Jinman Li ¹ Chunsheng Zhang ¹ Yang Lin ¹ Yimeng Liu ¹ Qingshuang Jin ^2* Tianhao Xiao ¹ Xiaoqi Liu ² Ying Zhang ²

• ¹ CNOOC (China) Limited Tianjin branch, Tianjin, China
• ² China University of Petroleum, Beijing, Beijing, China

Timely and accurate oil well production early warnings are essential for optimizing oilfield management and enhancing economic returns. However, most existing early warning methods rely on field-based experience to set thresholds and indicators, which lack a solid theoretical foundation. This study targets the high water cut stage of water-driven reservoirs, proposing the water cut as a key early warning indicator. We employ statistical methods to classify wells and determine adaptive warning thresholds tailored to different well characteristics. In addition, accurate water cut prediction is crucial for effective early warning. Traditional prediction methods often overlook the influence of static parameters, leading to prediction errors. To address this, we introduce a Long Short-Term Memory (LSTM) neural network model that integrates both dynamic and static well features, improving prediction accuracy. The proposed early warning model was applied to field wells, and the results demonstrated significant improvements in warning performance. Specifically, the classification-based adaptive warning thresholds and incorporation of dynamic and static parameters reduced false alarms and missed warnings, outperforming systems based solely on dynamic features.