Research On Key Techniques Of Precipitation Nowcasting And Meteorological Data Estimation Based On Deep Learning

For the deep learning technology in meteorology,this dissertation selects two key technologies for precipitation nowcasting and meteorological data estimation for research.Accurate precipitation nowcasting with high temporal and spatial resolution provides convenience for daily activities such as transportation and agricultural irrigation.It provides a reference for early warning of natural disasters such as floods and mudslides.However,due to the Spin-Up period in the current mainstream precipitation forecasting business system,it is impossible to provide accurate precipitation forecasts in the short temporary period.The precipitation nowcasting based on deep learning has become the key technology to fill the short temporary period forecast gap.Most meteorological data have multimodal characteristics.For a meteorological element,there are data obtained through different angles or observation principles.Combining these data can provide a more comprehensive understanding of the development process of the meteorological element.The current weather forecast models based on deep learning cannot fully use the multimodal characteristics of meteorological data.In response to this issue,this dissertation combines multimodal fusion and multi-task learning methods to improve the effect of precipitation nowcasting.The lack of meteorological data due to missing equipment deployment or equipment failure often occurs,which leads to the inability of daily business and deep learning technology applications.Among them,the lack of radar echo data has a huge impact.Radar echo data estimation based on other meteorological data has become a key technology to complement the missing.There is little research on radar echo data estimation based on deep learning.According to the characteristics of the radar echo data estimation,this dissertation introduces the attention mechanism and the self-attention mechanism to improve the estimation effect.The main results of this dissertation on the above two issues are summarized as follows:(1)For the precipitation nowcasting,this dissertation firstly combines multimodal fusion and spatiotemporal prediction,and proposes a new precipitation amount nowcasting model based on late fusion.This model fuses the spatiotemporal feature information flow of precipitation amount grid data,radar echo data,and reanalysis data from the spatial scale,and provides abundant meteorological spatiotemporal feature support for the 0-4hour precipitation amount nowcasting.The model uses Traj GRU as the RNN unit,and offers additional guidance for nowcasting through the internal connection of the generated optical flow field guidance network.Based on the research of precipitation nowcasting based on multimodal fusion,a multi-task learning strategy is adopted.A new multi-task learning precipitation nowcasting model based on spatiotemporal scale fusion is further proposed,which simultaneously realizes precipitation amount and precipitation intensity nowcasting.The model further combines multimodal fusion and spatiotemporal prediction to fuse the spatiotemporal feature information flow of the three types of data from the spatiotemporal scale.The dual-input dual-output MFSP-LSTM unit in the model retains the spatiotemporal information flow of the input data while fusing them at the hidden state level.The global spatiotemporal receptive field of the MFSP-LSTM unit is expanded by introducing the SAM unit.Multi-task learning strategies are used in training,and the three forecasts are learned in parallel,and the forecast results promote each other.Based on the natural precipitation nowcasting dataset in the southeastern coastal areas of China,depth experimental evaluation and analysis are carried out.The results showed that the precipitation amount nowcasting model based on late fusion could provide effective nowcasting,and the nowcasting effect exceeds the current operational system.The multi-task learning precipitation nowcasting model based on spatiotemporal scale fusion comprehensively improves the effect of the precipitation amount nowcasting,and has obvious advantages in the heavy precipitation forecast of the precipitation intensity nowcasting.(2)For the radar echo data estimation,a new autoencoder model is proposed based on attention mechanism and self-attention mechanism.The model uses the late fusion Attention-UNet as the primary model framework to integrate meteorological satellite data and numerical weather prediction data from the spatial feature level,and uses the attention mechanism at the high layer of the network to achieve the skip connection.This model introduces the visual Transformer module in the lower layer of the network,extracts the spatial long-distance dependence information in the lower layer of the network,and provides global feature support for generating data details.For the characteristics of radar echo data,a new loss function is designed to strengthen the model’s estimation of highvalue echo.Experimental evaluation and analysis are carried out based on the real radar echo data estimation dataset in northwestern France.The experimental results show that the model can improve the effect of radar echo data estimation.