Convection Forecast Enhanced by the Deep Learning of Radar Observation and Numerical Prediction

It is quite challenging to accurately forecast heavy rainfall induced mainly by the rapid growth of a convective system. Existing methods for convection forecast, such as numerical weather prediction (NWP) and radar echo extrapolation based methods, usually fail to represent the complex nonlinear dynamical and thermo-dynamical processes of convective systems, due to linear assumption (radar echo extrapolation) and physical parameterization based on statistics of large scale processes (NWP).

In this presentation, we introduce new technique developed in Shanghai Central Meteorological Observatory (SCMO)/Shanghai Meteorological Service (SMS) for 0-72h forecast of convection and rainfall, based on the deep learning of datasets from radar echo and numerical prediction. Specifically, for the 0-2 hours forecast, we build an trainable model based on Covolutional LSTM (ConvLSTM) for the nowcasting problem using radar echo dataset. In addition, we utilize the Group Normalization algorithm to refine the convergence performance of the optimization. Experiments show our model outperforms the traditional ConvLSTM and the extrapolation method COTREC. It is also expected that applying radar observation in conjunction with a physical understanding of the atmosphere by numerical model can improve prediction skill for convection in a longer range than that of COTREC. To make the 2-6 hours forecast, the results of radar extrapolation and the numerical model are combined to get an optimized forecast by relieving the spin-up problem of numerical model and the lack of nonlinear-mapping ability to extract inherent features of convection from radar image. Intensity and position correction are performed, which is then followed by the weight coefficient adjustment. For the 6-72 hours forecast, we use the machine learning method to better understand the forecast error of numerical model and make the corrected forecast by eliminating this error.

This technique is verified with numerical experiments for a series of convective cases occurred in Shanghai, which shows that machine learning on the datasets from radar and numerical model can enhance the convection and rainfall prediction compared to that without the use of deep learning.