| Title | Physics-Guided Deep Learning for Prediction of Geothermal Reservoir Performance |
|---|---|
| Authors | Zhen QIN, Anyue JIANG, Dave FAULDER, Trenton T. CLADOUHOS, and Behnam JAFARPOUR |
| Year | 2022 |
| Conference | Stanford Geothermal Workshop |
| Keywords | machine learning, physics-informed, neural networks, geothermal reservoir |
| Abstract | Data-driven predictive models have recently been developed and applied to the prediction of energy production from geothermal reservoirs. Traditional physics-based simulation models that are often used for predicting the performance of geothermal reservoirs require detailed information of reservoir conditions and properties, in-depth technical knowledge, and extensive computational efforts. Data-driven models, on the other hand, extract complex statistical patterns from representative training data and use them for prediction. Deep learning-based architectures offer powerful and fast prediction models that are amenable to flexible implementation and convenient deployment and application. Despite their efficiency, data-driven models suffer from three main limitations: 1) extensive data requirements, 2) lack of physical plausibility and interpretability, and 3) poor generalizability beyond training data. These limitations undermine the use of data-driven models for applications in complex science and engineering problems. We develop a novel physics-informed machine learning approach that integrates physics-based representations into recurrent neural network (RNN) models. A typical approach for enabling the incorporation of physics is to regularize the learning of trainable parameters by adding a physics-guided loss function as a regularization term. In this framework, the architecture of the deep learning model is designed based on the governing equations as well as expert knowledge of the physical system. We present a variant of the RNN as a physics-guided deep-learning model for predicting the dynamical responses of geothermal reservoirs. Inspired by the lumped-parameter modelling, the architecture of the predictive model is designed based on simplified governing equations (i.e., conservation of energy). Different components of heat-transfer terms are incorporated into the architecture of physics-informed RNN to approximately represent the behaviour of the underlying dynamics. We present the physics-informed RNN approach in detail and demonstrate its prediction performance by applying it to data from a field-scale geothermal reservoir model. |