Precipitation forecasting often struggles with zero-inflated data where non-zero events are rare. The Zero Inflation Diffusion Framework (ZIDF) proposes a novel approach by integrating Gaussian perturbation for smoothing, which enhances the predictability of sparse datasets. Additionally, ZIDF employs Transformer-based methods to capture temporal patterns effectively, coupled with diffusion-based denoising to maintain the integrity of the data structure.
In tests conducted with observational precipitation data from South Australia, ZIDF shows significant reductions in Mean Squared Error (MSE) by up to 56.7% and Mean Absolute Error (MAE) by 21.1% when compared to existing models like the Non-stationary Transformer. These results underscore the framework’s robust handling of sparse time series data and indicate its applicability in other fields facing similar zero inflation challenges, highlighting the need for advanced statistical methods in environmental data science.
👉 Pročitaj original: arXiv AI Papers
