The extreme precipitation event of July 2021 and the ensuing floods caused 41 deaths and over 2 billion euro in damages in Belgium alone. An impact-based early warning system can improve preparedness and reduce the societal and economic impacts of such extreme precipitation events, as it allows local authorities, emergency services and industry to make better-informed, timely decisions.

Early warning systems require forecasts for weeks ahead, as provided by global numerical weather prediction (NWP) models. However, such models generally fail to capture precipitation extremes, in part due to their limited spatial resolution. Km-scale NWP models, integrated in seamless observation-driven short-term prediction systems such as RMIB`s Project IMA, better represent these extremes, making them more suitable for high-resolution (urban) flood models. However, their time horizon is limited to 1-2 days, making them unsuitable for early warnings and proper management of extreme events.

This project aims to meet the need for a consistent and calibrated ensemble forecasting system by seamlessly combining models at different time horizons, and to integrate them in impact models.

Objectives

1. The realization of a real-time deep learning (DL) based multimodal quantitative precipitation estimate (QPE)
2. The creation of a seamless precipitation forecast product that is:
◦ frequently d, to ingest the most recent observations
◦ probabilistic and ensemble-based,
◦ sharp and calibrated
◦ based on state-of-the-art ensemble prediction systems: pySTEPS nowcasting, ACCORD’s limited area NWP, and ECMWF’s medium-range ensemble forecasts.
3. Integration of these forecasts in hydrological models to enable impact-based early warnings for extreme precipitation events.

Methodology

This project leverages the deep learning (DL) approach, highly effective for data fusion of multimodal observations, by extending it to the fusion or blending of three types of forecasts in a residual learning paradigm:
• Nowcasts based on multimodal observations,
• Short-range (0-3 days) NWP
• Medium-range (0-2 weeks) global NWP forecasts
We will integrate nowcasts for the first time driven by a multimodal DL-based QPE in our seamless prediction system.

Downscaling, calibration and blending will be considered in a single framework, and different DL architectures such as Unets and GANs will be compared.

The coupling to hydrological models will be implemented relatively early in the project, to ensure
that impact models can optimally use the ensemble forecasts at their full resolution and time
range, and to enable impact-based validation.

Runtime: 2022 - 2026