Review on Time Series Forecasting Based on Diffusion Models

Time-series prediction serves as a cornerstone for pattern mining and trend forecasting across diverse domains; this project can be applied to electric energy prediction, the financial field, and the new media field. However, it faces inherent challenges such as non-stationarity, long-range dependencies, and high-dimensional complexity. This paper presents a comprehensive survey of diffusion models in time-series prediction, focusing on clarifying the fundamental differences between traditional methods, conventional deep learning models, and the innovative diffusion paradigm. First, it constructs a "data-task-method" analytical framework to systematically map time-series characteristics to prediction tasks and corresponding method limitations. Second, it proposes a novel "module-architecture" taxonomy for diffusion model derivatives, distinguishing between mechanism-optimized and scenario-adapted methods. Third, it conducts multi-dimensional performance comparisons, verifying that diffusion models outperform traditional and deep learning baselines by 12%-25% in metrics like RMSE and MAE across traffic, meteorology, and industrial scenarios. Finally, future directions are outlined, including enhancing model interpretability, fusing multimodal data, integrating with large pretrained models, and lightweight design for edge deployment.