BROWSE
Detail
Conference Paper
Improving Zero-Shot Energy Forecasting: The Impact of Multi-Step Component Analysis on Time-Series Foundation Models
Cited - time in 
Share 
Share
- Authors
- SangJoon Lee, Yumin Hwang, Seok Kap Ko
- Issue Date
- 2026-10
- Citation
- International Conference on Information and Communication Technology Convergence (ICTC) 2026, pp.1-4
- Publisher
- IEEE
- Language
- English
- Type
- Conference Paper
- Abstract
- Time-series forecasting is critical for stable power system operations, yet deploying high-performance neural forecasters often faces challenges due to data scarcity and complex periodicity. Although recent Time-Series Foundation Models (TSFMs) enable robust zero-shot forecasting across domains, their accuracy in energy sectors can be further refined to handle distinct variability. This paper proposes a hybrid forecasting method that integrates TSFMs with multi-level STDR decomposition (Trend, Seasonality, Dispersion, and Remainder). Using public electricity peak demand data and synthetic signals, we evaluated the effectiveness of decomposing multiple periodicities—such as weekly and monthly cycles—prior to forecasting. Our experimental results demonstrate that incorporating STDR decomposition consistently improves forecast accuracy compared to direct foundation model predictions. Furthermore, for time-series data containing nested periodic signals, multi-level decomposition is shown to provide additional predictive gains. We also conduct a sensitivity analysis on synthetic trend amplitudes, revealing that the efficacy of multi-level decomposition is closely tied to the relative signal power of the periodic components. These findings provide practical insight for the adoption of foundation models in real-world energy applications where structural clarity and predictive robustness are paramount.
- Keyword
- Time-series forecasting, Time-series foundation model, Power Spectral Density, STDR de- composition, Peak energy demand
- KSP Keywords
- Data scarcity, Electricity peak demand, Energy applications, Energy forecasting, Forecast accuracy, Forecasting methods, High performance, Hybrid forecasting, Multi-level, Multi-step, Peak energy demand