Harvesting Earth's heat: A deep learning Odyssey for reservoir characterization and sustainable geothermal energy management

Abstract

Characterizing and predicting reservoir behavior pose significant challenges in geothermal reservoir engineering. For sustainable geothermal field management, the potential occurrence of thermal breakthroughs in producers during the injection of cold water necessitates a profound understanding of how production is influenced by the injection philosophy. The fractured nature of geothermal reservoirs adds complexity and nonlinearity to the relationship between production and injection wells. In this study, we explore alternative models to simulate reservoir behavior as substitutes for full reservoir simulations. Utilizing deep learning algorithms, we investigate two different architectures, namely the standard Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU). These models map injection flow rates at the injectors to tracer concentration data at the producers through layers, nodes, and activation functions. Training these models on a synthetic geothermal reservoir, Notably, the GRU architecture consistently demonstrates superior predictive capabilities across various scenarios for all three producers. Specifically, in configurations involving a target feed, the GRU model yields higher R2 values compared to its LSTM counterpart, indicating its effectiveness in capturing inter-well relationships. Furthermore, when considering all producers collectively, the GRU model exhibits a trend of lower test errors and higher R2 values, reaffirming its proficiency in modeling complex reservoir dynamics. These findings underscore the significance of GRU as a preferred choice for accurate reservoir behavior prediction. In addition to the aforementioned insights, our study contributes novel approaches to reservoir behavior prediction in the field of geothermal reservoir engineering. By leveraging deep learning algorithms, specifically the GRU and LSTM architectures, we introduce innovative methodologies for simulating reservoir behavior as alternatives to full reservoir simulations. Through comprehensive analysis, we demonstrate the superior predictive capabilities of the GRU model, particularly in capturing inter-well relationships and modeling complex reservoir dynamics.