The MG-TSD model addresses the challenge of instability in diffusion models for time series forecasting by utilizing multi-granularity data as targets. The model outperforms existing methods and provides reliable predictions without relying on external data. Extensive experiments on real-world datasets demonstrate the effectiveness of the approach.
Diffusion probabilistic models have shown promise in generative time series forecasting but face challenges due to their stochastic nature. The MG-TSD model leverages different granularity levels within data to guide the learning process, resulting in more precise forecasting results.
The proposed model introduces a novel multi-granularity guidance diffusion loss function and offers a practical implementation method. By using coarse-grained data as targets, the model reduces variability and improves prediction quality across various granularity levels.
Overall, the MG-TSD model demonstrates superior performance compared to existing time series prediction methods by effectively utilizing multi-granularity information to stabilize diffusion models for accurate forecasting.
翻譯成其他語言
從原文內容
arxiv.org
深入探究