الفهرس 
IntroductionOnly 14 pages are availabe for public view
 |  | from | 128 |  |  |
| | | from | 128 | | |
Abstract
Oil and gas reservoirs are characterized by qualitative and quantitative values using
pressure transient analysis. The well test is conducted by creating a flow disturbance in the
well and recording the related response of the bottom‐hole pressure. Well test analysis
consists of two main phases: (1) the recognition of the entire reservoir model, and (2) the
model parameter estimation. The objective of this study is to apply the Artificial Neural
Network (ANN) technology to identify the reservoir model. A multilayer neural network had
been used with back propagation optimization algorithm for the recognition process. The
required training and test datasets have been generated by using the analytical solutions of
commonly used reservoir models. Nine networks have been constructed; each one
differentiates among six boundary models. Most commonly found reservoir models of
different inner, outer boundary and reservoir medium are included (e.g. vertical, fracture
and horizontal wells; homogenous, dual porosity and radial composite reservoirs; and
infinite, one sealing fault, two sealing faults, rectangle and circle boundaries).
Each of the ANN of the 9 networks has been constructed by one input layer with either
100 or 200 input nodes, two hidden layers; each has 0.5 of the number of the input nodes
and one output layer with six nodes characterizing the different reservoir boundary models.
Different network structures and training intensity were tested during this work to arrive at
optimum network design.
The performance of the proposed ANN has been examined by the actual field data in
addition to simulated noisy and smooth datasets by two testing modes; individual network
testing and comparative modes. The results indicate that the proposed multilayer neural
network can recognize the reservoir models with acceptable accuracy. This work shows that
distributing the commonly used reservoir models into 9 networks and by using two hidden
layers for each network with large training datasets can yield very good model identification
even with low noise level data.