Intelligent predictions for flow pattern and phase fraction of a horizontal gas-liquid flow

In-situ measurement of phase fraction of a gas-liquid flow is closely related to the production efficiency in natural gas extraction. However, the measurement accuracy can be affected by the co-existed multiple flow patterns. This study proposes an intelligent strategy that identifies the flow pattern followed by a phase fraction prediction. For flow pattern recognition, we establish a bidirectional long short-term memory (BI-LSTM) network whose inputs are time-series phases of a Radio Frequency Sensor (RFS). The accuracy is 92.4 % over four classical flow patterns. The time-series phases of RFS are agreed well with the axial imaging from a Wire-Mesh Sensor (WMS). Two predictive models are developed for gas fraction: dimensionless analysis model (DAM) based on RFS and gas Froude number, and neural network model (NNM) with the phases of RFS and the recognized flow pattern. The mean absolute errors (MAE) are 3.2 % and 1.5 % for DAM and NNM, respectively. It is concluded that a NNM, incorporated with RFS and flow pattern by BI-LSTM, can intelligently predict gas fraction with high-accuracy. As the present strategy decouples the pattern recognition and gas fraction prediction into two networks, the complexity of a NNM is reduced which benefits the in-situ measurement practice.