📄 中文摘要
StaTS是一种用于概率时间序列预测的扩散模型，旨在解决固定噪声调度带来的中间状态难以逆转以及终态偏离近噪声假设的问题。通过交替更新，StaTS学习噪声调度和去噪器，包含谱轨迹调度器（STS），该调度器通过谱正则化学习数据自适应噪声调度，以提高结构保留能力和逐步改进预测效果。与以往方法不同，StaTS不仅依赖于时间域条件，还考虑了调度引起的谱降解，从而在不同噪声水平下更好地恢复结构信息。
StaTS：基于频率引导去噪器的自适应时间序列预测的谱轨迹调度学习
📄 English Summary
StaTS: Spectral Trajectory Schedule Learning for Adaptive Time Series Forecasting with Frequency Guided Denoiser
StaTS is a diffusion model designed for probabilistic time series forecasting that addresses the challenges posed by fixed noise schedules, which often lead to intermediate states that are difficult to invert and terminal states that deviate from the near noise assumption. By employing alternating updates, StaTS learns both the noise schedule and the denoiser, incorporating a Spectral Trajectory Scheduler (STS) that adapts the noise schedule based on data with spectral regularization. This approach enhances structural preservation and stepwise improvement in forecasting performance. Unlike previous methods that primarily rely on time-domain conditioning, StaTS also models schedule-induced spectral degradation, facilitating better structure recovery across varying noise levels.