Artificial Neural Networks for Geothermal Reservoirs: Implications for Oil and Gas Reservoirs

Abstract

Subsurface numerical models take a significant time to build and run. For this reason, the energy industry has been looking towards proxy models that could reduce model computational time. With the advancement of artificial neural network algorithms, building proxy models has become more efficient, and has enabled quick forecasting and quick reservoir management decision-making. In this study, we used a geothermal reservoir to evaluate the suitability of two deep learning algorithms, feed forward neural network and convolutional neural network, for proxy modeling. We used metrics such as the mean square error, losses, number of parameters for the model, and time to run, to compare the two deep learning algorithms. From our study, we determined that the convolutional neural network resulted in less error than the feed forward network and used less hyperparameters. However, the feed forward network was significantly faster than the convolutional neural network. The process of building the proxy model shows how a similar approach can be followed for oil and gas reservoir modeling and demonstrates the feasibility of neural networks in subsurface reservoir modeling and forecasting.