A Multiscale Recurrent Neural Network Model for Long-Term Prediction of Geothermal Energy Production

Abstract

Management and optimization of energy recovery from geothermal reservoirs rely on accurate prediction of energy production performance for alternative development scenarios. While physics-based reservoir simulation models are traditionally used as a comprehensive approach to predict the response of geothermal reservoirs to different production strategies, data-driven models can serve as efficient fit-for-purpose tools for rapid evaluation and screening of alternative production and development plans as well as for facilitating daily operation and surveillance decisions. We evaluate the use of recurrent neural networks (RNN), as machine learning architectures for representation and prediction of sequential/dynamic data, for long-term prediction of geothermal energy production. As a universal approximator that can capture complex and nonlinear trends in data, RNN has enjoyed great success in many applications. Since RNN primarily exploits statistical relations in training data to generate predictions, it can be challenged in applications where extrapolation beyond the training data range is needed. We explore new RNN architectures and training approaches to improve its generalization power for long-term prediction of geothermal reservoir performance. Since historical monitoring and performance observations in geothermal fields contain short-term and long-term patterns, we develop multiscale RNN architectures and the corresponding training implementations to learn both short-term and long-term trends in the data and use them for future predictions. We present the developed architecture and training process and show its application to simulated datasets from a field-scale model.