Research On Precipitation Nowcasting Based On Deep Learning

Timely and accurate prediction of rainfall area and intensity in advance is of great significance for the economic development of various industries and the convenience of people's daily life.The conventional precipitation nowcasting method is mainly based on weather dynamic modeling.With the rise of machine learning method,considering the clear mapping between radar echo intensity and precipitation intensity,it has become a research hotspot to extrapolate radar echo image based on optical flow method and various deep learning methods,so as to realize precipitation nowcasting,and shows great potential in prediction accuracy.With the support of National key R&D Program of China,the changes in radar echo image sequences are predicted based on deep learning method for precipitation forecasting.For the precipitation forecasting task,the meteorological dataset of the Pearl River Delta region is firstly collected and constructed,including radar observation echo image and numerical weather prediction products.At the same time,the related process of establishing dataset is introduced,including image preprocessing,data selection and normalization,dataset partitioning and data augmentation.Finally,precipitation forecasting method based on deep learning is proposed,including the deep learning precipitation nowcasting method based on radar echo map and the deep learning precipitation nowcasting method based on fusion of radar echo map and numerical weather prediction data.For the deep learning precipitation nowcasting method based on radar echo map,a precipitation nowcasting model which incorporates an improved spatiotemporal recurrent neural network and a residual Unet network with an attention mechanism is proposed.In the improved spatiotemporal recurrent neural network,the end-to-end encoder-decoder network is utilized to extract spatiotemporal feature information at different spatial scales.In order to address the problem that convolution operation is difficult to capture the growth and decay process and rotation transformation of weather due to its local invariance,a sub network is constructed to actively learn the local neighborhood set information at different locations.Then,an improved Unet module is integrated into the output of spatiotemporal recurrent neural network to address the problem of distortion in radar echo prediction map.In addition,the weighted loss function is designed to improve the performance on the unbalanced precipitation dataset.The experimental results show that the spatiotemporal recurrent neural network incorporating the residual Unet network with the attention mechanism can greatly improve the prediction accuracy and image authenticity.In order to further improve the accuracy of precipitation nowcasting,the method of fusing radar echo map and numerical weather prediction data for precipitation prediction is investigated.Firstly,a multi-source data fusion prediction network model based on spatiotemporal convolutional attention is proposed,which uses spatial attention mechanism and channel attention mechanism to fuse the spatiotemporal feature of numerical weather prediction data and radar echo map.The experimental results show that the model with multi-source data fusion has higher prediction accuracy.Then,in order to improve the prediction computation efficiency of the multi-source data fusion model,a convolutional transformer based fusion precipitation nowcasting model is proposed,which utilizes spatial position encoding and multi-head convolutional self-attention to realize the extraction and fusion of spatiotemporal feature from radar echo map and numerical weather prediction data.The experimental results show that the method can significantly improve the efficiency of prediction computation while ensuring the prediction accuracy.