Abstract
This research addresses the challenges in using machine learning (ML) to assess and optimize production in unconventional wells, where computational costs (dependence on the accuracy of physical models) and complexity of wellbore design create significant challenges for decision-making and field development. A novel Hybrid Data-Physics (HDP) architecture is proposed that integrates data-driven models with physics equations embedded in a deep neural network (DNN). This approach optimizes both network and physical parameters simultaneously to predict short and long-term production rates, thereby generating more realistic operational scenarios.
Using a comprehensive dataset from nearly 1300 Duvernay wells within the Western Canadian Sedimentary Basin (WCSB), the proposed HDP model significantly improves production performance estimations by refining physical parameters through iterative neural network processes. The model performs exceptionally well even with limited data, particularly in unconventional wells where geological complexities pose challenges for traditional simulation methods. By incorporating simpler equations such as decline curves, the HDP model bypasses the need for complex physics, capturing hidden complexities and operational trends.
This study highlights the HDP model's transformative potential in production optimization, merging data analytics with physics-based modeling to enhance operational insights and decision-making. This pioneering approach reduces uncertainty and adapts to dynamic conditions, offering a robust, efficient tool for unconventional reservoir management. Additionally, the HDP model's ability to integrate various data types and adapt to evolving conditions underscores its versatility and practical applicability in real-world scenarios. Through this integration, the model not only enhances predictive accuracy but also provides a scalable solution that can be tailored to diverse reservoir conditions and operational needs.
