Speaker
Description
This study aims to improve the accuracy and reliability of shale gas production predictions by integrating a Conformer model with 3D reservoir properties, considering well parameters and the geological characteristics surrounding the horizontal wells. A comprehensive analysis of 672 horizontal wells in the Duvernay Formation was conducted to validate the enhanced accuracy and robustness of the proposed approach, introducing new alternatives for predictive modelling in unconventional resource extraction. A novel approach for accurate production prediction was introduced, combining a 3D geo-parameterization technique with a hybrid 3D Conformer module. The 3D Geo-Parameterization method was developed to tokenize the formation properties surrounding the horizontal wells. These geological and well operation features were integrated with a multi-head self-attention-based Conformer model and fused into a unified representation of the horizontal wells' productivity. The integrated multimodality features were subsequently input into three distinct models—LightGBM, XGBoost, and CatBoost—each associated with a specific sampler. Outputs from the three models were finally stacked to predict the well’s production. A thorough field study involving 672 horizontal wells in the Duvernay Formation was carried out, demonstrating that the newly developed method achieves an impressive coefficient of determination (R²) of 0.86 for predicting shale gas production over 12 months. This signifies a remarkable 15.1% average improvement in R² and 30.5% mean absolute percentage error (MPAE) decline compared to traditional methods, such as LightGBM and Artificial Neural Network, which relied solely on tabular data and utilized cumulative production figures. This advancement highlights the superiority of the novel 3D geo-parameterization technique and the integration of a 3D Conformer model, demonstrating their effectiveness in unravelling complex geological factors. Incorporating multi-modal inputs and utilizing a hybrid fusion architecture significantly boosts the model's predictive accuracy by reinforcing the relationships between diverse features and shale gas production. Simultaneously, the integration of a self-attention mechanism within the 3D-Conformer architecture plays a crucial role in emphasizing and utilizing the distribution of properties near the wellbore region, thereby enhancing the model's performance. This innovative approach establishes a new benchmark for predictive modelling in unconventional resource extraction. It emphasizes the importance of utilizing unstructured geo-property distribution to enhance the accuracy of production forecasts. This study highlights the transformative potential of combining advanced machine learning architectures with 3D reservoir models for shale gas production prediction. The results advocate for further investigation into hybrid models specifically designed for unconventional resources, emphasizing the benefits of multi-modality parameterization for enhanced performance. Future work will focus on optimizing the model with real-time production data and assessing its adaptability to various geological formations, offering a novel perspective and advancing the application of machine learning in petroleum engineering.
| Country | Canada |
|---|---|
| Student Awards | I would like to submit this presentation into both awards |
| Acceptance of the Terms & Conditions | Click here to agree |
