Hybrid AI Models for Wet Gas Compressor Modelling

Abstract Wet gas compressors can play an important role in achieving more effective and sustainable drainage of existing gas reservoirs. Research in this niche area is about complex multiphase fluid-thermodynamics and involves sophisticated and very expensive test equipment and testing. A better understanding of the mechanisms that occur when a multiphase fluid is manipulated by turbo impellers is essential not only for optimization of the compressor design, bult also for developing models to predict compressor performance and then enabling effective control and monitoring during field operation. Traditionally the effort of learning wet gas compressor performance has been based on conventional turbomachinery theory combined with extensive full-scale testing on wet gas conditions. Performance interpretation under such condition relays on empirical data and involves analysis of complex multi-dimensional patterns. The main goal of this paper is to describe the building of a surrogate model for wet gas compressors to predict their performance using experimental test measurements. The uniqueness of the proposed approach lays in hybrid method combining physics-based models for dry gas compressors and data obtained from testing. The AI model used to build surrogate model is a particular type of neural network called tabular neural network, TabNet. As its name suggests, this model is adapted to tabular data which is the type of data obtained in the test facilities. Physics laws can be integrated into AI models in various ways. In this work, the loss function is modified to enforce affinity laws governing dry gas compressor data, ensuring the model balances data fitting with adherence to these laws. The approach was evaluated across multiple compressors, and the results demonstrated that TabNet is capable of achieving high accuracy when compared to traditionally computed ground truth values. Furthermore, in scenarios with limited data availability, the physics-informed TabNet significantly outperformed the classical TabNet, delivering substantially more accurate predictions and offering a viable solution where the standard approach struggled to produce reliable results.