Data-Driven Models for Precipitation Forecasting: A Two-Step Approach

Recent advances in data-driven weather models have shown that these models can match or even outperform traditional dynamical models for certain variables, such as 500 hPa geopotential height and 850 hPa temperature. However, when it comes to precipitation forecasting, the success of these models has been limited. Precipitation forecasting poses unique challenges due to its highly non-linear nature and its non-normal distribution. Additionally, many prominent data-driven weather models rely on ERA5 reanalysis data, which often fails to accurately represent the true characteristics of precipitation.

To address these challenges, we propose a two-step approach. First, we leverage existing data-driven weather models to forecast wind, temperature, and humidity at 13 vertical levels for specific locations. These forecasts are then used as input features for fully-connected neural networks, which are trained to diagnose precipitation. We use IMERG (Integrated Multi-satellitE Retrievals for GPM) data as the target dataset, allowing for improved accuracy in precipitation forecasts by building upon the strengths of existing models.

Our preliminary tests using over-parameterized neural networks achieve a median absolute error of 0.12 mm/hr in the West Pacific region. These results, along with findings from previous literature, indicate that our approach outperforms generalized linear models, random forests, and under-parameterized neural networks. This demonstrates the potential for improvements in precipitation forecasting through our two-step method, which integrates data-driven weather models with standalone over-parameterized neural networks specifically designed for precipitation forecasting.