Recurrent Neural Networks for Predicting the Dynamic Response of Geothermal Reservoirs from Monitoring Data

Abstract:

Management and optimization of energy production from geothermal reservoirs rely on accurate prediction of energy production performance for alternative development scenarios. Although physics-based modeling provides a comprehensive prediction approach, construction of a reservoir model is not a trivial task and involves integration of multiple sources of data as well as geologic and flow modeling expertise. Moreover, in many cases complex coupled multi-physics processes are needed to properly represent hydraulic, thermal, geomechanical, and geochemical effects. Given the uncertainty in the description of reservoir models as well as physical processes and the related properties (model input parameters), the resulting predictions are subject to a significant level of uncertainty. Despite these limitations, physics-based simulation models are widely accepted as a standard prediction approach for long-term planning and management of geothermal reservoirs and other subsurface flow and transport systems.

Data-driven models rely on statistical patterns and dependencies in the collected data from various sources to develop a predictive model. A powerful class of data-driven models that has enjoyed great success in many fields is neural network, especially deep learning models. For dynamical systems where the response data is represented as correlated time-series (sequential data), recurrent neural networks (RNN) are more effective for capturing the temporal trends in the data. RNN architectures represent a directed graph with a temporal sequence that can be used to model dynamic data. An important advantage of RNN is its internal state (memory), which is used to process data sequences of variable lengths. In this work, we develop coupled sequence-to-sequence RNN models for predicting the dynamic response of geothermal reservoirs using past monitoring measurements. The developed RNN architecture consists of an encoder that captures the dynamics within the time-series input and a decoder RNN that exhibits the dynamic behavior in the prediction. We present the workflow, including the RNN architecture and its training process, and evaluate its performance by applying to several examples, including real field data from geothermal reservoirs.