[KongchangAI]
ModelCorrectCast框架

CorrectCast

A decoupled residual correction framework for improving precipitation nowcasting accuracy. It freezes the backbone network of a base forecaster and trains only a lightweight corrector to estimate and subtract residuals. It supports multimodal context and scale-aware modulation, and provides a unified correction interface across different forecasting models.

Timeline (last 90 days)

Nov 18

CorrectCast is a decoupled residual correction framework that refines completed forecast outputs while keeping the base forecaster frozen.

Unverified50%
Nov 18

CorrectCast trains only a lightweight corrector, providing a unified correction interface for different forecast models.

Unverified50%
Nov 18

CorrectCast's corrector is conditioned on the base forecast, historical quantitative precipitation estimates (QPE) and precipitable water vapor (PWV) observations, and uses scale-aware modulation to estimate residual corrections.

Unverified50%
Nov 18

On the Mississippi River Basin (MRB) and Contiguous United States (CONUS) datasets, CorrectCast improves six base forecasters (covering extrapolation, deterministic methods and generative methods).

Unverified50%
Nov 18

CorrectCast's absolute gain in average CSI is 0.0233 on MRB and 0.0290 on CONUS.

Unverified50%
Nov 18

The paper does not discuss the corrector's performance on extreme precipitation events, nor its transferability across different climate regions.

Unverified50%

All Facts (6)

Source Articles