An innovative approach for oil well bottomhole pressure forecasting using Kolmogorov-Arnold Neural Networks (KANs): A case study in an offshore oilfield

This study introduces a novel method for bottomhole pressure (BHP) prediction utilizing the Kolmogorov-Arnold Network (KAN) machine learning algorithm. The KAN model is an interpretable neural network that employs learnable non-linear activation functions to model data peculiarities. To enhance reservoir management and optimize production results, the reservoir engineer must accurately forecast dynamic reservoir properties such as BHP and its temporal changes. BHP trends are conventionally acquired by utilizing pressure gauges in designated wells and incorporating the data into reservoir simulation or nodal analysis models. Conducting these pressure surveys can be challenging due to the associated expenses and missed production opportunities when performed on a flowing well lacking a permanent downhole gauge. The KAN model was utilized to forecast BHP for a single well in the Volve oilfield using exclusively surface-measured data and engineered features. We employed several machine learning algorithms, including linear regression, KAN, neural networks and tree-based models (XGBoost and GBM), enabling reservoir engineers to prioritize the acquisition of specific surveillance data based on the highly correlated features within the models. The models were evaluated using the coefficient of determination ( R 2 ), mean absolute error (MAE), and root mean squared error (RMSE) metrics, with the KAN model demonstrating superior performance compared to the other regression models. The KAN model also outperformed the full-field history-matched reservoir simulation model provided by the oilfield operator and gave the best results when compared to actual downhole gauge data, demonstrating its superior predictive capabilities.