Summary In cementing operations, conventional methods for predicting wellhead pump pressure, which rely on empirical and numerical models, often fall short regarding real-time accuracy and adaptability to different geological blocks and cementing techniques. These shortcomings hinder precise control of the wellhead pressure range, impacting operational safety, efficiency, and on-site personnel guidance. Furthermore, traditional offline machine-learning models cannot handle concept drift due to changing downhole fluid conditions during cementing. This paper frames wellhead pump pressure prediction as an online regression forecasting problem that utilizes the strengths of incremental and online ensemble learning with the Hoeffding tree regressor to overcome these challenges. First, data features are partitioned into subsets to create base learners, and then global and local models are deployed to address concept drift. The dynamic integration of these models significantly improves adaptability and performance. Experimental results confirm the superiority of our method over existing approaches. Compared with the suboptimal online models, mean squared error (MSE) is reduced for pump pressure prediction in five cementing data sets by 74.65%, 75.09%, 59.27%, 65.21%, and 70.01%, respectively. Additionally, MSE is reduced by 52.46%, 51.18%, 42.93%, 57.12%, 56.23%, 64.83%, and 79.51%, respectively, in seven open data sets, which showcases the broad applicability of our method to other regression tasks.
Read full abstract