Edinburgh Napier University

  This paper presents the application of feed forward neural networks to the performance modeling of a gas transmission compressor. Results of compressor model testing suggest that compressor speed can be estimated to within ± 2.5 %. The neural network property of function approximation is used to predict compressor speed for given process constraints and instrument input sets. The effects of training set size, instrument noise, reduced input sets and extrapolation from the training domain, are quantified. Various neural network architectures and training schema were examined. The embedding of a neural network into an expert system is also discussed. A neural network can be retrained to reflect changing compressor characteristics. A global saving in compressor fuel gas of 1% could prevent the production of 6 million tonnes of CO2 per year.