Horizontal wells are widely used in unconventional resource development, where keeping the trajectory within a thin target interval is critical for reservoir contact and drilling-risk control. However, stratigraphic interpretation remains challenging because field decisions may be constrained by delayed measurements, noisy logging responses, and expert judgment. In this study, we propose a multi-scale boundary-enhanced Temporal Convolution-Transformer for real-time stratigraphic position prediction within an approximately 10-m-thick target interval. The model uses elemental logging-while-drilling data and natural gamma-ray responses to identify the current stratigraphic position relative to the target sublayers and adjacent overlying and underlying strata, while estimating signed distances to the upper and lower target boundaries. A main temporal convolutional block first extracts stable local sequential patterns from elemental and gamma-ray responses. A multi-scale boundary-enhancement branch then applies parallel one-dimensional convolutional kernels of 3, 5, and 9 along the logging sequence to enhance stratigraphic transition responses at different scales. The Transformer encoder further captures longer-range contextual dependencies along the horizontal well trajectory. To address class imbalance and difficult downward layer crossing, we adopt target-interval-focused sampling, reduce redundant far-boundary target-layer samples, preserve hard lower-boundary samples, and apply class-aware loss reweighting. Using 32 elemental components and gamma-ray responses, the model achieved a Macro-F1 score above 0.90 and a downward layer-crossing recall of 93% on two blind wells excluded from model training, with zero mutual misclassification between the adjacent overlying and underlying strata. Compared with Bi-LSTM and Transformer baselines, it improved the F1 score by 23% and the recognition of downward layer crossing by 41%. These results suggest that the proposed framework may support real-time stratigraphic warning and geosteering decision-making.