| Precipitation nowcasting refers to the prediction of precipitation intensity and distribution in the next 0-2 hours,and accurate precipitation nowcasting is crucial to human life and economic development.Due to the nonlinearity,complexity,and timeliness of precipitation,traditional forecasting methods are difficult to meet the accuracy requirements of meteorological business.With the continuous development of deep learning,using deep learning models for precipitation nowcasting has become a key point in the research of meteorological field.In addition,these models have higher prediction accuracy than traditional meteorological forecasting methods.Although deep learning models have made great progress in precipitation nowcasting,there are still some problems,such as poor generalization performance and insufficient prediction accuracy.In order to further improve the accuracy of precipitation nowcasting,this thesis improves the deep learning methods for precipitation nowcasting based on the temporal and spatial characteristics of radar echo images.The specific contents are as follows:(1)In order to improve the generalization ability of the models on spatial features,many researchers have introduced convolutional neural networks to extract the spatial features of radar image sequences.However,due to the complex spatial features of radar image sequences,convolutional neural network is inherently difficult to describe their complete spatial dependencies,making it impossible to effectively model spatial features.Therefore,this thesis utilizes the spatial correlation of the radar image sequence,and uses statistical theory to design a feature extraction method based on spatial correlation(FESC),which is used to extract the spatial features of the input sequence in the model.The experimental results show that FESC can effectively improve the prediction accuracy of the model and the accuracy of precipitation nowcasting.(2)Currently,many studies on precipitation nowcasting use an Encoder-Decoder structure based on recurrent neural networks.This structure enhances the capabilities of the model in the area of feature extraction and processing to a certain extent,but it is still difficult to capture the global spatial dependencies and trajectory motion features in radar images,resulting in poor performance in precipitation nowcasting of medium to high intensity.In order to further improve the ability of the model to capture spatial trajectory distribution features and learn temporal features,this thesis improves the Encoder-Decoder structure and proposes an end-to-end deep spatio-temporal fusion network(DST-FN).DST-FN includes a spatio-temporal trajectory prediction network and a time information prediction network,which independently extracts and uniformly models the spatial trajectory features and temporal information features.Meanwhile,the feature of input sequence in the spatio-temporal trajectory prediction network is extracted by the FESC method.In addition,this thesis designs a trajectory gated recurrent model with spatial self-attention module(Spatial SelfAttention Module Traj GRU,SAM-Traj GRU)for the recurrent structure in the DST-FN network.Based on the trajectory gated recurrent unit(Traj GRU),SAM-Traj GRU introduces a spatial selfattention mechanism to memorize features with global spatial dependencies.Experimental results demonstrate that,through the collaboration of deep spatio-temporal fusion networks and spatial selfattention mechanisms,DST-FN can accurately predict the distribution characteristics and motion trajectories of radar echoes,and the prediction results outperform existing models on the evaluation metrics for deep learning model and the precipitation forecast. |
