EGU24-5526, updated on 08 Mar 2024
https://doi.org/10.5194/egusphere-egu24-5526
EGU General Assembly 2024
© Author(s) 2024. This work is distributed under
the Creative Commons Attribution 4.0 License.

Designing an Early-Warning System to Forecast Extreme Climate Conditions Using Data-Driven Approaches with Machine-Learning and Deep-Learning Methods

Afshin Shafei and Francesco Cioffi
Afshin Shafei and Francesco Cioffi
  • Sapienza university of Rome, Civil, Environmental and Hydraulic Engineering, Rome, Italy (francesco.cioffi@uniroma1.it)

This study introduces a methodology for enhancing early-warning systems (EWS) for climate variables such as temperature, humidity, and precipitation. These systems are crucial for predicting hydrological extremes, including heat waves and floods. Traditional forecasting methods face challenges due to the complex nature of climate systems, limitations of global circulation models, and computational demands, often resulting in predictions with coarse spatial and temporal resolutions.

Our approach integrates advanced Machine Learning (ML) models with comprehensive data collection for global climate forecasting and regional downscaling. The methodology centers on the use of the ERA5 reanalysis dataset from the European Center for Medium-Range Weather Forecasts (ECMWF) and the CMCC dataset, which provides high-resolution climate data.

The core of our global forecasting relies on FourCastNet, a cutting-edge deep learning model developed by NVIDIA. Utilizing Fourier Neural Networks, FourCastNet excels in generating high-resolution global climate forecasts quickly and accurately. It offers a lead time of up to 96 hours for various atmospheric variables, with a specific focus on precipitation forecasts up to 36 hours ahead. This model’s ability to handle complex climate patterns makes it ideal for initial global forecasting.

For regional downscaling, we employ Stacked Super-Resolution Convolutional Neural Network (SRCNN) and Super-Resolution Generative Adversarial Network (SRGAN) models, which are trained on the CMCC dataset. This dataset contains dynamically downscaled ERA5 reanalysis and has a 2.2 km spatial resolution and a 6-hourly temporal resolution, matching the temporal resolution of FourCastNet outputs. This compatibility enables seamless linking of global and regional forecasts. The downscaling aims to increase spatial resolution by eight times, providing detailed local climatic insights.

All computational models and simulations are conducted on the Google Cloud platform. This platform provides the necessary computational resources, including GPUs, to handle the large-scale processing of climate datasets and the execution of complex ML models efficiently.

In summary, this methodology combines advanced ML models and detailed data collection from both ERA5 and CMCC datasets for both global forecasting and regional downscaling. This integrated approach aims to deliver accurate, high-resolution predictions of climate variables, significantly enhancing the capabilities of early-warning systems. The selection of suitable high-resolution datasets for training downscaling models is a key step, ensuring the generation of detailed regional forecasts.

How to cite: Shafei, A. and Cioffi, F.: Designing an Early-Warning System to Forecast Extreme Climate Conditions Using Data-Driven Approaches with Machine-Learning and Deep-Learning Methods, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-5526, https://doi.org/10.5194/egusphere-egu24-5526, 2024.