Diffusion-Based Weather Time Series Forecasting: A Novel Conditional Guidance Approach for Probabilistic Predictions

Weather forecasting is a critical task for various sectors, including agriculture, transportation, and disaster management. Traditional methods like ARIMA and Long Short-Term Memory (LSTM)-based models have limitations in capturing non-linear dependencies and uncertainty in multi-variable time series data. This paper proposes a novel approach using diffusion models for probabilistic weather time series forecasting, conditioned on historical data and seasonal embeddings. Diffusion models, originally designed for image generation, are adapted here to generate future sequences by progressively adding and removing noise, enabling robust handling of stochasticity in climate variables such as temperature, humidity, wind speed, and pressure. The model incorporates conditional guidance to align predictions with observed patterns, improving accuracy for short-term forecasts and providing confidence intervals for better decision-making. Experiments on the Daily Delhi Climate dataset demonstrate superior performance, with the proposed model achieving MAE of approximately 1.0, RMSE of 1.3, and MAPE of 5% across variables, outperforming baselines like LSTM and ARIMA by 30-50%. Ablation studies confirm the efficacy of conditioning, and cross-validation shows low variance, indicating stability. This work contributes to generative modeling in time series, offering a scalable framework for uncertain environments. Future extensions could include spatio-temporal integration for regional forecasts.