Prediction Of Radar Echo Sequences For Rainfall Forecasts

Nowcasting precipitation plays a crucial role in people's daily life and production.At present,the methods of precipitation nowcasting based on deep learning have become a popular research topic in the meteorological community.In this thesis,facing the meteorological station's demand for rainfall forecasting,a prediction model of radar echo sequences based on conditional generative adversarial network(CGAN)is proposed and explored by a way of video prediction,and the research results is combined with engineering applications.The main work and contributions of this thesis are summarized as the following.(1)Research on radar echo sequences prediction model based on conditional GANUnder the framework of adversarial learning,a generator based on spatio-temporal LSTM(ST-LSTM)and a discriminator based on 3D convolutional neural network are constructed.Both of them form the model of CGAN.By introducing a joint loss function that combines the adversarial loss with the content loss,the CGAN model can be trained by playing the min-max game.The generator takes the historical sequences of the radar echo images as input,uses the stacked ST-LSTMs to extract features layer-by-layer,and guides the generation of prediction sequences through the parameter control.When the training phase is finished,the generator will be used for radar echo sequence prediction.By adding Gaussian random noise to the input of the discriminator,the system stability can be effectively improved.The sample image sequences are processed gradually by 3D convolution layers and fully connected layers,and therefore the spatial and temporal information can be effectively extracted simultaneously,and finally the probability of authenticity about the input sequence can be predict based on the Sigmoid function.The comparison experiments between the proposed method and the Pred RNN model on artificial synthesis,natural video sequences and radar echo sequences show that the prediction results of the two models are relatively sharp,and the predicted shape,boundary and motion trajectory are basically consistent with the real situation.However,the prediction of texture details in the model of this thesis looks a bit sharper,and the model has stronger generalization ability and better robustness.(2)Application of radar echo sequence prediction model for meteorological stations' for rainfall forecastingTo explore the application value of the proposed model,the prediction experiment of echo sequence is carried out based on the historical data about rainfall and radar echo gathered from a large area coverage of the test scenarios.Therefore,the effectiveness of this model in engineering application is further verified.In order to achieve a large range of rainfall prediction,the large-scale radar image corresponding to the experimental area is divided into overlapping sub-blocks suitable for model input,and each sub-block sequence is fetched into the trained model respectively to generate the corresponding prediction results of the same period;all of the prediction results about the sub-blocks are further combined to obtain the prediction results of the large-scale radar echo sequence corresponding to the experimental area.The experimental results show that even if the cumulative rainfall forecast is for the next one hour,the model can still accurately predict the radar shape,boundary and movement trajectory of the main rainfall area;The prediction of the radar echo intensity is consistent with the actual situation at the initial stage of the forecast.As the forecasting time increases,the accuracy of the detailed prediction of the echo intensity decreases.