Speaker
Description
(1) Purpose and scope of the study;
With the development of artificial intelligence technologies, AI-based oil and gas productivity prediction has become a hotspot. However, traditional Estimated Ultimate Recovery (EUR) intelligent prediction is mostly data-driven without physical constraints, leading to limited model generalization ability and large discrepancies between predicted EUR and actual EUR. Therefore, conducting research on deep shale gas EUR prediction based on physical constraints-data-driven dual constraints has become an indispensable step for efficient and intelligent development of shale gas currently.
(2) Methods, Steps and Processes;
Here, we construct a physics-constrained-data-driven machine learning framework for predicting well EUR in Duvernay shale reservoir near Fox Creek, Alberta, Canada. First, data of core analysis, geochemic well logging, treatments and production are collected. Reservoir porosity, saturation, and permeability are derived from core-logging integrations, geochemical parameters are sourced from geochemical experiments, and rock mechanical parameters are obtained through geomechanical experiments. Then, these parameters are used as input features to train machine learning model with actual well EUR. Meanwhile, the modified Arps equation is embedded into LSTM in differentiable form to capture long-term trends of production data, transforming it into physics-constrained models.
(3) Results, Insights and Conclusions ;
A study was conducted based on 30 Duvernay shale gas wells with complete time series of production data. The training model employs porosity, permeability, gas saturation, TOC, brittleness index, shale thickness, and buried depth as input parameters, while select well production and EUR as output parameters with a modified Arps equation as constraints. A shale-gas-EUR prediction model based on physical constraints-data-driven dual constraints was finally established. The results indicate that the shale gas well decreasing curve can be accurately characterized by a non-homogeneous fractional order differential equation. The prediction efficiency of 12-month production and EUR using the proposed model for the examined well is improved by an average of 22% and 34%, respectively, compared with that of the purely data-driven model. Furthermore, the model was used to predict 12-month shale gas production from a fractured horizontal well that was producing in January 2024 and was not included in the training model, with a prediction accuracy of more than 88% compared to the actual production. Therefore, the accuracy of this physically constrained EUR prediction model is fully demonstrated. The workflow of this research can be used to guide the oil and gas production prediction in other sedimentary basins.
| Country | China |
|---|---|
| Acceptance of the Terms & Conditions | Click here to agree |
