APPLYING DEEP LEARNING TO PETROLEUM WELL DATA 1
Applying Deep Learning to Petroleum Well Data
Janette Garcia, Akash Levy, Albert Tung, Ruomeng (Michelle) Yang, and Verena Kaynig-Fittkau
Abstract—In this work, we explore deep learning methods
to perform time series prediction on petroleum well output.
We successfully trained restricted Boltzmann machines (RBMs),
fully-connected networks (FCNs), convolutional neural networks
(CNNs), and recurrent neural networks (RNNs) on petroleum
well data to accomplish this task. A comparison of our results
indicate that neural networks are a useful tool for understanding
the behavior of petroleum wells and may in many cases outper-
form previous methods for predicting future oil well data, both
in time efficiency and accuracy.
I. PROBLEM STATEMENT
O
UR goal for this work is to improve upon the existing
techniques used by petroleum engineers to analyze and
appraise oil wells. Appraising an oil company is a lengthy
investigation process; the production profiles of oil wells can
be complex, driven by reservoir physics, marked by a variety
of operational events and obfuscated by data noise. Petroleum
engineers can usually decipher the production profiles of oil
wells, understand their underlying behavior, forecast their
expected production, and identify opportunities for perfor-
mance improvements. However, the investigation process is
time-consuming. This opens a vast portfolio of unidentified
opportunities to optimize this analytic process. Thus Quantum
Reservoir Impact (QRI), a petroleum reservoir management
company, has proposed to explore whether deep learning
algorithms, known to be effective in pattern recognition and
object classification, can be used to understand and predict
the behavior of oil wells. Our objective is to determine the
viability of neural networks in predicting future outcomes,
specifically with time series data representing oil production
in petroleum wells.
II. INTRODUCTION
First introduced in the early 1980s, neural networks are
a programming paradigm within deep learning that enable
computers to learn from observational data. They leverage the
physiology and architecture of the human brain [1]. Neural
networks are generally presented as systems of interconnected
neurons that send messages to each other and are used to
approximate functions that depend on large amounts of input
[2]. Their connections have numeric weights that are tuned
based on experience, allowing them to adapt to inputs and to
learn [3].
Deep learning involves relaying information from the input
layer through multiple hidden layers, where representations are
stored in the form of matrices and processed mathematically
at each layer [4]. Various neural networks are classified by
A special thanks to: Quantum Reservoir Impact (QRI), Institute for Applied
Computational Science (IACS) at Harvard University, Harvard John A.
Paulson School of Engineering and Applied Sciences.
differences in the mathematical operations and layout archi-
tecture between the hidden layers. Each network is designed
and optimized for jobs such as image classification, pattern
recognition, and time series prediction.
A variety of neural networks present rich architectures to
model sequential time series data of oil wells. We were par-
ticularly interested in the performance of restricted Boltzmann
machines (RBMs), fully-connected networks (FCNs), convolu-
tional neural networks (CNNs), and recurrent neural networks
(RNNs). In this paper, we implement the aforementioned
four deep learning models to explore the potential for neural
networks to forecast this type of time series data. The task
involved training the neural networks to understand oil well
data and tweaking the network parameters to yield optimal
results.
In the following sections, we first describe prior work
relevant to the task at hand. We then discuss the neural
networks we utilized to accomplish the task. After this, we
explain the dataset and preprocessing methods involved in
training and testing. We finally report experimental results and
include discussion of our work.
III. RELATED WORK
Deep learning models have been shown to perform well
at many different tasks. Image classification is a well-known
example; deep learning has been used to classify handwritten
digits [5] and recognize traffic signs [6] with a high level of
accuracy. Recent research on time-series prediction has been
successfully achieved on electroencephalography (EEG) data
that measures cortical brain activity with temporal resolution
[7] and ultra-short-term wind prediction [8]. These experi-
ments rely on several kinds of neural networks, including deep
recurrent networks, fully-connected networks, and convolu-
tional neural networks. Work has also been conducted using
recurrent neural networks on electricity demand forecasting
[9]. However, problems with neural network training and
time efficiency have been cited [10]. We anticipated that by
implementing several neural network models and optimizing
hyper-parameters on a graphics processing unit (GPU), we
would be able to leverage the highly adaptive and expressive
properties of neural networks and uncover useful knowledge
about the models and petroleum wells.
IV. METHODS
We implemented four machine learning architectures: re-
stricted Boltzmann machines (RBMs), fully-connected net-
works (FCNs), convolutional neural networks (CNNs), and
recurrent neural networks (RNNs). Below, we discuss the
history of each model as well as the technicalities involved in
applying them to our task. We then explain the data provided