Spatio-temporal graph neural networks to improve precipitation forecasts from numerical models

• Published in: Soft computing (Berlin, Germany) Vol. 29; no. 9-10; pp. 4481 - 4494
• Main Authors: Yousaf, Umair, De Rango, Alessio, Furnari, Luca, D’Ambrosio, Donato, Senatore, Alfonso, Mendicino, Giuseppe
• Format: Journal Article
• Language: English
• Published: Heidelberg Springer Nature B.V 01.05.2025
• Subjects: Accuracy; Artificial intelligence; Artificial neural networks; Climate change; Error correction; Graph neural networks; Machine learning; Meteorological data; Neural networks; Numerical models; Numerical weather forecasting; Precipitation; Spatial resolution; Weather forecasting; Weather stations; United States > US; China
• ISSN: 1432-7643; 1433-7479
• DOI: 10.1007/s00500-025-10635-7

Weather forecasting has always been a fascinating challenge due to the randomness, difficulty and multiple intersections in the spatiotemporal distribution of several atmospheric processes. The classical approach for weather forecasting is based on numerical weather prediction (NWP). However, this approach is subject to errors and criticism as it is unable to predict severe events accurately, especially if located in complex areas characterized by steep orographic effects or strong air-sea interactions. In this work, a Deep Learning (DL) methodology has been applied to improve the one-day ahead precipitation accuracy of the Weather Research and Forecasting (WRF) NWP system by correcting the prediction error. The WRF data consists of a spatial resolution of 2 km and refers to a portion of the Calabria region (Italy). A weather station network of 22 gauges has been considered inside this latter area. The meteorological data for the whole year of 2021 and 4 months of 2022 is considered for training and evaluation respectively. The developed DL model is based on Spatio-Temporal Graph Neural Network (called WRF-GNN). The improved prediction has been compared with observed precipitation data from the rain gauge network, the WRF output, Random Forest (WRF-RF), XGBoost (WRF-XGB), and another Artificial Neural Network (WRF-ANN) model. The WRF-GNN significantly enhanced the prediction accuracy compared to the WRF and WRF-ANN models, with an improvement from + 16 to + 34% with respect to WRF, by minimizing the error compared to observations.