Page 175 - Intelligent Digital Oil And Gas Fields

134 Intelligent Digital Oil and Gas Fields

The DL approach—also known as the deep structured learning or hierarchi-
cal learning—is new and at the forefront of ML research. The main idea is to
move ML closer to its AI roots. An example of a DL technique is a deep
neural network (DNN), which combines a multilayer network with mul-
tiple hidden layers organized into a graph or network (Fig. 4.15), compared
with the “classic” NN with a single feed-forward “hidden” layer (see Fig.
4.8A). The main advantage of a DNN over a classic, single-layer NN is
the ability to abstract high-level data from extremely complex data sets.
For more information on DL and DNN see Goodfellow et al. (2016).
It is encouraging to see that the concepts of DL and DNN have recently
started to find their way into E&P research and development, production
optimization, and predictive modeling. Crnkovic-Friis and Erlandson
(2015) train DNNs to learn the relationships between the geological param-
eters of nonconventional reservoirs (e.g., thickness, porosity, water satura-
tion, vitrinite reflectance, total organic carbon, brittleness, etc.) and average
estimated ultimate recovery (EUR) of an asset.
The DNN model was trained, validated, and tested on a region in the
Eagle Ford shale and has included both oil and dry gas wells. The DNN
model significantly outperforms both volumetric estimates and type-curve
region averages in terms of EUR prediction. However, the most important
advantage over traditional decline/type-curve analysis is probably that the
DNN model requires geological data only, which means the model can
be used in the exploration stage. In contrast, type-curve analysis requires
production data to predict EUR, which is only available after a region
has been producing for a while.

Input layer Output layer

Hidden layer 1 Hidden layer 3
Hidden layer 2
Fig. 4.15 Schematic of a DNN architecture with three hidden layers.