Rethinking multivariate modeling in long-term forecasting: an efficient univariate framework with power decomposition...

Area of Science:

• Machine Learning
• Time Series Analysis
• Data Science

Background:

• Long-term time series forecasting (LTSF) is crucial for industries.
• Current multivariate approaches face challenges like data drift and noise.
• Univariate modeling offers a potentially more effective and efficient alternative.

Purpose of the Study:

• To propose PDCNet, a novel univariate model for LTSF.
• To enhance predictability and adapt to pattern drifts efficiently.
• To demonstrate the superiority of univariate modeling for LTSF tasks.

Main Methods:

• Developed Power Decomposition (P3D) to disentangle time series based on data power distribution.
• Introduced Predictability-ACF analysis for optimal decomposition thresholds.
• Designed a lightweight online Post-Calibration mechanism to adapt to pattern drifts without retraining.

Main Results:

• PDCNet consistently outperforms state-of-the-art models on univariate LTSF tasks.
• Achieved top-tier multivariate performance through simple aggregation.
• P3D improved accuracy by an average of 15% in 68.57% of cases; calibration further boosted accuracy by 16% in calibrated regions.
• Reduced GPU memory usage by up to 95% compared to multivariate models.

Conclusions:

• Capturing internal temporal dependencies within each variable is an efficient and practical approach for LTSF.
• PDCNet serves as a competitive baseline, offering superior performance with a significantly reduced memory footprint.
• The proposed methods enhance predictability and adaptability in time series forecasting.