Inferring Interwell Connectivity in Fractured Geothermal Reservoirs Using Neural Networks

Abstract

The goal of this study was to use machine learning methods to characterize the fractured nature of geothermal reservoirs, specifically neural networks. A series of simulations were performed on discrete fracture networks with different fracture characters. For each of the fracture networks, a neural network was trained using injection rates from two injectors as input and produced tracer concentration at one producer as output. A sensitivity analysis was applied on the trained models to quantify the significance that each injector has on the producer which was then translated to a connectivity index describing the injector-producer relationship. The connectivity indices from the neural network were compared to connectivity indices obtained using the ACE algorithm as well as simulated peak arrival time of the tracer. In many cases the neural network connectivity was in agreement with the connectivity from peak arrival times, but the results were not conclusive in determining if neural networks could in fact quantify the strength of connection between wells.