TCN-Autoformer: A Hybrid Deep Learning Framework for High-Precision Curve Reconstruction

Summary Formation resistivity is a key parameter reflecting pore structure, fluid distribution, and formation heterogeneity and plays a vital role in subsurface fluid characterization. However, existing resistivity reconstruction methods struggle to effectively capture the nonstationarity and weak periodic characteristics of resistivity curves in complex formations, resulting in limited reconstruction accuracy under varying geological conditions. To address these challenges, we propose a hybrid deep learning model that integrates a temporal convolutional network (TCN) with an Autoformer architecture (TCN-Autoformer), enabling feature extraction across different depth intervals of formation resistivity curves and modeling of long-range dependencies. Specifically, the TCN module employs dilated causal convolutions to capture localized response patterns in resistivity sequences, such as short-scale fluctuations induced by fluid invasion effects and lithological variations. Meanwhile, the Autoformer module leverages an autocorrelation-based mechanism to identify periodic electrical variations arising from formation structural rhythms and spatial fluid distribution in both the time and frequency domains. This process enables an explicit decomposition of trend and periodic components, thereby enhancing the accuracy of formation resistivity curve reconstruction. Blind-well tests show that the TCN-Autoformer achieves R2 values of 0.9406 and 0.9547 for deep lateral resistivity (RD) and shallow lateral resistivity (RS) prediction, outperforming the standard transformer by 4.2% and 2.3%, respectively, while reducing mean squared error (MSE) and mean absolute error (MAE) by 20%–40%. These results indicate that the proposed model can effectively capture the complex features and weak periodic dependencies of formation resistivity curves, providing a quantifiable data-driven tool for high-precision reconstruction of resistivity curves.