BROWSE
Detail
Journal Article
시계열 파운데이션 인공지능 모델 기술 동향
- Issue Date
- 2026-04
- Citation
- 전자통신동향분석, v.41, no.2, pp.63-72
- ISSN
- 1225-6455
- Publisher
- 한국전자통신연구원
- Language
- Korean
- Type
- Journal Article
- Abstract
- Time-series forecasting, which aims to predict future values from historical observations, is fundamental to applications in domains such as weather, energy, and finance. Recent advances in deep learning have shifted research focus from traditional statistical approaches toward neural forecasting models, particularly transformer-based architectures. However, deploying high-performance neural forecasters in real-world settings often requires large, domain-specific datasets, which are difficult to obtain in security- and privacy-sensitive sectors such as energy. Inspired by large-language models that generalize across tasks without domain-specific fine-tuning, recent work has introduced time-series foundation models (TSFMs) trained on large and diverse collections of public time series data to enable zero-shot forecasting across domains. This study surveys the technical evolution of time-series forecasting, from classical statistical models to contemporary AI-based approaches. We review representative TSFMs developed in industry, comparing their architectures, training data scale, feature support, and degree of openness. Our analysis reveals common architectural patterns, data scaling trends, and open challenges in deploying TSFMs, particularly in data-constrained and security-sensitive domains. In addition, we discuss emerging large-scale datasets and benchmarking efforts,and outline practical considerations for adopting TSFMs in real-world forecasting applications.
- Keyword
- TSFM, 공변량 예측, 다변량 예측, 시계열 예측, 제로샷 예측, 확률 모델
- KSP Keywords
- Architectural patterns, Classical statistical, Data Scaling, Domain-specific, Fine-tuning, Forecasting model, High performance, Large-scale datasets, Open challenges, Real-world, Statistical Model
This work is distributed under the term of Korea Open Government License (KOGL)
(Type 4: : Type 1 + Commercial Use Prohibition+Change Prohibition)
(Type 4: : Type 1 + Commercial Use Prohibition+Change Prohibition)
