Extreme weather events such as heavy storms and heavy precipitation have a large impact on our society. Due to both societal developments and climate change, it is expected that their impact will increase in the future. Especially in the case of rapidly developing extreme weather phenomena, society would benefit from forecasts at short time intervals, i.e., nowcasting. In principle, such forecasts can be made with an NWP model. However, an NWP model is expensive to run both in computational resources and in run-time, so that in practice the highest rate at which forecasts can be produced is about once per hour with today’s technologies. Machine learning (ML) techniques have been used as a much more computational efficient replacement for, e.g., parameterizations in NWP models [2, 4, 5]. Therefore, using state-of-the-art ML techniques, it might also be possible to develop a system with which forecasts can be produced efficiently many times per hour and even on-demand.
The aim of this MSO project was to investigate whether it is possible to design a short-term forecasting system that updates the most recent forecast from the NWP model with the latest observational data using physics-guided ML techniques. Such a system fits well in the requirements for Early Warning System as described by the World Meteorological Organization  and it will be a prototype for making nowcast products for the KNMI Early Warning Center (EWC) .
In section 1 we consider the unique characteristics of weather forecasting data and their implications for the application of ML techniques. In section 2 we describe constraints that current ML techniques have in so far these are relevant to their application to weather forecasting. Hereafter, we describe how we designed the prototype and the results that we obtained with it in sections 3 and 4, and we conclude with a number of recommendations for an operational system based on our prototype design in section 5.
C. A. Severijns & G. A. Pagani
. Short-Term Forecasting of High-Impact Weather with Physics-Guided Machine Learning : Technical Report
KNMI number: TR-403, Year: 2023, Pages: 12