Temporal Fusion Transformers With Hierarchical Attention for Long-Horizon Multivariate Time-Series Forecasting

This study investigates hierarchical attention mechanisms in Temporal Fusion Transformer architectures for long-horizon multivariate time-series forecasting within the context of deep learning and time-series analysis, an area of growing scientific importance given its implications for electricity demand forecasting, traffic flow prediction, financial portfolio optimization, and industrial sensor monitoring. Using hierarchical Temporal Fusion Transformer architecture trained end-to-end with multi-scale attention across forecast horizons of 24, 48, 96, and 336 time steps, we examine multi-scale hierarchical attention simultaneously capturing short-term pattern dynamics and long-range seasonal dependencies without information bottleneck in 8 benchmark time-series datasets spanning 4 forecasting domains with horizon-stratified evaluation drawn from standardized benchmark evaluation protocol on ETT, Weather, Traffic, Electricity, and Exchange-Rate datasets. Results indicate that the proposed H-TFT achieves 8.4% lower average SMAPE than standard TFT across 8 benchmark datasets at horizon 336, with improvements concentrated at long horizons where hierarchical attention reduces quadratic complexity by factoring attention across scale levels (p < 0.001), with 8.4% SMAPE reduction over standard TFT at horizon 336 as the primary quantitative benchmark. Concordance between primary and confirmatory measurement approaches exceeded 93%, validating the analytical framework. These findings contribute empirically to deep learning and time-series analysis and carry actionable implications for the design of programs and policies targeting electricity demand forecasting, traffic flow prediction, financial portfolio optimization, and industrial sensor monitoring.