Deep insight: an efficient hybrid model for oil well production forecasting using spatio-temporal convolutional networks and Kolmogorov–Arnold networks

Abstract

As global energy demand grows, the oil and gas industry faces increasing challenges in optimizing production while achieving sustainability. Accurate oil well production forecasting is essential for effective resource management and operational decision-making. However, traditional mathematical models struggle with the nonlinear and dynamic characteristics of production data, while existing hybrid neural networks often lack sensitivity to operational changes and suffer from overcomplexity due to numerous parameters. This study proposes a novel hybrid model, TCN-KAN, combining temporal convolutional networks (TCN) and Kolmogorov-Arnold networks (KAN), to address these challenges. By integrating feature selection informed by reservoir engineering expertise and Spearman correlation analysis, the model effectively reduces input dimensionality while ensuring physically meaningful feature representation. Experimental results demonstrate the TCN-KAN model's superior ability to capture nonlinear interactions and long-term temporal dependencies, achieving the highest predictive accuracy among tested models. Additionally, a modified whale optimization algorithm (WOA) is employed for hyperparameter tuning, further enhancing the model's robustness. Validation on Volvo oil field data (2008-2016) highlights the model's operational sensitivity and practical value, providing actionable insights for optimizing oilfield management strategies.