Lithology identification supports reservoir characterization and well-log interpretation. Random sample splits can overestimate performance because adjacent samples from the same well are correlated, while logging-channel differences may be underused. We propose a Residual Feature-Gated BiGRU framework (RFG-BiGRU) for blind-well lithology identification. The framework combines singular spectrum analysis for signal conditioning, a learnable feature gate, residual feature fusion, and bidirectional GRU sequence modeling with temporal attention. RFG-BiGRU was evaluated on the six-class North Sea subset of the FORCE 2020 well-log dataset. The two designated blind wells, 31/2-19 S and 31/2-8, were excluded from model development, while the remaining 13 wells were used for development. RFG-BiGRU achieved a mean accuracy of 89.85%, a mean Macro-F1 of 55.69%, and a mean weighted F1-score of 89.16%. Supplementary five-fold evaluation across the 13 development wells yielded a mean Macro-F1 of 37.69% for RFG-BiGRU under a fixed configuration, with differences in preprocessing and weighting between model pipelines. In a separate matched five-seed comparison on the two blind wells, Full RFG-BiGRU, Plain BiGRU, and BiGRU + Temporal Attention achieved mean Macro-F1 values of 55.78%, 55.35%, and 55.71%, respectively. The Full model’s mean difference from Plain BiGRU was 0.43 percentage points with a sample standard deviation of 2.50 points, so the single-run 3.32-point difference is not treated as a stable repeated-run advantage.