Research On Short-term Precipitation Prediction Based On Spatiotemporal Prediction Network And Convolutional Neural Networ

The precipitation caused by severe convection weather is one of the inducement to natural disasters,which is closely related to people’s life.In recent years,such increasingly frequent disasters have posed a serious threat to the life and property security.Therefore,nowcasting has attracted extensive attention worldwide.With the development of the meteorological monitoring equipment,a high focus has been placed on the radar-based nowcasting technology,whereas the precision and time-effect could not meet the business requirements.In this dissertation,an improved quality control method of radar and two deep-learning-based nowcasting algorithms have been proposed in order to promote the precision of nowcasting.To sum up,the main contributions of this dissertation are summarized as follows:(1)Exploiting a more effective method of radar quality control.In this dissertation,the composite reflectivity is computed to alleviate the severe beam blockage that the radar experiences by the nearby mountains at lower elevation scans.And then,the k-nearest neighbor algorithm is applied to raster the composite reflectivity to a resolution of 0.005°.To mitigate the noise,the global average value of the reflectivity is utilized to identify the background noise and filter the systematic noise signals with the revised methods.In addition,a nowcasting dataset covering 2017-2019 is made to verify our proposed algorithms.(2)Proposing an asynchronous spatiotemporal encoder network for radar extrapolation.Taking differences in spatial and temporal characteristics into consideration,two special encoders are employed to capture the distinct spatial and temporal features respectively.Specifically,the temporal encoder is based on the deformable convolution and Convolutional GRU to capture the trajectory variation between time steps and the spatial encoder is to acquire the distribution variation of the radar reflectivity among the sequences with multilevel receptive fields.And then,such feature maps are integrated through a cross fusion module with the residual connection.The ultimate predictions would be generated with Pixel Shuffle iteratively.Extensive experiments have testified the effectiveness and superiority of the proposed algorithm.(3)Proposing a deep spatial-temporal fusion network for precipitation estimation.As the poor utilization in spatiotemporal characteristics from the Z-R relationship,the neighbor rectified unit is designed to improve the current feature extraction with the previous feature.Through the stacked units,the comprehensive feature maps with ample spatiotemporal information would be acquired.Subsequently,the precipitation would be regressed with the global average pool and 1 × 1 convolution.Extensive experiments illustrate that the proposed model achieve a better performance than the existing precipitation estimation methods.