Well Production Forecasting in Volve Field Using Kolmogorov–Arnold Networks

Accurate oil production forecasting is essential for optimizing field development and supporting efficient decision-making. However, traditional methods often struggle to capture the complex dynamics of reservoirs, and existing machine learning models rely on large parameter sets, resulting in high computational costs and limited scalability. To address these limitations, we propose the Kolmogorov–Arnold Network (KAN) for oil production forecasting, which replaces traditional weights with spline-based learnable activation functions to enhance nonlinear modeling capabilities without large-scale parameter expansion. This design reduces training costs and enables adaptive scaling. The KAN model was applied to forecast oil production from wells 15/9-F-11 and 15/9-F-14 in the Volve field, Norway. The experimental results demonstrate that, compared to the best-performing baseline model, the KAN reduces the Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE) by 78.5% and 89.5% for well 15/9-F-11 and by 80.1% and 91.8% for well 15/9-F-14, respectively. These findings suggest that the KAN is a robust and efficient multivariate forecasting method capable of capturing complex dependencies in oil production data, with strong potential for practical applications in reservoir management and production optimization.