirs are effectively improved, and the problems of poor model generalization, high training cost, unstable prediction and the like in a traditional method are solved. The method is clear in structure and moderate in calculation amount, has good interpretability and visualization ability, and can be widely applied to the fields of oil reservoir yield prediction, intelligent oil field management, production optimization decision making and the like.

Description

Construction of Oil Production Prediction Model Based on Domain Knowledge Transfer in Multi-Reservoir Scenarios Technical Field

[0001] This invention belongs to the interdisciplinary field of oil and gas field development engineering and artificial intelligence, specifically involving the construction of an oil production prediction model based on domain knowledge transfer in multi-reservoir scenarios. Background Technology

[0002] As a crucial energy source and chemical raw material, petroleum production fluctuations directly impact oilfield development efficiency and energy supply-demand balance. Traditional petroleum production forecasting primarily relies on empirical formulas for reservoir engineering, 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, utilizing machine learning methods for petroleum production forecasting has become a research hotspot. Among these, deep learning and ensemble learning models can automatically learn nonlinear relationships from large amounts of historical data, significantly improving prediction accuracy. This approach ensures both model accuracy and stability while reducing the cost of modeling small-sample reservoirs, exhibiting good scalability and engineering application value. Summary of the Invention

[0003] To overcome the shortcomings of the existing technology, the present invention aims to provide a method for constructing an oil production prediction model based on domain knowledge transfer in multi-reservoir scenarios. This method protects data privacy while alleviating the problem of data scarcity, and utilizes a deep learning model architecture to achieve high-precision oil production prediction. Its main steps include:

[0004] 1. Construction of an oil production prediction model based on domain knowledge transfer in multi-reservoir scenarios, characterized by the following steps:

[0005] S1 collects historical oil production data from different oil fields and preprocesses the collected data to construct training and test sets for each oil field.

[0006] S2. Construct a time-series prediction model for oil production based on the XGBoost algorithm as a core prediction tool for domain knowledge transfer;

[0007] S3. Build a domain knowledge transfer 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 XGBoost model constructed in step S2, train local models in each source reservoir and extract domain knowledge features from the models.

[0009] S5. Based on the domain knowledge transfer framework built in step S3, transfer the domain knowledge of the source reservoir local model to the target reservoir model, and perform model adaptation training on the target reservoir training set; use the transferred and adapted target reservoir model and each source reservoir local model to predict production on the corresponding reservoir test set, carry out model evaluation calculation, and obtain model evaluation index results.

[0010] S6. Based on the evaluation index results of the migration adaptation model and the local model of the source reservoir 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 for each reservoir.

[0011] 2. The method according to claim 1, wherein the characteristics of the historical oil production and related data of different reservoirs collected in step S1 include date, average bottom hole pressure, average bottom hole temperature, average tubing pressure differential, average annular pressure, average choke size percentage, average wellhead pressure, average wellhead temperature, choke size pressure differential, and daily oil production.

[0012] 3. The method according to claim 1, wherein the data preprocessing in step S1 includes: performing preliminary cleaning on the collected historical oil production and related data of multiple oil reservoirs, identifying and removing abnormal data using the three-standard-deviation method; using linear interpolation to process missing values ​​in the data to ensure the continuity of the time series data; and using cubic Lagrange interpolation to correct data where the characteristic daily liquid production and daily oil production values ​​are negative, thereby ensuring the rationality of the core production data and the accuracy of subsequent model training.

[0013] 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:

[0014] S2.1. Using the XGBoost regression algorithm as the core framework, an oil production prediction model for knowledge transfer in multiple oil reservoir domains is constructed. The model consists of an input feature layer, a tree node splitting layer, a feature importance calculation layer, and an output prediction layer.

[0015] S2.3 The optimizer for the model uses the built-in optimizer of XGBoost for iterative parameter updates; the loss function is the squared error, and the evaluation metrics include root mean square error, mean absolute error, and coefficient of determination, as shown in the following formulas:

[0016]

[0017] Among them, y i This represents the actual oil production value. Let n be the predicted oil production value from the model, and n be the sample size.

[0018] According to the method described in claim 1, step S3, which involves establishing a domain knowledge transfer framework, determining the knowledge transfer paths between the source reservoir and the target reservoir, and initializing the model parameter transfer strategy and adaptation rules, further includes:

[0019] S3.1 Based on the scale and development maturity of reservoir data, source reservoirs and target reservoirs are divided. Local model training units are established for each source reservoir, knowledge adaptation units are established for the target reservoirs, and a knowledge transfer server is built. The server includes a domain knowledge storage module and a transfer strategy scheduling module.

[0020] S3.2 Source Reservoir Knowledge Extraction and Distribution: After the local model training unit of each source reservoir completes the initial training based on the training data, it extracts the core domain knowledge and transmits it to the knowledge storage module on the server side. The server side distributes the source reservoir domain knowledge with the highest matching degree to the target reservoir knowledge adaptation unit according to the similarity of the geological characteristics between the target reservoir and each source reservoir, so as to ensure that the initial knowledge of the target reservoir is compatible with the source reservoir.

[0021] S3.3 Source Model Optimization and Knowledge Update: Each source reservoir local model training unit continuously uses newly added training data to iteratively optimize the model. After every 100 rounds of iterative training, the updated feature importance weights, hyperparameter adjustment results, and other related knowledge are transmitted to the server in real time to update the source reservoir knowledge base in the knowledge storage module.

[0022] S3.4 Knowledge Screening and Adaptation: After receiving updated domain knowledge from the source reservoirs, the server first screens the knowledge, removing parameters that conflict with the geological characteristics of the target reservoir. A "feature mapping normalization" strategy is then used to adapt and transform the importance weights of the source reservoir features according to the dimensions of the target reservoir features. Next, based on the similarity weights between each source reservoir and the target reservoir, a weighted average method is used to fuse the knowledge from multiple source reservoirs, as shown in the following formula:

[0023]

[0024] Among them, W target W represents the feature importance weight of the target reservoir after fusion. source-i s represents the feature importance weight of the i-th source reservoir. i G represents the similarity weight of geological features between the i-th source reservoir and the target reservoir, M is the number of source reservoirs participating in knowledge fusion, and G is the number of source reservoirs participating in knowledge fusion. target G represents a comprehensive index of the geological characteristics of the target oil reservoir. source-i This represents a comprehensive index of the geological characteristics of the i-th source oil reservoir;

[0025] S3.5 Target Model Knowledge Deployment: Deploy the fused and adapted domain knowledge to the target reservoir knowledge adaptation unit, update the parameter configuration of the target reservoir model, and complete one domain knowledge migration cycle.

[0026] 5. The method according to claim 1, characterized in that, in step S5, knowledge transfer and model adaptation training are completed based on the domain knowledge transfer framework, local models are trained in each source reservoir and the transfer model is adapted in the target reservoir, the trained source reservoir local model and the target reservoir transfer adaptation model are used to predict the production of the corresponding reservoir test set, and model evaluation calculation is performed to obtain the model evaluation index calculation results, including the following sub-steps:

[0027] S5.1. Complete source model training and knowledge transfer based on the domain knowledge transfer framework: Train the XGBoost local model on each source reservoir training set until the model loss converges, determine the end of training and record the number of source model training times; Based on the transfer strategy in step S3, transfer the domain knowledge of the source reservoir to the initial model of the target reservoir, and perform adaptation training on the target reservoir training set. The number of adaptation training times is consistent with the number of source model training times to ensure the fairness of the control experiment.

[0028] S5.2. Use the trained source reservoir local model to predict oil production on the corresponding source reservoir test set and the target reservoir migration adaptation model to predict oil production on the target reservoir test set, and output the production prediction results for each reservoir.

[0029] S5.3 Obtain the production prediction values ​​of the source reservoir local model and the target reservoir migration adaptation model, as well as the actual production values ​​of the corresponding reservoirs. Use three complementary evaluation indicators—root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²)—to evaluate and calculate the predicted oil production data and the actual oil production values.

[0030] In step S6, the calculation results of the evaluation indicators of the source reservoir local model and the target reservoir migration adaptation model are analyzed and compared. The predictive performance of the model is evaluated from multiple perspectives, including error control, trend fitting, and feature adaptability, and the final optimal model is obtained. Among them, the error control dimension prioritizes models with RMSE below 50 and MAE below 30. The trend fitting dimension is judged by the overlap between the predicted curve and the actual curve. The feature adaptability dimension ensures that the importance weight of the core production features in the target reservoir model ranks in the top 3. Finally, the optimal prediction model for each reservoir is comprehensively selected.

[0031] The advantages of this invention compared to the prior art are mainly reflected in:

[0032] (1) Strong ability to collaborate on multi-source reservoir data and transfer domain knowledge: This invention achieves feature sharing and model transfer across reservoirs by jointly modeling and transferring historical production data from multiple reservoirs. Compared with traditional single-reservoir modeling methods, this invention can significantly improve the generalization performance and adaptability of predictions by relying on the characteristic patterns of other reservoirs in reservoirs with scarce data.

[0033] (2) Comprehensive construction of time series features, capable of capturing complex dynamic patterns: This invention introduces lag features, moving statistical features, exponential weighted average features, as well as sine and cosine seasonal features and trend features, forming a multi-scale, multi-dimensional time series feature system that can comprehensively depict the periodic changes, trend fluctuations and short-term disturbances of oil production over time, thereby enhancing the model's time series expression capability.

[0034] (3) Robust model structure and efficient training: The XGBoost gradient boosting tree model is used as the core predictor, which features high accuracy, resistance to overfitting, and strong interpretability. The model has high training efficiency and can quickly complete the training and optimization of a large amount of reservoir data while ensuring accuracy, making it suitable for real-time prediction needs in engineering sites.

[0035] In summary, this invention proposes a petroleum production prediction model based on domain knowledge transfer in multi-reservoir scenarios. Through multi-source reservoir data fusion, time-series feature construction, and an XGBoost-based ensemble learning model, high-precision petroleum 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 from single-reservoir to multi-reservoir scenarios. Attached Figure Description

[0036] 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.

[0037] Figure 1 is an overall flowchart of the method of the present invention;

[0038] Figure 2 is a schematic diagram of the structure of a neural network based on self-attention;

[0039] Figure 3 is a schematic diagram of the transfer learning framework;

[0040] Figure 4 shows the trend of model output prediction. Detailed Implementation

[0041] Example 1:

[0042] Referring to Figure 1, the construction of the oil production prediction model for multi-reservoir scenarios according to the present invention includes the following steps:

[0043] S1. Collect historical oil production data from different oil reservoirs, perform time formatting, missing value imputation, and outlier detection on the collected data. Construct training and testing sets for each oilfield.

[0044] S2. Construct a time-series oil production prediction model based on a self-attention mechanism, which can be used as a tool for implementing transfer learning algorithms;

[0045] S3. Based on the model constructed in step S2, a transfer learning framework is built, using the XGBoost model as the core prediction model, and modeling is performed using the training set data. Iterative training is conducted using standardized training data, and model performance is monitored on the validation set to prevent overfitting.

[0046] S4. Apply the trained model to the test set data, calculate the root mean square error, mean absolute error, and coefficient of determination, and comprehensively evaluate the model's prediction accuracy and stability.

[0047] S5. After completing the model training, construct future time series input features using historical data from the last 30 days, including lag terms, moving statistical features, and the mean and variance features of sensor parameters, and predict the trend of oil production changes in the next 30 days using the trained model.

[0048] S6. Analyze and compare the calculation results of the model evaluation indicators in step S5 to determine the final prediction model.

[0049] In this embodiment, step S1 involves real-time collection of relevant historical oil production data, including parameters such as date, daily liquid production, pump inlet pressure, porosity, permeability, saturation, horizontal section length, pressure breakage, and daily oil production. Next, test and training sets are divided based on four oilfield datasets. The collected data is then preprocessed as follows:

[0050] S1.1. Perform preliminary cleaning on the collected data to remove outliers exceeding three times the standard deviation; Table 1 shows a partial sample of processed data from an oil field:

[0051] Table 1 Example of Historical Oil Production Sample Data

[0052]

[0053] S1.2. The collected data is initially cleaned, and missing values ​​are processed by linear interpolation to ensure the continuity of the time series data. For data with negative daily liquid and oil production values, cubic Lagrange interpolation is used to correct them to ensure the rationality of the core production data and the accuracy of subsequent model training.

[0054] In this embodiment, step S2 involves constructing a time-series oil production prediction model based on a self-attention mechanism. This model aims to capture the dynamic patterns of oil production parameters and other characteristics changing over time. Figure 2 shows the structure of this model. The model construction specifically includes the following sub-steps:

[0055] S2.1. Using the XGBoost regression algorithm as the core framework, an oil production prediction model for knowledge transfer in multiple oil reservoir domains is constructed. The model consists of an input feature layer, a tree node splitting layer, a feature importance calculation layer, and an output prediction layer.

[0056] S2.2 Setting model parameters: The objective function is squared error regression, the number of trees is 1200, the learning rate is 0.01, the tree depth is 5, the sample sampling rate is 0.85, the feature sampling rate is 0.85, the regularization parameter is 0.2, the L1 regularization is 0.05, the L2 regularization is 0.5, the random seed is 42, and the number of parallel computing threads n_jobs is set to -1;

[0057] S2.3 The optimizer for the model uses the built-in optimizer of XGBoost for iterative parameter updates; the loss function is the squared error, and the evaluation metrics include root mean square error, mean absolute error, and coefficient of determination, as shown in the following formulas:

[0058]

[0059] Where yi represents the actual oil production value. Let n be the predicted oil production value from the model, and n be the sample size.

[0060] In this embodiment, step S3 involves building a domain knowledge transfer framework (see Figure 3), determining the knowledge transfer path between the source reservoir and the target reservoir, and initializing the model parameter transfer strategy and adaptation rules. Specifically, this includes the following:

[0061] S3.1 Based on the scale and development maturity of reservoir data, source reservoirs and target reservoirs are divided. Local model training units are established for each source reservoir, knowledge adaptation units are established for the target reservoirs, and a knowledge transfer server is built. The server includes a domain knowledge storage module and a transfer strategy scheduling module.

[0062] S3.2 Source Reservoir Knowledge Extraction and Distribution: After the local model training unit of each source reservoir completes the initial training based on the training data, it extracts the core domain knowledge and transmits it to the knowledge storage module on the server side. The server side distributes the source reservoir domain knowledge with the highest matching degree to the target reservoir knowledge adaptation unit according to the similarity of the geological characteristics between the target reservoir and each source reservoir, so as to ensure that the initial knowledge of the target reservoir is compatible with the source reservoir.

[0063] S3.3 Source Model Optimization and Knowledge Update: Each source reservoir local model training unit continuously uses newly added training data to iteratively optimize the model. After every 100 rounds of iterative training, the updated feature importance weights, hyperparameter adjustment results, and other related knowledge are transmitted to the server in real time to update the source reservoir knowledge base in the knowledge storage module.

[0064] S3.4 Knowledge Screening and Adaptation: After receiving updated domain knowledge from the source reservoirs, the server first screens the knowledge, removing parameters that conflict with the geological characteristics of the target reservoir. A "feature mapping normalization" strategy is then used to adapt and transform the importance weights of the source reservoir features according to the dimensions of the target reservoir features. Next, based on the similarity weights between each source reservoir and the target reservoir, a weighted average method is used to fuse the knowledge from multiple source reservoirs, as shown in the following formula:

[0065]

[0066] Among them, W target W represents the feature importance weight of the target reservoir after fusion. source-i s represents the feature importance weight of the i-th source reservoir. i G represents the similarity weight of geological features between the i-th source reservoir and the target reservoir, M is the number of source reservoirs participating in knowledge fusion, and G is the number of source reservoirs participating in knowledge fusion. target G represents a comprehensive index of the geological characteristics of the target oil reservoir. source-i This represents a comprehensive index of the geological characteristics of the i-th source oil reservoir;

[0067] S3.5 Target Model Knowledge Deployment: Deploy the fused and adapted domain knowledge to the target reservoir knowledge adaptation unit, update the parameter configuration of the target reservoir model, and complete one domain knowledge migration cycle.

[0068] In this embodiment, step S5 involves completing knowledge transfer and model adaptation training based on a domain knowledge transfer framework. Local models are trained in each source reservoir, and the transfer model is adapted in the target reservoir. The trained source reservoir local models and target reservoir transfer adaptation models are used to predict production on the corresponding reservoir test sets, and model evaluation calculations are performed to obtain the model evaluation index calculation results. This includes the following sub-steps:

[0069] S5.1. Complete source model training and knowledge transfer based on the domain knowledge transfer framework: Train the XGBoost local model on each source reservoir training set until the model loss converges, determine the end of training and record the number of source model training times; Based on the transfer strategy in step S3, transfer the domain knowledge of the source reservoir to the initial model of the target reservoir, and perform adaptation training on the target reservoir training set. The number of adaptation training times is consistent with the number of source model training times to ensure the fairness of the control experiment.

[0070] S5.2. Use the trained source reservoir local model to predict oil production on the corresponding source reservoir test set and the target reservoir migration adaptation model to predict oil production on the target reservoir test set, and output the production prediction results for each reservoir.

[0071] S5.3 Obtain the production predictions of the source reservoir local model and the target reservoir migration adaptation model, along with the actual production values ​​of the corresponding reservoirs. Use three complementary evaluation indicators: root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²). 2 The predicted oil production data and the actual oil production data are evaluated and calculated.

[0072] S5.4 Obtain the predicted and actual production values ​​for each oilfield using the global and local models. Calculate and evaluate the predicted and actual oil production data using three complementary evaluation indicators: relative error, coefficient of determination, and mean square error. The model evaluation parameter formulas are as follows:

[0073]

[0074] Among them, y i For the true value, For predicted values, is the mean of the actual values, and n is the total number of samples;

[0075] In this embodiment, step S6 comprehensively evaluates and compares the prediction results of different reservoir area models, evaluates the prediction effect of the model from multiple perspectives such as error distribution, trend fitting and relative accuracy, and determines the final optimal model structure and parameter configuration.

[0076] For historical production data from different reservoirs, the same time-series feature construction method and XGBoost model configuration were used for training and validation. Mean absolute error, coefficient of determination, and mean square error were calculated on the test set. The evaluation results of each reservoir model are shown in the table below:

[0077]

[0078] Figure 4 is a schematic diagram comparing the actual and predicted production of the model of the present invention on the reservoir test set.

[0079] As shown in the figure, the unified transfer model of this invention can more accurately capture the trend of reservoir production changes, and its prediction error is significantly smaller than that of the independently trained models in the high fluctuation range. The results indicate that by introducing multi-scale time series features and a domain knowledge transfer mechanism, the model of this invention can achieve higher fitting accuracy and stronger generalization ability in multi-reservoir data scenarios. Compared with the single-reservoir independently trained model, this invention achieves significant improvements in all evaluation metrics, including R...2 The average value increased by about 5% to 8%, while MAE and MSE decreased by more than 60% respectively.

[0080] In summary, the oil production prediction method based on time series feature construction and transfer modeling proposed in this invention demonstrates superior prediction accuracy and stability in multi-reservoir collaborative prediction tasks, verifying the effectiveness and engineering applicability of the method.

Claims

1. Construction of an oil production prediction model based on domain knowledge transfer in multi-reservoir scenarios, characterized in that, Includes the following steps: S1 collects historical oil production data from different oil fields and preprocesses the collected data to construct training and test sets for each oil field. S2. Construct a time-series oil production prediction model based on the XGBoost algorithm as the core prediction tool for domain knowledge transfer; S3. Build a domain knowledge transfer 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 XGBoost model constructed in step S2, train local models in each source reservoir and extract domain knowledge features from the models. S5. Based on the domain knowledge transfer framework built in step S3, transfer the domain knowledge of the source reservoir local model to the target reservoir model, and perform model adaptation training on the target reservoir training set; use the transferred and adapted target reservoir model and each source reservoir local model to predict production on the corresponding reservoir test set, carry out model evaluation calculation, and obtain model evaluation index results; S6. Based on the evaluation index results of the transferred and adapted model and the source reservoir local 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 for each reservoir.

2. The method according to claim 1, characterized in that, The characteristics of historical oil production and related data collected in step S1 for different reservoirs include date, average bottom hole pressure, average bottom hole temperature, average tubing pressure differential, average annular pressure, average choke size percentage, average wellhead pressure, average wellhead temperature, choke size pressure differential, and daily oil production.

3. The method according to claim 1, characterized in that, The data preprocessing in step S1 includes: preliminary cleaning of the collected historical oil production and related data from multiple reservoirs; identification and removal of outlier data using the three-standard-deviation method; linear interpolation processing for missing values ​​in the data to ensure the continuity of the time series data; and correction of data with negative daily liquid and oil production values ​​using cubic Lagrange interpolation to ensure the rationality of core production data and the accuracy of subsequent model training.

4. The method according to claim 1, characterized in that, Step S2 involves constructing a time-series oil production prediction model, which includes: S2.1 Using the XGBoost regression algorithm as the core framework, a knowledge transfer-oriented oil production prediction model for multiple oil reservoirs is constructed. 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 Setting model parameters: the objective function is squared error regression, the number of trees is 1200, the learning rate is 0.01, the tree depth is 5, the sample sampling rate is 0.85, the feature sampling rate is 0.85, the regularization parameter is 0.2, L1 regularization is 0.05, L2 regularization is 0.5, the random seed is 42, and the number of parallel computing threads n_jobs is set to -1; S2.3 The model optimizer uses the built-in XGBoost optimizer for iterative parameter updates; the loss function is squared error, and the evaluation metrics include root mean square error, mean absolute error, and coefficient of determination, with the formulas as follows: Among them, y i This represents the actual oil production value. Let n be the predicted oil production value from the model, and n be the sample size.

5. The method according to claim 1, characterized in that, In step S3, a domain knowledge transfer framework is established, the knowledge transfer paths between source and target reservoirs are determined, and model parameter transfer strategies and adaptation rules are initialized. This further includes: S3.1, dividing source and target reservoirs based on reservoir data scale and development maturity, establishing local model training units for each source reservoir, establishing knowledge adaptation units for the target reservoir, and building a knowledge transfer server. The server includes a domain knowledge storage module and a transfer strategy scheduling module. S3.2, Source reservoir knowledge extraction and distribution: After initial training based on training data, the local model training units of each source reservoir extract core domain knowledge and transmit it to the server-side knowledge storage module. The server distributes the source reservoir domain knowledge with the highest matching degree to the target reservoir knowledge adaptation unit based on the similarity of geological features between the target reservoir and each source reservoir. S3.3, Source Model Optimization and Knowledge Update: Each source reservoir's local model training unit continuously uses newly added training data to iteratively optimize the model. After every 100 rounds of iteration, the updated feature importance weights, hyperparameter adjustment results, and other domain knowledge are transmitted to the server in real time to update the source reservoir knowledge base in the knowledge storage module. S3.4, Knowledge Screening and Adaptation: After receiving the updated domain knowledge from the source reservoirs, the server first screens the knowledge, removing parameters that conflict with the geological characteristics of the target reservoir. It then uses a "feature mapping normalization" strategy to adapt and convert the source reservoir feature importance weights according to the target reservoir's feature dimensions. Finally, based on the similarity weights between each source reservoir and the target reservoir, a weighted average method is used to fuse the knowledge from multiple source reservoirs, as shown in the following formula: Among them, W target W represents the feature importance weight of the target reservoir after fusion. source-i s represents the feature importance weight of the i-th source reservoir. i G represents the similarity weight of geological features between the i-th source reservoir and the target reservoir, M is the number of source reservoirs participating in knowledge fusion, and G is the number of source reservoirs participating in knowledge fusion. target G represents a comprehensive index of the geological characteristics of the target oil reservoir. source-i S3.5, Target Model Knowledge Deployment: Deploy the fused and adapted domain knowledge to the target reservoir knowledge adaptation unit, update the parameter configuration of the target reservoir model, and complete one domain knowledge migration cycle.

6. The method according to claim 1, characterized in that, In step S5, knowledge transfer and model adaptation training are completed based on the domain knowledge transfer framework. Local models are trained in each source reservoir and the transfer model is adapted in the target reservoir. The trained source reservoir local model and target reservoir transfer adaptation model are used to predict the production of the corresponding reservoir test set and perform model evaluation calculation to obtain the model evaluation index calculation results. The steps include the following sub-steps: S5.1, Source model training and knowledge transfer are completed based on the domain knowledge transfer framework: XGBoost local model is trained on each source reservoir training set until the model loss converges, the training is determined to be over and the number of source model training times is recorded. 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 training times is the same as the number of training times of the source model to ensure the fairness of the control experiment. S5.

2. Using the trained source reservoir local model on the corresponding source reservoir test set and the target reservoir migration adaptation model on the target reservoir test set, respectively, predict oil production and output the production prediction results for each reservoir; S5.

3. Obtain the production prediction values ​​of the source reservoir local model and the target reservoir migration adaptation model and the actual production values ​​of the corresponding reservoirs, and use three complementarity evaluation indicators: root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²). 2 The predicted oil production data and the actual oil production data are evaluated and calculated.

7. The method according to claim 1, characterized in that, In step S6, the calculation results of the evaluation indicators of the source reservoir local model and the target reservoir migration adaptation model are analyzed and compared. The predictive performance of the model is evaluated from multiple perspectives, including error control, trend fitting, and feature adaptability, and the final optimal model is obtained. Among them, the error control dimension prioritizes models with RMSE below 50 and MAE below 30. The trend fitting dimension is judged by the overlap between the predicted curve and the actual curve. The feature adaptability dimension ensures that the importance weight of the core production features in the target reservoir model ranks in the top 3. Finally, the optimal prediction model for each reservoir is comprehensively selected.

Images