DEVELOPMENT AND TESTING OF UNIVERSAL PRESSURE DROP MODELS IN PIPELINES USING ABDUCTIVE AND ARTIFICIAL NEURAL NETWORKS

Determination of pressure drop in pipeline system is difficult. Conventional methods
(empirical correlations and mechanistic methods) were not successful in providing
accurate estimate. Artificial Neural Networks and polynomial Group Method of Data
Handling techniques had received wide recognition in terms of discovering hidden
and highly nonlinear relationships between input and output patterns. The potential of
both Artificial Neural Networks (ANN) and Abductory Induction Mechanism (AIM)
techniques has been revealed in this study by generating generic models for pressure
drop estimation in pipeline systems that carry multiphase fluids (oil, gas, and water)
and with wide range of angles of inclination. No past study was found that utilizes
both techniques in an attempt to solve this problem. A total number of 335 data sets
collected from different Middle Eastern fields have been used in developing the
models. The data covered a wide range of variables at different values such as oil rate
(2200 to 25000 bbl/d), water rate (up to 8424 bbl/d), angles of inclination (-52 to 208
degrees), length of the pipe (500 to 26700 ft) and gas rate (1078 to 19658 MSCFD).
For the ANN model, a ratio of 2: 1: 1 between training, validation, and testing sets
yielded the best training/testing performance. The ANN model has been developed
using resilient back-propagation learning algorithm. The purpose for generating
another model using the polynomial Group Method of Data Handling technique was
to reduce the problem of dimensionality that affects the accuracy of ANN modeling. It
was found that (by the Group Method of Data Handling algorithm), length of the pipe,
wellhead pressure, and angle of inclination have a pronounced effect on the pressure
drop estimation under these conditions. The best available empirical correlations and
mechanistic models adopted by the industry had been tested against the data and the
developed models.
Graphical and statistical tools had been utilized for comparing the performance of
the new models and other empirical correlations and mechanistic models.

Thorough verifications have indicated that the developed Artificial Neural Networks
model outperforms all tested empirical correlations and mechanistic models as well as
the polynomial Group Method of Data Handling model in terms of highest correlation
coefficient, lowest average absolute percent error, lowest standard deviation, lowest
maximum error, and lowest root mean square error.
The study offers reliable and quick means for pressure drop estimation in
pipelines carrying multiphase fluids with wide range of angles of inclination using
Artificial Neural Networks and Group Method of Data Handling techniques.
Graphical User Interface (GUI) has been generated to help apply the ANN model
results while an applicable equation can be used for Group Method of Data Handling
model. While the conventional methods were not successful in providing accurate
estimate of this property, the second approach (Group Method of Data Handling
technique) was able to provide a reliable estimate with only three-input parameters
involved. The modeling accuracy was not greatly harmed using this technique.