Hybrid SINDy Model for Unconventional Reservoir Production Forecasting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for forecasting well production in unconventional reservoirs, such as shale gas and tight oil formations, are inadequate due to their inability to effectively capture complex physics and operational dynamics, leading to unreliable production forecasts and a lack of robust, scalable solutions for optimal reservoir management.
Innovation Solution
A hybrid data-driven and physics-informed model using sparse nonlinear regression (SNR) methods, specifically sparse identification of nonlinear dynamics (SINDy), is employed to identify rate-pressure relationships and forecast well production by analyzing bottomhole pressure and flowrate data, allowing for automated and practical forecasting without relying on mechanistic fracture or simulation model assumptions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional decline curve analysis (DCA) and rate transient analysis (RTA) methods are used for production forecasting, then the forecasting process is simple and manual, but they fail to capture complex physics and nonlinear processes in unconventional reservoirs, resulting in unreliable forecasts
Solution Approach 1:
The patent merges data-driven machine learning methods with physics-based reservoir models to create a hybrid forecasting system. The neural network learns from historical production data while the physics model incorporates fundamental reservoir equations, combining the pattern recognition strength of ML with the physical accuracy of traditional models to achieve reliable forecasts for complex unconventional reservoirs
Solution Approach 2:
The forecasting system uses a composite approach by integrating multiple modeling paradigms (data-driven and physics-based) into a unified framework. This composite model structure allows the system to leverage the advantages of both approaches: the adaptability of machine learning to complex nonlinear behaviors and the physical consistency of traditional reservoir models
2Productivity
If numerical simulation models incorporating complex physics are used, then the model captures detailed reservoir physics, but it is not conducive for continuous model updates and production optimization for thousands of wells every day
Solution Approach 1:
The patent segments the forecasting problem by using a two-tier approach: a lightweight data-driven model for rapid predictions and a more complex physics-based model for calibration and validation. This segmentation allows daily production optimization across thousands of wells using the fast model while periodically updating parameters using the more accurate physics model, achieving both speed and reliability
Solution Approach 2:
The patent replaces the computationally intensive numerical simulation mechanics with a machine learning-based predictive system. The neural network is trained on physics-based simulation data and then used to rapidly predict production without running full numerical simulations, substituting the mechanical computation process with a statistical inference process that is much faster for operational use
3Ease of operation
If traditional RTA methods with explicit analytical formulations are used, then the diagnostic process is systematic, but non-linearities require pseudo-pressure and pseudo-time transformations and manual three-stage processes, reducing efficiency
Solution Approach 1:
The patent implements self-service automation where the machine learning model automatically performs diagnostic analysis, parameter estimation, and production forecasting without requiring manual intervention through traditional three-stage RTA processes. The system self-calibrates using available data and continuously updates predictions, eliminating the need for manual diagnostics, model calibration, and forecasting steps that consume significant time
Data Source
AI summary
A method of forecasting production of a well penetrating a reservoir in a subterranean formation, including: receiving a plurality of bottomhole pressures for the well; receiving a plurality of flowrates for the well; determining rate normalized pressure (RNP) data for the well over a period of time based on the plurality of bottomhole pressure and the plurality of flowrates; performing a sparse identification of nonlinear dynamics (SINDy) analysis on the RNP data to identify a relationship between flowrate and bottomhole pressure for the well, wherein the SINDy analysis is based on a plurality of physics features; providing a forecast of future production for the well based on the identified relationship between flowrate and bottomhole pressure for the well; and producing fluids from the reservoir based, at least in part, on the forecast of future production.