Research On Nowcasting Method Based On Convolutional Recurrent Neural Network

Nowcasting forecast mainly refers to the forecast of precipitation in the next 0-2hours,which is of great value to transportation,agriculture,urban disaster prevention and other fields.The key problem of nowcasting is to accurately extrapolate the radar echo image in the future.The traditional method used for radar echo extrapolation is mainly the optical flow method,but the prediction result of the optical flow method has a relatively large deviation because the three assumptions of the method are not valid in the actual radar echo images.At the same time,the traditional method only uses the current radar data,and does not make good use of the large amount of existing historical data.Recent years,with the development of deep learning,scholars have proposed many deep learning algorithms for extrapolation of radar echo.These models are mainly based on the time and space characteristics of the radar echo sequence for prediction,but these models still have some drawbacks,such as the gradual dissipation of large echoes,fuzzy prediction results,and poor prediction of heavy rainfall.This dissertation analyzes the existing radar extrapolation model and points out that it has three problems: the loss function cannot guide the model to generate images with similar styles to the real image;the sequence model's hidden states will 'shift' when they are updated;the model has feature loss when extracting data features.Therefore,this dissertation proposes improvements in three aspects: introducing a multi-target loss function;optimizing the transfer mode of the sequence model state;using skip-connections to improve the model's feature extraction capability of radar data.Comparative experiments show that the above three schemes can improve the predictive ability of existing models.On the basis of the above improvement scheme,this dissertation proposes a radar echo extrapolation model based on hierarchical Conv RNN,which uses a asynchronous update strategy to reduce the distribution offset and information loss of the sequence model's hidden state.The experimental results show that the hierarchical Conv RNN model obtains better prediction results than existing models and requires less training cost.This dissertation further improves the Conv RNN model from the aspects of model parameters,model structure and loss function optimization.The experimental results show that the above improvements can improve the prediction performance of the hierarchical Conv RNN model.Finally,this dissertation also participated in the design and implementation of a nowcasting system based on radar echo extrapolation,used to display and compare the prediction results of various radar echo extrapolation algorithms.