Combining Deep Learning with Variational Mode Decomposition and Multifractals for Urban Precipitation Nowcasting
Résumé
Short-term precipitation forecasting for the next 0 to 6 hours (also known as nowcasting) is essential for decision-making of weather-dependent activities to mitigate socio-economic impacts. Accurate and reliable precipitation nowcasting can serve as an early warning of massive flooding, as well as a guide for water-related risk management based on scenarios. Although precipitation nowcasting is not new, it is challenging and complicated due to the extreme variability of precipitation over a wide range of space time scale. The traditional theory-driven numerical weather prediction (NWP) methods confront numerous obstacles, including an insufficient representation of physical processes, notably due to large scale truncations, and the need of substantial computing resources. On the other hand, data-driven deep learning models attempt to establish a relationship between input and output data to predict future precipitation without regarding the underlying physical processes. In general, they provide a unique output therefore with no information on the uncertainty generated by the nonlinearity of the involved processes. A sort of hybrid approach has been developed using stochastic models physically based on the cascade paradigm, particularly within the framework of universal multifractals (UM), which can be also used to assess the prediction of precipitation nowcasting performance across scales.
AGU FIn this communication, we present a rather systematic study of the combination (VMD-RNN) of variational mode decomposition (VMD) with variants of recurrent neural network (RNN), such as long short-term memory (LSTM), gated recurrent unit (GRU), bidirectional LSTM (BiLSTM) and bidirectional GRU (BiGRU). We underline the role of multifractals to determine the optimal level of the mode decomposition. This is tested on a rather long 5 minutes time series of data in the Paris area (January 1, 2001 to December 31, 2020, i.e. a total of 7305 data). We conclude this communication by discussing several axes of improvement.