Evaluation of the ConvGRU Deep Learning Method for Convective Weather Nowcasting

Nowcasting and early warning of severe convective weather play a crucial role in prevention of meteorological disasters. Although there has been much progress over the past several decades, nowcasting remains challenging. This study evaluates the ConvGRU deep learning method for convective storm nowcasting using long-term operational radar data. First, the nowcasting problem is transformed into a sequence prediction problem. Then, we use the modified recurrent neural network named ConvGRU to build an encoder-decoder model which will be fed long-term radar data to complete the training process. The radar reflectivity data are a mosaic of data from six operational radars over the Beijing–Tianjin–Hebei region during the period of 2010-2016. The trained neural network is then used to make nowcasts in next one hour by every six minutes for the testing dataset (i.e. radar data in 2017). The experimental results show that, compared with the traditional extrapolation forecast method CTREC, the proposed deep learning method improves nowcasting accuracy.