This study presents application of Temporal Fusion Transformer (TFT) architecture for sequence prediction in Directed Energy Deposition Laser-wire (DED-Lw) processes. The TFT architecture was adapted and optimized for binary classification of process stability using multivariate time series data from DED-Lw process parameters. The methodology combines enhanced training techniques including AdamW optimization with weight decay, cosine annealing learning rate scheduling, gradient clipping, and increased dropout regularization. Results demonstrate a good performance, with the optimized model achieving 98.88% overall accuracy, 100% precision for stable class detection, and perfect recall (100%) for unstable class identification; a critical requirement for manufacturing safety. This represents a 24.03% improvement over the initial implementation. Analysis of attention mechanisms reveals a transformation from highly localized temporal focus to distributed attention patterns that effectively capture longer range dependencies. The model demonstrates particular strength in transition detection, accurately identifying stability state changes with precise timing. Comprehensive evaluation across confidence thresholds confirms robust performance regardless of threshold selection. Feature importance analysis during transitions highlights specific process parameters that serve as leading indicators for stability changes, offering potential for early warning system development. This work establishes TFT-based classification as a promising approach for additive manufacturing process monitoring while acknowledging current limitations in dataset diversity that warrant further investigation for broader generalizability across different DED-Lw systems and materials.