Machine Learning Surrogate Models for Well Interference Forecasting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Hydraulic fracturing treatments induce complex network geometries, making it challenging to evaluate well productivity using physics-based modeling approaches, especially with decreased perforation cluster spacing and increased treatment size, leading to uncertain well production forecasting due to well interference and near-wellbore fracture competition.
Innovation Solution
A data-driven approach utilizing machine learning models to forecast well interference and predict infill or child well production by collecting and processing input data, developing and tuning machine learning models, and performing model ensemblement to provide real-time forecasting and optimization of well completion and reservoir features.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If physics-based modeling approaches are used to evaluate well productivity, then detailed fracture network geometries can be analyzed, but computational complexity and time requirements increase significantly
Solution Approach 1:
The patent creates simplified surrogate models that replicate the behavior of complex physics-based models. These surrogate models are trained on datasets generated from detailed fracture network simulations, allowing them to predict well productivity with comparable accuracy but at a fraction of the computational cost. The surrogate models serve as efficient copies that capture the essential relationships without requiring full physics-based calculations.
Solution Approach 2:
The patent transforms the complex physics-based modeling problem into a machine learning prediction problem by changing the parameters from physical quantities requiring differential equation solutions to statistical patterns learned from training data. This parameter transformation allows the system to trade detailed physical modeling for data-driven predictions that are computationally efficient while maintaining predictive accuracy.
2Productivity
If decreased perforation cluster spacing and increased treatment size are used, then well production potential increases, but well interference and fracture competition effects become more complex and harder to model
Solution Approach 1:
The patent uses surrogate models that are trained to replicate the complex interactions in dense fracture networks. These models learn the patterns of well interference and fracture competition from training data generated with detailed fracture geometries, enabling them to predict production in complex scenarios without requiring explicit modeling of every fracture interaction.
Solution Approach 2:
The patent segments the complex fracture network problem into manageable training samples and features. By dividing the training data into subsets representing different fracture configurations and well spacing scenarios, the machine learning model can learn incremental patterns and generalize to complex cases that would be difficult to model as a single unified system.
3Stability of the object's composition
If conventional modeling approaches are used for well production forecasting, then traditional assumptions can be maintained, but they fail to address measurement errors and data variation in field production data
Solution Approach 1:
The patent changes the modeling approach from deterministic physics-based equations to probabilistic machine learning models that can accommodate measurement errors and data variations. The machine learning models learn from actual field data including its imperfections, making them more robust to the variability and noise present in real-world production measurements.
Solution Approach 2:
The machine learning models automatically adapt to the characteristics of the training data, including measurement errors and variations. Rather than requiring manual adjustment of assumptions to account for data quality issues, the models self-adjust during training to learn the actual relationships present in the field data, making them more reliable for forecasting.
Data Source
AI summary
Disclosed are systems and methods for obtaining input data comprising properties associated with at least one parent well and a child well associated with the at least one parent well, dividing the input data into a training data subset, a validation data subset, and a test data subset, selecting at least one machine learning model using the training data subset, the validation data subset, and the test data subset based on k-fold cross validation, tuning hyper parameters for each of the at least one machine learning model, and generating a learning output using the at least one machine learning model and the hyper parameters for each of the at least one machine learning model, the learning output indicating a test root-mean-square error (RMSE) and a training RMSE.