Neural Network Identifies Misallocated Reservoir Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Reservoir simulations face inaccuracies due to misallocated historical production data, which complicates the prediction of reservoir performance in complex geological settings like petroleum reservoirs, where analytical solutions are intractable and standard CFD software is insufficient.
Innovation Solution
A method and system using a neural network algorithm trained with high-confidence reservoir data to identify and correct misallocated historical production data by comparing it with high-confidence reservoir sample data, replacing outliers, and retraining the algorithm to improve predictive accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If standard CFD software is used for reservoir simulation, then computational fluid dynamics can be applied, but it is insufficient for complex geological settings with asymmetric, irregular geometries and flow variable interdependencies
Solution Approach 1:
The patent creates a digital twin (virtual copy) of the reservoir that replicates its complex geological features, flow dynamics, and production behavior. This digital twin is trained using historical production data and can be queried to predict future performance, thereby capturing the adaptability needed for complex geometries while maintaining reliability through data-driven validation.
Solution Approach 2:
The patent replaces traditional mechanical CFD solving approaches with a machine learning-based system. Instead of repeatedly solving complex partial differential equations, a neural network model is trained once on historical data and then used for rapid predictions, substituting the mechanical computational process with an intelligent system that better handles geological complexity.
2Reliability
If historical production data is used to train reservoir simulation models, then predictive ability can be improved, but misallocated historical production data adversely affects the predictive ability
Solution Approach 1:
The patent implements a feedback mechanism where the digital twin's predictions are continuously compared against actual production data. When discrepancies are detected, the system identifies potentially misallocated historical data and corrects it, then retrains the model. This closed-loop feedback ensures that only accurate, well-allocated data is used for training, thereby improving predictive ability while filtering out measurement errors.
Solution Approach 2:
The patent performs preliminary validation and cleaning of historical production data before using it to train the reservoir simulation model. By pre-processing the data to identify and correct misallocations, the system ensures that the training data is of high quality, which directly improves the predictive ability of the resulting model.
3Productivity
If a neural network model is trained with reservoir historical production data, then predictions of reservoir performance can be made, but misallocated data in the training set reduces prediction accuracy
Solution Approach 1:
The system uses feedback loops to continuously validate the neural network's predictions against actual reservoir performance. When prediction errors are detected, the system traces back to identify misallocated historical data in the training set, corrects these allocations, and retrains the model. This iterative feedback process ensures that the neural network maintains high prediction capability while using only high-quality, accurately allocated data.
Data Source
AI summary
A method for training a predictive reservoir simulation in which high-confidence reservoir sample data is used to identify misallocated historical production data used in the simulation. A neural network algorithm is trained with high-confidence reservoir historical production data. High-confidence reservoir sample data is obtained by at least one sensor at a reservoir location over a time interval, after which the reservoir historical production data is parametrically varied over the time interval to determine a time-indexed discrepancy between the reservoir historical production data and the high-confidence reservoir sample data over the time interval. The time-indexed discrepancy and a defined threshold discrepancy are then used as inputs to a machine learning process to further train the neural network algorithm to identify reservoir historical production data whose discrepancy exceeds the threshold discrepancy and thereby constitutes misallocated historical production data. The misallocated data is later back allocated to respective wells by back propagation algorithm.