Carbon dioxide flooding production and absorption profile prediction method based on random sampling and RNN (Recurrent Neural Network)

Abstract

The invention discloses a carbon dioxide flooding production and absorption profile prediction method based on random sampling and RNN, and relates to the technical field of oil and gas field development, and the method comprises the steps: S1, constructing a high-fidelity carbon dioxide production and absorption profile data set; s2, performing preprocessing operation on the data; s3, constructing an RNN model with physical constraints; s4, dynamically correcting the prediction deviation through error inversion, and optimizing the model; s5, industrial-grade real-time decision support is carried out by deploying an online prediction system; according to the carbon dioxide flooding production and absorption profile prediction method based on random sampling and RNN provided by the invention, the problems of low efficiency, low precision and poor interpretability of an existing carbon dioxide flooding production and absorption profile prediction technology are solved.

Description

A method for predicting carbon dioxide displacement absorption profiles based on random sampling and RNN. Technical Field

[0001] This invention relates to the field of oil and gas field development technology, and in particular to a method for predicting carbon dioxide flooding production and absorption profiles based on random sampling and RNN. Background Technology

[0002] Carbon dioxide flooding (CFD) is a key technology for enhancing oil and gas recovery (EOR) in oil and gas field development, particularly suitable for low-permeability reservoirs, high-water-cut reservoirs, and depleted reservoirs where conventional waterflooding has proven ineffective. It plays an irreplaceable role in improving oil recovery and extending reservoir lifespan. In the industrial application of CFD, the reservoir's production-absorption profile is a core factor determining the effectiveness and economics of the development plan. Accurate prediction of this profile is crucial for optimizing injection and production parameters, mitigating gas channeling risks, and ensuring efficient reservoir development.

[0003] For predicting the production and absorption profiles of carbon dioxide flooding, existing technologies mainly fall into three categories: First, traditional numerical simulation methods, which rely heavily on commercial software such as Eclipse for reservoir simulation. These methods require the establishment of complex geological models, and a single calculation can take tens of hours, resulting in extremely low computational efficiency and failing to meet the real-time decision-making needs of oilfield development. Second, traditional empirical formula methods, which fit statistical patterns based on historical development data and have the advantage of fast computation speed, but have a very narrow scope of application. Because they do not fully consider key influencing factors such as reservoir heterogeneity and changes in injected phase, they can only be adapted to specific geological conditions and development regimes. Once the operating conditions change, the prediction accuracy will drop sharply, making it difficult to meet the development needs of complex reservoirs. Third, machine learning prediction methods. Currently, there is a scarcity of dedicated datasets for predicting the production and absorption profiles of carbon dioxide flooding in the publicly available literature, with sample sizes generally below 4,000 sets. Moreover, existing models mostly adopt a purely data-driven construction approach, lacking geological mechanism constraints, and the prediction results are prone to deviating from physical laws, making it difficult to guarantee the reliability of the predictions.

[0004] In summary, existing carbon dioxide flooding oil production profile prediction technologies have significant defects and shortcomings: First, they suffer from efficiency bottlenecks; traditional numerical simulations and empirical formula methods cannot meet the needs of dynamic adjustments to injection-production regimes, making it difficult to support rapid iteration of oilfield development plans. Second, they have limitations in accuracy; purely data-driven machine learning models are prone to overfitting and cannot guarantee that the prediction results conform to fundamental physical laws such as the law of conservation of mass, resulting in poor prediction accuracy stability. Third, they have poor interpretability; machine learning black-box models struggle to pinpoint the source of prediction errors and cannot provide effective guidance for practical development measures such as reservoir well location adjustment and stratified control. Summary of the Invention

[0005] The purpose of this invention is to provide a method for predicting the production and absorption profile of carbon dioxide flooding based on random sampling and RNN, so as to solve the problems of low efficiency, low accuracy and poor interpretability of existing carbon dioxide flooding oil production and absorption profile prediction technologies.

[0006] To achieve the above objectives, this invention provides a method for predicting carbon dioxide flooding production and absorption profiles based on random sampling and RNN, comprising the following steps: S1, constructing a high-fidelity carbon dioxide production and absorption profile dataset; S2, performing preprocessing operations on the data; S3, constructing a physically constrained RNN model; S4, dynamically correcting prediction bias and optimizing the model through error inversion; S5, providing industrial-grade real-time decision support by deploying an online prediction system.

[0007] Preferably, in S1: S11, determine input variables: select static parameters and dynamic development parameters characterizing reservoir heterogeneity; there are 21 static parameters and 30 dynamic development parameters, which together constitute a 51-dimensional parameter space; S12, set parameter boundaries: based on the geological characteristics of the target reservoir, set the value range of the variables, including rock porosity, rock permeability, reservoir thickness, and injection well injection rate. The rock porosity is 1.15~0.20, the rock permeability is 0.1~10 mD, the reservoir thickness is 0.8~6 m, and the injection well injection rate is 10000~40000 m³ / h. 3 / day; S13. The 51-dimensional parameter space is sampled using the Latin hypercube sampling algorithm to obtain a set of sample points, and 10,000 geological models are generated from the set of sample points.

[0008] Preferably, in S11: the 21 static parameters include the rock porosity of 7 target layers, the rock permeability of 7 target layers, and the effective reservoir thickness of 7 target layers; the 30 dynamic development parameters are the injection rate of the injection well in 30 time steps.

[0009] Preferably, in S2: S21, Set the injection regime: Apply a uniform injection rate of 60 months to 10,000 geological models, setting the injection rate to change every two months, that is, uniformly apply the injection rate of 30 time steps to the geological models from a time series perspective; S22, Data extraction: Extract the dynamic response data of each geological model using the Eclipse numerical simulator. The dynamic response data of each geological model includes the simulation results of 30 time steps, the gas injection rate and liquid / gas production rate of 21 subdivided layers / networks; S23, Set the sequence length: Set the time series length according to the time correlation of reservoir dynamic response. , through the front Historical data prediction at each time step S24. Dataset partitioning: Each geological model is treated as one sample, and the 10,000 samples are randomly divided into training and test sets in a 7:3 ratio.

[0010] Preferably, in S3: S31, Constructing an RNN module: Constructing a multi-layer RNN network, introducing a dropout mechanism between the multi-layer RNN networks, setting the dropout rate to 0.2 to prevent overfitting of the physically constrained RNN model; S32, Designing a fully connected output module: Constructing two fully connected networks after the RNN module as a fully connected output module; the first fully connected network maps the hidden output of the RNN module from 128 dimensions to 64 dimensions, introducing the ReLU activation function to enhance the nonlinear fitting ability of the physically constrained RNN model; S33, After the fully connected output module, designing a physical constraint mechanism using the ResNet residual connection idea to form a physically constrained RNN model.

[0011] Preferably, in the multi-layer RNN network of S31: the input dimension is set to 51 dimensions, that is, the input dimension is the dimension of the fused feature vector, and the hidden layer dimension is set to 128 dimensions; by using the batch_first=True mode, the input data is adapted to the standard form, which includes the number of samples, sequence length, and feature dimension, specifically expressed as follows: ,in The sample size is 1, with each geological model considered as one sample. The sequence length represents the data extracted from each layer at 30 time steps for each geological model, for a total of 21 layers. This indicates that the feature dimension is 51.

[0012] Preferably, the expression for the physical constraint mechanism is: In the formula, This represents the final prediction result of the physically constrained RNN model; The dynamic features of the input sequence at the previous time step are represented by the parameters related to the production and absorption profile, which include static and dynamic parameters. This represents the prediction increment of the fully connected output module.

[0013] Preferably, in S4; S41, calculate error metrics: input the test set into the physically constrained RNN model to obtain the prediction results, restore the real physical values ​​through inverse normalization, and calculate the core error metrics, including mean absolute error (MAE) and mean absolute percentage error (MAPE); the mean absolute error reflects the absolute deviation between the predicted value and the true value, and the mean absolute percentage error reflects the relative deviation, avoiding error evaluation distortion caused by differences in numerical magnitude; S42, perform error analysis: statistically analyze the mean MAE and MAPE of the test set to evaluate the overall performance of the physically constrained RNN model; analyze the error change trend of a single sample at different time steps to identify the key stages of error accumulation; S43, parameter optimization and adjustment, including hyperparameter adjustment and training set ratio adjustment.

[0014] Preferably, in S43: Hyperparameter tuning: Based on the results of error analysis, the hyperparameters of the physically constrained RNN model are adjusted, including the hidden layer dimension of the RNN module, the learning rate, and the sequence length.

[0015] Training set proportion adjustment: If or If so, the proportion of the training set increases by 150%; or If so, the proportion of the training set increases by 100%; or If the training set percentage increases by 50%, then the training set percentage will increase by 50%. or If so, no adjustment will be made.

[0016] Preferably, in S5; S51, Model Deployment: Deploy the optimized RNN model with physical constraints as an online prediction system, supporting real-time data input and result output; S52, Rolling Prediction Warm-up: Use the initial real monitoring data as input to the optimized RNN model with physical constraints to complete model initialization; S53, Continuous Prediction: From Starting from each time step, the input features at the current moment include the prediction results of the previous moment, combined with real-time acquired dynamic parameters (such as bottom hole flowing pressure) and injection phase discrimination parameters, and continuously output the prediction results of the production and absorption profile for each subsequent time step; S54, Output results: The online prediction system displays the production and absorption profile data and the trend of production and absorption profile data in real time, providing decision support for reservoir development adjustments (such as injection phase switching and bottom hole flowing pressure control).

[0017] Therefore, the carbon dioxide flooding production and absorption profile prediction method based on random sampling and RNN adopted in this invention has the following beneficial effects: (1) The dataset has significant advantages and supports the generalization ability of the model: 10,000 large-scale sample datasets are constructed through Eclipse numerical simulation, covering 60 months of dynamic evolution under multiple working conditions, and the data fidelity is high; the parameter system includes 21 static parameters and 30 dynamic development parameters, which comprehensively characterize the reservoir heterogeneity and development dynamic characteristics, providing sufficient and high-quality feature support for model learning, enabling the model to adapt to different geological working conditions and development systems, and has strong generalization ability; (2) The prediction accuracy is high and conforms to physical laws: On the one hand, through the deep integration of static and dynamic parameters, the model input features are ensured to be comprehensive and the key factors affecting the production and absorption profile are effectively captured; on the other hand, the physical constraint mechanism is introduced to make the prediction results based on the historical real state and to avoid deviating from the physical rules of reservoir displacement. (2) The prediction accuracy is greatly improved. The average MAE and MAPE of the full test set are both at a low level. (3) The computational efficiency is greatly improved to meet the real-time decision-making requirements: The prediction method based on the deep learning RNN model does not require solving complex multiphase flow equations. A single prediction only takes milliseconds. Compared with the hours or even days of traditional numerical simulation, the computational efficiency is improved by tens of thousands of times. The online prediction system supports rolling prediction and can output the production and absorption profile change trend in real time, which fully meets the real-time decision-making requirements in the dynamic development of reservoirs. (4) It is highly adaptable and has a wide range of applications: The dataset is simulated by random combination of injection rate change system of injection wells over 60 months, covering multidimensional geological conditions. After the model is trained on the dataset, it can adapt to carbon dioxide flooding development scenarios with different heterogeneous reservoirs and different injection rate switching frequencies. It can be widely used in the prediction of carbon dioxide flooding production and absorption profiles of various reservoirs. Its application range is far greater than that of traditional methods and existing machine learning methods.

[0018] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0019] Figure 1 is an overall flowchart of the carbon dioxide flooding production and absorption profile prediction method based on random sampling and RNN of the present invention; Figure 2 is a schematic diagram of the physically constrained RNN model of the embodiment of the present invention. Detailed Implementation

[0020] The following detailed description of embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0021] Please refer to Figure 1-2. The method for predicting carbon dioxide production and absorption profiles based on random sampling and RNN includes the following steps: S1, constructing a high-fidelity carbon dioxide production and absorption profile dataset.

[0022] S11. Determine Input Variables: Select static and dynamic development parameters characterizing reservoir heterogeneity. There are 21 static parameters, including rock porosity, rock permeability, and effective reservoir thickness for seven target formations. There are 30 dynamic development parameters, representing the injection rate of the injection well over 30 time steps. These 21 static and 30 dynamic parameters constitute a 51-dimensional parameter space. S12. Set Parameter Boundaries: Based on the geological characteristics of the target reservoir, set the value ranges for the variables, including rock porosity, rock permeability, reservoir thickness, and injection rate. Rock porosity is 1.15–0.20 mD, rock permeability is 0.1–10 mD, reservoir thickness is 0.8–6 m, and injection rate is 10,000–40,000 m³ / h. 3 / day; S13. The 51-dimensional parameter space is sampled using the Latin hypercube sampling algorithm to obtain a set of sample points, and 10,000 geological models are generated from the set of sample points.

[0023] S2. Perform preprocessing operations on the data.

[0024] S21. Setting the injection regime: Apply a uniform injection rate of 60 months to 10,000 geological models, setting the injection rate to change every two months, that is, uniformly apply the injection rate of 30 time steps to the geological models from the time series; S22. Data extraction: Extract the dynamic response data of each geological model through the Eclipse numerical simulator. The dynamic response data of each geological model includes the simulation results of 30 time steps, the gas injection rate and liquid / gas production rate of 21 subdivided layers / networks; The model used in this invention includes 7 major layers (corresponding to the above 21 (3 types) of static data), each major layer is further divided into 3 minor layers, for a total of 21 subdivided layers; In terms of rock porosity in the static data, 7 static parameters correspond to 7 layers, and each static data is assigned to 1 layer (3 subdivided layers share this one static data); S23. Setting the sequence length: Set the time series length according to the time correlation of reservoir dynamic response. , through the front Historical data prediction at each time step S24. Dataset partitioning: Each geological model is treated as one sample, and the 10,000 samples are randomly divided into training and test sets in a 7:3 ratio.

[0025] S3. Construct an RNN model with physical constraints.

[0026] S31. Building the RNN module: Construct a multi-layer RNN network, setting the input dimension to 51 dimensions (i.e., the input dimension is the dimension of the fused feature vector), and the hidden layer dimension to 128 dimensions; using batch_first=True mode, the input data is adapted to the standard form, which includes the number of samples, sequence length, and feature dimension, specifically represented as follows: ,in The sample size is 1, with each geological model considered as one sample. The sequence length represents the data extracted from each layer at 30 time steps for each geological model, for a total of 21 layers. The feature dimension is 51. A dropout mechanism is introduced between the multi-layer RNN network, with a dropout rate of 0.2, to prevent overfitting of the physically constrained RNN model. S32, Design a fully connected output module: Two fully connected networks are built after the RNN module as the fully connected output module. The first fully connected network maps the hidden output of the RNN module from 128 dimensions to 64 dimensions, and introduces the ReLU activation function to enhance the nonlinear fitting ability of the physically constrained RNN model. S33, After the fully connected output module, a physical constraint mechanism is designed using the ResNet residual connection concept to form a physically constrained RNN model. The expression for the physical constraint mechanism is: In the formula, This represents the final prediction result of the physically constrained RNN model; This represents the dynamic features of the input sequence at the previous time step, i.e., the parameters related to the production and absorption profile; This represents the prediction increment of the fully connected output module. Regarding the physical constraint mechanism: ① Model structure: A residual network structure is used to force the predicted value to satisfy the law of conservation of mass—predicted value = historical state + increment. The inherent meaning is to control the conservation of mass and ensure that the predicted value does not deviate from the physical constraints of geology; ② Data generation: Training data is generated by combining parameters through orthogonal experimental design to ensure that the data is based on geological statistical laws; ③ Error correction: The prediction deviation is dynamically corrected through the RNN error inversion module, returning to the range allowed by physical laws.

[0027] S4. Through error inversion, dynamically correct prediction bias and optimize the model.

[0028] S41. Calculate error metrics: Input the test set into the physically constrained RNN model to obtain the prediction results. Restore the true physical values ​​through inverse normalization and calculate the core error metrics, including Mean Absolute Error (MAE) and Mean Absolute Percentage Error (MAPE). MAE reflects the absolute deviation between the predicted and true values, while MAPE reflects the relative deviation, avoiding distortion in error assessment due to differences in magnitude. S42. Perform error analysis: Calculate the mean MAE and MAPE of the test set to evaluate the overall performance of the physically constrained RNN model. Analyze the error change trend of individual samples at different time steps to identify key stages of error accumulation. S43. Parameter optimization and adjustment, including hyperparameter adjustment and training set ratio adjustment. Hyperparameter adjustment: Based on the error analysis results, adjust the hyperparameters of the physically constrained RNN model, including the hidden layer dimension, learning rate, and sequence length. If the error analysis results show underfitting, increase the hidden layer dimension; if the error analysis results show overfitting, add dropout. Training set ratio adjustment: If... or If so, the proportion of the training set increases by 150%; or If so, the proportion of the training set increases by 100%; or If the training set percentage increases by 50%, then the training set percentage will increase by 50%. or If so, no adjustment is needed.

[0029] S5. Provide industrial-grade real-time decision support by deploying an online prediction system.

[0030] S51. Model Deployment: Deploy the optimized RNN model with physical constraints as an online prediction system, supporting real-time data input and result output; S52. Rolling Prediction Warm-up: Use the initial real monitoring data as input to the optimized RNN model with physical constraints to complete model initialization; S53. Continuous Prediction: From... Starting from each time step, the input features at the current moment include the prediction results of the previous moment, combined with real-time acquired dynamic parameters (such as bottom hole flowing pressure) and injection phase discrimination parameters, and continuously output the prediction results of the production and absorption profile for each subsequent time step; S54, Output results: The online prediction system displays the production and absorption profile data and the trend of production and absorption profile data in real time, providing decision support for reservoir development adjustments (such as injection phase switching and bottom hole flowing pressure control).

[0031] Using existing technologies such as Eclipse for reservoir simulation, each independent simulation process takes approximately 3 minutes. To obtain 3000 sets of model data through simulation, the entire process would take about 4-5 days. However, the online prediction system of this invention can complete the prediction of 3000 sets of model data in a very short time of 1-2 minutes. This comparison fully demonstrates the significant efficiency advantage of the online prediction system of this invention, greatly shortening the time required for data prediction and enabling the originally time-consuming simulation work to be completed quickly.

[0032] Therefore, this invention employs the aforementioned method for predicting carbon dioxide flooding production and absorption profiles based on random sampling and RNN. A large-scale dataset of 10,000 samples is constructed using Eclipse numerical simulation, covering a 60-month dynamic evolution process under multiple operating conditions, ensuring high data fidelity. The parameter system includes 21 static parameters and 30 dynamic development parameters, comprehensively characterizing reservoir heterogeneity and dynamic development characteristics, providing sufficient and high-quality feature support for model learning. This enables the model to adapt to different geological conditions and development regimes, exhibiting strong generalization ability. Deep fusion of static and dynamic parameters ensures comprehensive model input features, effectively capturing key factors influencing the production and absorption profile. The introduction of a physical constraint mechanism ensures that the prediction results evolve reasonably based on historical real conditions, avoiding deviations from the physical laws of reservoir displacement and significantly improving prediction accuracy. The system demonstrates low performance across the entire test set, with both the average MAE and MAPE remaining at low levels. The prediction method, based on a deep learning RNN model, eliminates the need to solve complex multiphase flow equations, requiring only milliseconds for a single prediction – a computational efficiency improvement of tens of thousands of times compared to the hours or even days required for traditional numerical simulations. The online prediction system supports rolling predictions and can output real-time trends in production and absorption profile changes, fully meeting the real-time decision-making needs in dynamic reservoir development. The dataset simulates 60 months of injection rate changes in injection wells using random combinations, covering multidimensional geological conditions. After training on this dataset, the model can adapt to different heterogeneous reservoirs and carbon dioxide flooding development scenarios with varying injection rate switching frequencies. It can be widely applied to predicting production and absorption profiles for carbon dioxide flooding in various reservoirs, with an applicability far exceeding traditional methods and existing machine learning methods.

[0033] 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 preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for predicting carbon dioxide flooding production and absorption profiles based on random sampling and RNN, characterized in that, Includes the following steps: S1. Construct a high-fidelity carbon dioxide production and absorption profile dataset; S2. Perform data preprocessing operations; S3. Construct a physically constrained RNN model; S4. Optimize the model by dynamically correcting prediction biases through error inversion; S5. Provide industrial-grade real-time decision support by deploying an online prediction system.

2. The method for predicting carbon dioxide flooding production and absorption profiles based on random sampling and RNN according to claim 1, characterized in that, In S1: S11, Determine input variables: Select static and dynamic development parameters characterizing reservoir heterogeneity; there are 21 static parameters and 30 dynamic development parameters, forming a 51-dimensional parameter space; S12, Set parameter boundaries: Based on the geological characteristics of the target reservoir, set the value range of the variables, including rock porosity, rock permeability, reservoir thickness, and injection well injection rate. Rock porosity is 1.15~0.20, rock permeability is 0.1~10 mD, reservoir thickness is 0.8~6 m, and injection well injection rate is 10000~40000 m³ / h. 3 / day; S13. The 51-dimensional parameter space is sampled using the Latin hypercube sampling algorithm to obtain a set of sample points, and 10,000 geological models are generated from the set of sample points.

3. The method for predicting carbon dioxide flooding production and absorption profiles based on random sampling and RNN according to claim 2, characterized in that, In S11: 21 static parameters include rock porosity, rock permeability, and effective reservoir thickness of 7 target layers; 30 dynamic development parameters are the injection rate of the injection well over 30 time steps.

4. The method for predicting carbon dioxide flooding production and absorption profiles based on random sampling and RNN according to claim 3, characterized in that, In S2: S21, Set Injection Schedule: Apply a uniform injection rate for 60 months to 10,000 geological models, setting the injection rate to change every two months; S22, Data Extraction: Extract dynamic response data for each geological model using the Eclipse numerical simulator. The dynamic response data for each geological model includes simulation results for 30 time steps, gas injection rates and liquid / gas production rates for 21 subdivided stratigraphic / network layers; S23, Set Sequence Length: Set the time series length based on the temporal correlation of reservoir dynamic response. , through the front Historical data prediction at each time step S24. Dataset partitioning: Each geological model is treated as one sample, and the 10,000 samples are randomly divided into training and test sets in a 7:3 ratio.

5. The method for predicting carbon dioxide flooding production and absorption profiles based on random sampling and RNN according to claim 4, characterized in that, In S3: S31, Build the RNN module: Construct a multi-layer RNN network and introduce a dropout mechanism between the multi-layer RNN networks, setting the dropout rate to 0.2; S32, Design the fully connected output module: Build two fully connected networks after the RNN module as the fully connected output module; the first fully connected network maps the hidden output of the RNN module from 128 dimensions to 64 dimensions, and introduces the ReLU activation function to enhance the nonlinear fitting ability of the physically constrained RNN model; S33, After the fully connected output module, design a physical constraint mechanism using the ResNet residual connection idea to form a physically constrained RNN model.

6. The method for predicting carbon dioxide flooding production and absorption profiles based on random sampling and RNN according to claim 5, characterized in that, In the S31 multi-layer RNN network: the input dimension is set to 51 dimensions, and the hidden layer dimension is set to 128 dimensions; the batch_first=True mode is used to adapt the input data to a standard form, which includes the number of samples, sequence length, and feature dimensions, specifically represented as follows: ,in The sample size is 1, with each geological model considered as one sample. The sequence length represents the data extracted from each layer at 30 time steps for each geological model, for a total of 21 layers. This indicates that the feature dimension is 51.

7. The method for predicting carbon dioxide flooding production and absorption profiles based on random sampling and RNN according to claim 6, characterized in that, The expression for the physical constraint mechanism is: In the formula, This represents the final prediction result of the physically constrained RNN model; It represents the dynamic features of the input sequence at the previous time step; This represents the prediction increment of the fully connected output module.

8. The method for predicting carbon dioxide flooding production and absorption profiles based on random sampling and RNN according to claim 7, characterized in that, In S4; S41, calculate error metrics: input the test set into the RNN model with physical constraints to obtain the prediction results, restore the true physical values ​​through the inverse normalization operation, and calculate the core error metrics, including mean absolute error (MAE) and mean absolute percentage error (MAPE). The mean absolute error (MAE) reflects the absolute deviation between the predicted and actual values, while the mean absolute percentage error (MAPE) reflects the relative deviation. S42. Perform error analysis: Statistically calculate the mean MAE and MAPE of the test set to evaluate the overall performance of the physically constrained RNN model; analyze the error trends of individual samples at different time steps to identify key stages of error accumulation. S43. Optimize and adjust parameters, including hyperparameter adjustment and training set ratio adjustment.

9. The method for predicting carbon dioxide flooding production and absorption profiles based on random sampling and RNN according to claim 8, characterized in that, In S43: Hyperparameter tuning: Based on the results of error analysis, adjust the hyperparameters of the physically constrained RNN model, including the hidden layer dimension, learning rate, and sequence length of the RNN module; Training set ratio adjustment: If or Then the proportion of the training set increases by 150%; like or Then the proportion of the training set increases by 100%; like or Then the proportion of the training set increases by 50%; like or If so, no adjustment will be made.

10. The method for predicting carbon dioxide flooding production and absorption profiles based on random sampling and RNN according to claim 9, characterized in that, In S5; S51, Model Deployment: Deploy the optimized RNN model with physical constraints as an online prediction system, supporting real-time data input and result output; S52, Rolling Prediction Warm-up: Use the initial real monitoring data as input to the optimized RNN model with physical constraints to complete model initialization; S53, Continuous Prediction: From... Starting from each time step, the input features at the current moment include the prediction results of the previous moment, combined with the dynamic parameters and injection phase discrimination parameters collected in real time, and continuously output the prediction results of the production and absorption profile at each time step; S54, Output results: The online prediction system displays the production and absorption profile data and the changing trend of the production and absorption profile data in real time, providing decision support for reservoir development adjustment.

Images