Feasibility of predicting performance characteristics of multi-stage centrifugal natural gas compressors, an effort on using data-mining techniques

Abstract

Performance curves of turbomachinery are the essential materials required for getting a deep insight into the operation of these facilities under different operating conditions. For driving these performance curves, a sort of expensive and rigorous experiments are required on the turbomachinery. This study investigates the possibility of predicting the performance curves of multi-stage centrifugal natural gas compressors with offshore applications using data mining techniques (e.g., Linear Regression), thereby eliminating the need for the required experiments for this objective. The main focus of this article has been specifically on predicting the polytropic efficiency and polytropic head performance curves of such compressors. The database of the project includes an extensive information bank related to different centrifugal compressors in various operating conditions. A large portion of the database is used for training the predictive models, and the remaining is used for tests and validations. Besides, cross-validation has also been done to ensure the reliability of the results. The study results prove that it is possible to predict the performance curve of centrifugal compressors with limited samples, although marginal errors are inevitable. The results showed that the trained model has a better performance in high flow rates. Accordingly, the maximum RMSE of the head in high flow rates was 3 kJ/kg and in low flow rates 11 kJ/kg. In addition, the maximum RMSE of efficiency in high flow rate was 0.02 and in low flow rate was 0.24.