Physics-informed sequential data augmentation for three-phase flow modeling

Accurately describing oil-gas-water three-phase flow characteristics of multiphase pumps is desirable for the structure design and condition adjustment in crude oil extractions. Considering limited labeled data available and complex flow process features, a physics-informed sequential data augmentation method is proposed to generate enough samples for the modeling of three-phase flow processes. First, using the process knowledge from the multiphase transportation and quantified uncertainty from Monte Carlo dropout strategy, the complex regions of each condition are respectively identified and locally augmented samples. Then, all collected data and generated candidates are utilized as the original samples of the time variational autoencoder model for the global augmentation. Finally, a gated recurrent unit model is built to predict the flow rates of multiphase pumps. Experimental results demonstrate the proposed method can capture the complex flow characteristics by introducing physics-informed augmented data. • A physics-informed two-stage data augmentation method is proposed to generate enough samples for the modeling of three-phase flow processes. • The complex dynamic features of three-phase flow processes can be elaborately described by the incorporation of physical knowledge and quantified uncertainty. • The complex regions of each condition are identified and locally augmented samples in the first stage. • All samples of different conditions are collected for the further global augmentation using time variational autoencoder architecture. • The proposed method can better capture both local and global flow characteristics in different conditions.