Abstract
Long-term time-series forecasting (LTSF) has received an increasing
attention for its significant challenges and real-world applications.
However, the previous studies under-explore the hierarchical timestamp
information in LTSF. This information is crucial, especially for LTSF as
failing to incorporate it may result in missing the global perspective
of time series and important long-term trending effects, such as weekly
and seasonal patterns. Therefore, we propose an interpretable
hierarchical model called VH-NBEATS, which advances the N-BEATS model by
addressing the aforementioned problem. VH-NBEATS comprises two essential
blocks: the hierarchical timestamp block and the harmonic seasonal block
to capture multi-diluted and trending effects. To address the high
variability of time series, VH-NBEATS involves a stochastic autoencoder
which significantly improves the standard deterministic approach. The
experimental results are evaluated on five real-world datasets, showing
state-of-the-art results for LTSF. We also prove that the VH-NBEATS
framework can be easily incorporated into other ones, such as PathTST,
leading to enhanced performance.