Edinburgh Napier University

  This research presented in this thesis examines the application of feed forward neural
networks to the performance control of a gas transmission compressor. It is estimated that a
global saving in compressor fuel gas of 1% could save the production of 6 million tonnes of
CO2 per year.
Current compressor control philosophy pivots around prevention of surge or anti-surge
control. Prevention of damage to high capital cost equipment is a key control driver but
other factors such as environmental emissions restrictions require most efficient use of fuel.
This requires reliable and accurate performance control.
A steady state compressor model was developed. Actual compressor performance
characteristics were used in the model and correlations were applied to determine the
adiabatic head characteristics for changed process conditions.
The techniques of neural network function approximation and pattern recognition were
investigated. The use of neural networks can avoid the potential difficulties in specifying
regression model coefficients. Neural networks can be readily re-trained, once a database is
populated, to reflect changing characteristics of a compressor.
Research into the use of neural networks to model compressor performance
characteristics is described. A program of numerical testing was devised to assess the
performance of neural networks. Testing was designed to evaluate training set size, signal
noise, extrapolated data, random data and use of normalised compressor coefficient data on
compressor speed estimates. Data sets were generated using the steady state compressor
model. The results of the numerical testing are discussed.
Established control paradigms are reviewed and the use of neural networks in control
l'Iystems were identified. These were generally to be found in the areas of adaptive or model
predictive control. Algorithms required to implement a novel compressor performance
control scheme are described. A review of plant control hierarchies has identified how the
Mdwme might be implemented. The performance control algorithm evaluates current
!,!'Ocells load and suggests a new compressor speed or updates the neural network model.
{'ornpressor speed can be predicted to approximately ± 2.5% using a neural network
h,lt1l'd model predictive performance controller. Comparisons with previous work suggest
l'1l1t 'IlUal global savings of 34 million tonnes of CO2 emissions per year. A generic, rotating
machinery performance control expert system is proposed