Physics-augmented water cut measurement and forecasting based on swirling dual-differential pressure and Transformer-BiLSTM

To address the difficulty of real-time water-cut monitoring and the challenge of anticipating future changes, this study proposes a physics-augmented method for water-cut measurement and prediction. First, an online measurement device based on a swirl dual-differential-pressure model is installed in the field to perform real-time wellhead measurements and data acquisition. Continuous water-cut data are obtained from the differential-pressure signals generated when the wellhead fluid flows through the swirling structure, and a filtering algorithm is applied to mitigate noise caused by changing operating conditions, keeping the daily mean relative error within 5 %. Then, based on the long-term monitoring data collected during field tests, a Transformer–BiLSTM (Bidirectional Long Short-Term Memory) prediction model is introduced, which jointly uses pressure, temperature, flow rate, and historical water cut as input features. On this basis, a dual-model framework is built that enables real-time water-cut prediction driven by multiple features and, when future feature values are unavailable, allows trend forecasting of water cut. The results show that, in the multi-feature prediction mode, most relative errors are within 1 %, while one-week-ahead water-cut forecasts maintain relative errors below 10 % and can effectively track the evolution of water cut. Generalization tests have demonstrated the applicability of the proposed method across different fluid types. However, constrained by the data acquisition duration and sample size, the model's generalization capability under extreme operating conditions warrants further exploration. The findings demonstrate that combining the swirl dual-differential-pressure model with the Transformer–BiLSTM network can significantly improve the accuracy and continuity of water-cut monitoring and short-term prediction, providing reliable support for production optimization and risk early warning in digital oilfields.