TimeFlowDiffuser: A Hierarchical Diffusion Framework with Adaptive Context Sampling for Multi-horizon Time Series Forecasting

doi:10.1007/978-3-032-04549-2_20

TimeFlowDiffuser: A Hierarchical Diffusion Framework with Adaptive Context Sampling for Multi-horizon Time Series Forecasting

2026-01-01

0

PRE

AI

W

Wei Li *

DOI：10.1007/978-3-032-04549-2_20

Original

Original request for help

Save

Abstract

En 中文

Time series forecasting presents significant challenges in predicting complex temporal patterns across varying horizons. We introduce TimeFlowDiffuser, a novel framework that adapts diffusion models for time series forecasting through a hierarchical structure with adaptive context sampling. Our approach incorporates: (1) a Hierarchical Temporal Resolution module that processes time series at multiple scales; (2) an Adaptive Context Sampling mechanism that dynamically selects relevant historical context; (3) a Frequency-Aware Conditioning component that handles different frequency components; and (4) a Multi-Horizon Generation strategy for efficient prediction at various time horizons. Experiments on five benchmark datasets demonstrate that TimeFlowDiffuser consistently outperforms state-of-the-art methods, achieving average improvements of 7.2% in MSE and 6.3% in MAE, with particularly strong performance on long-horizon forecasting tasks. Our approach shows enhanced robustness to missing values and distributional shifts, with computational trade-offs discussed in Sect. 5.6.

Keywords:

Time series forecasting

Diffusion models

Multi-scale modeling

Adaptive sampling

Multi-horizon prediction

TimeFlowDiffuser: A Hierarchical Diffusion Framework with Adaptive Context Sampling for Multi-horizon Time Series Forecasting

Abstract

Journal

Organization