Integrated Data-Driven and Physics-Based Techniques for Improved/Enhanced Huff and Puff Gas Injection in Multiporosity Shale Oil Reservoirs

Abstract

I anticipate that the revolution of data analytics in the oil and gas industry will have a profound global impact. By employing a high volume of collected data and computational power, my goal is to contribute to reservoir engineering practices in a more efficient and optimal way. This thesis presents innovative research in the field of shale reservoir management, with a primary focus on neural network applications encompassing reservoir characterization, production forecasting, surrogate modeling, and improve/enhanced oil recovery optimization while honoring physics-based reservoir data. To achieve my goal: 1. I present a novel integration of machine learning (ML) and petrophysical analysis through the development of a hybrid data-driven technique for reservoir characterization. This approach utilizes a hybrid ML model to calculate brittleness indices in shale reservoirs based on mineralogical data and well logs. The hybrid model employs a modified Pickett plot to identify oil-saturated brittle sweet spots amenable to successful hydraulic fracturing. 2. I develop a novel Sequence-to-Sequence (Seq2Seq) Long Short-Term Memory (LSTM) model for oil production forecasting. I compare results with traditional reservoir engineering techniques. The model, based on natural language processing techniques, accurately predicts future oil production rates by analyzing historical data sequences at daily intervals. 3. I optimize huff and puff (H-n-P) gas injection in shale reservoirs, using an innovative approach that integrates sequence-based proxy reservoir simulation with customized deep reinforcement learning (DRL). This strategy significantly reduces simulation time while facilitating decision-making during H-n-P gas injection projects. 4. I demonstrate that DRL can be used for other applications such as optimizing numerical tuning while simultaneously performing numerical simulation. Similarly, I use DRL for water flood optimization. I use data of the Eagle Ford Shale in Texas to demonstrate points 1, 2 and 3. I use parallel computing to demonstrate the applications of point 4. I conclude that methods employed in this thesis, which integrate physics-based and data-driven techniques, provide effective and innovative alternatives for reservoir management. These approaches leverage the power of the universal approximation theorem, demonstrating that neural networks can approximate any continuous function, enhancing thus the capabilities of reservoir management.