<p>As oil and gas exploration increasingly targets highly heterogeneous unconventional reservoirs, conventional well-log-based lithofacies identification methods often fail to adequately characterize local abrupt variations and long-range depositional dependencies within complex geological architectures such as thin interbeds. To address this challenge, this study proposes a hybrid KNN-Transformer model that integrates the local feature selection capability of the K-nearest neighbors (KNN) algorithm with the global sequence modeling strength of the Transformer architecture. The method first employs KNN to evaluate and weight locally discriminative features from raw logging curves, selecting the most sensitive parameters for lithofacies discrimination. Subsequently, the self-attention mechanism within the Transformer encoder is utilized to fuse long-range contextual information across the entire well section, thereby dynamically modeling complex lithofacies sequences. Experiments conducted on key well data from shale oil reservoirs in the southern Sichuan Basin demonstrate that the proposed model achieves a high identification accuracy of 95% on the test set, significantly outperforming benchmark models, including standalone Transformer, KNN, CNN, and LSTM. It exhibits particular superiority in preserving local details and delineating boundaries within thin interbeds. This work provides a high-precision and interpretable technical pathway for intelligent lithofacies identification in strongly heterogeneous reservoirs, offering substantial practical value for advancing automated log interpretation.</p>