IMPROVING THE METHOD FOR DETERMINING THE FUEL GAS CONSUMPTION FOR COMPRESSION USING PHYSICALLY MOTIVATED FEATURES

Vollständiger Abstract

Worum geht es in dieser Arbeit?

This paper addresses the pressing issue of improving the energy efficiency of gas transportation systems by developing a highly accurate method for predicting fuel gas consumption by gas compressor units at booster compressor stations. To assess the operating efficiency of booster compressor station equip-ment, a method for reliably determining the fuel gas consumption of gas compressor units is required. Traditional methods of standardization and forecasting, based on industry standards, are characterized by significant errors due to the use of average coefficients and a lack of adaptability to the actual technical condition of equipment under conditions of degradation of both a gas turbine engine and a centrifugal compressor. Numerous scientific papers have demonstrated that the use of machine learning methods can improve forecasting accuracy, including for fuel and energy resource consumption by equipment. However, the authors typically use physical parameters as training features. The authors systematically compared physical (directly measured) and physically motivated (calculated) parameters as input features for machine learning models. The study was conducted at two production facilities: a central booster compressor station (over 14,000 process modes) and a booster compressor station (over 119,000 process modes) within a comprehensive gas treatment facility. Twelve classes of models were tested, including linear regression and ensemble methods. It was found that aggregating thermodynamic parameters allows for a reduction in the input space dimension while preserving the physical essence of the process. The statistical significance of the selected features was confirmed using mutual information and Spearman's rank correlation methods. The results of a numerical experiment on delayed samples (20 %) demonstrate the undeniable advantage of physically motivated sets. The LightGBM model provided the best accuracy: the weighted mean absolute percentage error was 1.71 % for central booster compressor stations and 2.29 % for booster compressor stations. The obtained results exceed the accuracy of standard methods, which allows us to recommend these models for integration into intelligent dispatch control systems and digital twins of gas production and transportation facilities.