Abstract: The majority of oilfields in China are currently undergoing high water cut development stages, highlighting the significance of accurate well production prediction in the oilfield development process. Precise well production prediction plays a pivotal role in evaluating well development potential, refining well operation strategies, and formulating effective oilfield development plans. This study aims to construct a short-term production data prediction model tailored for the water-flooding development stage. Leveraging the attention mechanism and the long short-term memory (LSTM) network algorithm, the model integrates various preprocessing techniques such as the K-nearest neighbor estimation method and the box plot method to refine production dynamic data. Subsequently, the Attention-based Long and Short-Term Memory (A-LSTM) algorithm, LSTM algorithm, and Support Vector Regression (SVR) algorithm are employed to predict the daily oil production of individual wells using short-term production data. The Huber loss function serves as a metric to gauge the disparity between predicted values and actual values during the prediction process. Results indicate that the well production prediction model based on the A-LSTM algorithm outperforms the other two algorithms in accurately predicting the daily oil production of individual wells with short-term production data. The application of this model holds promise for providing further guidance in oilfield water injection development endeavors.
 

Attention mechanism; Machine learning;·production prediction;·oilfield development