Abstract
Accurate prediction of reservoir parameters is crucial for enhancing oil exploration efficiency and resource utilization. Although existing deep learning methods have made some progress in reservoir parameter prediction, they still face accuracy limitations in multi-task prediction. Additionally, the black-box nature of these models limits their interpretability, impacting trust and acceptance in practical applications. To address these challenges, this study proposes a Multi-TransFKAN model based on a Transformer architecture and an improved Kolmogorov-Arnold Network (KAN) framework for reservoir parameter prediction and interpretability analysis. By integrating Fourier functions in place of B-spline functions within the KAN framework, the model effectively captures complex periodic and nonlinear features. Combined with Monte Carlo Dropout and SHAP frameworks, it further enhances prediction accuracy and interpretability. Experimental results show that in test wells, the average RMSE values for porosity (PHIF), shale volume (VSH), and water saturation (SW) are 0.053, 0.049, and 0.062, respectively. Compared to other methods, the proposed model reduces RMSE by 52.5% and increases R(2) by 10.7%, demonstrating significant improvements in prediction accuracy. These findings highlight the model's capability to deliver more reliable predictions and a clearer understanding of the factors influencing reservoir parameters. Therefore, the Multi-TransFKAN model not only enhances the accuracy of reservoir parameter prediction but also improves model transparency and reliability in real-world applications through advanced interpretability techniques.