Recurrent Neural Networks for Permanent Downhole Gauge Data Analysis

Abstract

AbstractThe rich information contained in permanent downhole gauge (PDG) data has drawn attention from researchers. Previous work has demonstrated feature-based machine learning to be promising for PDG data analysis. The recent rise of deep learning was powered by techniques including recurrent neural networks (RNNs), which have been shown useful for processing sequential information. In this work, we explored how RNN can be utilized to analyze PDG data for better reservoir characterization and modeling, by examining two specific RNN structures: nonlinear autoregressive exogenous model (NARX) and standard RNN.RNN is a special type of artificial neural network designed for sequential data processing. Unlike a nonrecurrent neural network, the hidden layers in RNN take memories of previous computations (e.g. hidden layers or outputs computed in previous time) into account. In that way, the information contained in the measurements prior to a certain time can be used to model the response at that time, i.e. the convolutional effects are modeled by the recurrent structure of the RNN. Another favorable property of RNN is that it requires no assumptions on physics in advance. Both the inputs and outputs come directly from the raw measurements, and no handcrafted features need to be extracted from the forward physics model. Compared with feature-based machine learning, RNN is more powerful for the modeling where the forward model has not been determined.In this work, RNN was first tested on a series of synthetic and real flow rate-pressure datasets. The patterns learned by RNN from those data helped correctly identify the reservoir model and forecast the reservoir performance. RNN was also applied on temperature transient data, to demonstrate its advantage over feature-based machine learning with no assumptions on the physics model. The study also showed that RNN has the noise tolerance and computational efficiency that make it a promising candidate to analyze PDG data in practice.