Diffusion-Based Probabilistic Weather Time Series Forecasting

Breaking Through Traditional Model Limitations

Traditional weather forecasting models like ARIMA and LSTM have obvious limitations when handling multivariate time series, particularly in capturing nonlinear dependencies and quantifying prediction uncertainty. This research successfully broke through these limitations through an innovative approach—adapting diffusion models originally used for image generation to time series forecasting.

The core idea of diffusion models is to generate sequences by gradually adding and removing noise, a process that naturally reflects prediction uncertainty. The method provides conditional guidance on historical data and seasonal features, with seasonal features encoded through Fourier embedding. The benefit of this design is that the model can simultaneously generate future sequences of multiple meteorological variables such as temperature, humidity, and wind speed, capturing their complex interactions.

On the daily climate dataset from Delhi, India, the model demonstrated excellent performance. Mean Absolute Error (MAE) was approximately 1.0, Root Mean Square Error (RMSE) approximately 1.3, and Mean Absolute Percentage Error (MAPE) only 5%. Compared to traditional baseline models like LSTM and ARIMA, performance improved by 30-50%. More importantly, the model can generate probabilistic forecasts and confidence intervals, providing more valuable information for risk assessment and decision-making than single-point predictions.

Academic Publication

This research has been accepted by the 4th International Conference on Electronic Information Technology (EIT 2025).

Paper Information:

• Paper Number: 225081522130902170
• Author: Qianhe Yang
• Conference: 4th International Conference on Electronic Information Technology (EIT 2025)
• Conference Date: August 22-24, 2025
• Location: Chengdu, China
• Journal Code: IEEE (979-8-3315-7605-9)
• Indexing: IEEE Xplore, EI, Scopus

Complete Project Report