| Precipitation forecasting,especially precipitation nowcasting,has been an important and difficult research subject in the field of meteorology.Although the traditional precipitation nowcasting methods have achieved certain results,there are problems such as complicated models and error accumulation,which lead to limited prediction accuracy and timeliness.In recent years,the emergence of deep learning has provided a new way of solving the problem of precipitation nowcasting.By learning and analyzing a large amount of historical meteorological data,deep networks can fully capture the relevant features and make accurate precipitation forecasts.However,the existing methods based on deep learning mainly focus on rainy regions with abundant rainfall samples,while the annual rainfall in Lanzhou City is lower than the national average,with fewer rainfall events and fewer corresponding rainfall samples.Therefore,this thesis considers combining the ensemble algorithm for few samples and the deep network for capturing nonlinear trends to construct an end-to-end precipitation nowcasting model for areas with few rainfall.The model consists of a multi-scale spatio-temporal feature extraction network for processing radar data,a deep regression forest for performing ensemble learning,and a coding network for obtaining weights.Finally,the model proposed in this thesis is applied to the precipitation nowcasting in Lanzhou city,and compared with the results of the European Centre for Medium-Range Weather Forecasts(ECMWF)and other comparative models.The experiments show that the model proposed in this thesis is more accurate in predicting short-term precipitation and is more robust in both time and space dimensions. |
