Variational Hierarchical N-BEATS Model for Long-Term Time-Series Forecasting.

Area of Science:

• Machine Learning
• Time Series Analysis
• Forecasting

Background:

• Long-term time-series forecasting (LTSF) presents significant challenges.
• Existing LTSF methods often overlook hierarchical timestamp information, hindering the capture of weekly and yearly patterns.

Purpose of the Study:

• To propose VH-NBEATS, an interpretable variational hierarchical model extending N-BEATS.
• To effectively incorporate hierarchical timestamp information into LTSF models.
• To enhance LTSF performance by capturing hierarchical seasonal and trending effects.

Main Methods:

• Developed VH-NBEATS, featuring a hierarchical timestamp block and a harmonic seasonal block.
• Integrated a variational autoencoder (VAE) to handle high time-series variability.
• Evaluated the model on seven diverse real-world datasets.

Main Results:

• VH-NBEATS achieved state-of-the-art (SOTA) performance across all tested LTSF datasets.
• Demonstrated the effectiveness of the hierarchical timestamp block in capturing complex temporal patterns.
• Showcased the plug-and-play capability of the hierarchical timestamp block with existing LTSF methods.

Conclusions:

• VH-NBEATS offers a significant advancement in LTSF by effectively utilizing hierarchical timestamp data.
• The proposed hierarchical timestamp block can be integrated with various forecasting models to improve performance.
• The variational approach enhances robustness in handling volatile time-series data.