A Hybrid Machine Learning - Numerical Weather Prediction Approach for Rainfall Prediction

Authors:

Abstract
Accurate rainfall predictions are extremely important to make informed decisions e.g., to adapt to floods, droughts and to manage water resources, plan agriculture etc. However, rainfall predicted by conventional numerical weather prediction (NWP) models frequently falls short of the required accuracy for many of these applications. In this paper, we propose a novel Machine Learning (ML) based hybrid approach that effectively leverages nowcasts and forecasts of atmospheric states – chosen based on both physics and data-driven principles – obtained from the NWP-based Global Forecast System (GFS) model. We investigate three ML-based problem formulations for rainfall prediction: rain-no rain classification, multiclass classification, and regression. For each of these formulations, the performance of three ML models to predict rainfall for lead-times up to 7-days is evaluated using observed daily rainfall collected over 4 months from 55 stations across India. The proposed approach yields up to 25% improvement in rainfall prediction when compared with the GFS baseline.

This publication has 11 references indexed in Scilit: