Latent Diffusion Model for Quantitative Precipitation Estimation and
Forecast at km Scale
Abstract
Numerical weather prediction models often struggle to accurately
simulate high-resolution precipitation processes, due to resolution
limits and difficulties in representing convection and cloud
microphysics. Data-driven methods often offer more accurate
approximation of these unresolved processes by learning from
high-fidelity referential data. Yet, existing approaches fail to yield
fine-resolution, spatially coherent, reliable ensemble
simulations/predictions, particularly for high-impact extreme events. To
address these limitations, we develop a latent diffusion modeling (LDM)
framework for quantitative precipitation estimation and forecast. The
LDM leverages low-resolution (25 km) circulation and topographic
information to estimate precipitation at a 4 km resolution. The latent
diffusion model (LDM) learns the probability distribution of
precipitation patterns by first compressing high-resolution spatial data
into a compact, Quasi-Gaussian latent space. It then gradually refines
estimates through a deep neural network parameterized reverse diffusion
process, effectively capturing complex precipitation dynamics and
delivering superior ensemble predictions. This approach enables LDMs to
outperform traditional deep learning models like convolutional neural
nets and generative adversarial neural nets, particularly for extreme
events, while avoiding issues such as mode collapse, blurring artifacts,
or underestimation of extremes. Compared to the dynamical method (WRF
and GFS), the LDM offers significant performance gains, particularly for
extreme precipitation events, improving over 30% in root mean squared
error and over 40% in critical success index. For the extreme
precipitation (> 300 mm/d) in California on October 25,
2021, LDM can provide effective forecasting up to 7 days in advance,
forced by circulation prediction from a data-driven weather forecasting
model.