光流法及其在临近预报中的应用

短临降水预报对于暴雨和强对流天气监测预警服务具有重要意义，使用雷达回波外推方法进行短临降水预报是目前较为常用的预报方法之一，但是传统的雷达回波外推方法普遍存在数据利用率低、外推准确性差和外推模糊等问题。针对上述问题，利用陕西全省雷达拼接数据资料，选择深度学习中编码器-解码器结构，以卷积长短期记忆网络(Convolutional Long Short-Term Memory Network, ConvLSTM)作为循环单元，构造了基于全局通道注意力的ConvLSTM预测网络(Global Channel Attention based ConvLSTM, GCA-ConvLSTM)；此外，为进一步提高GCA-ConvLSTM预测网络的拟合能力，使用集成学习算法对其进行改进，通过装袋算法对数据集进行采样，训练3个GCA-ConvLSTM预测网络作为基学习器，使用加权投票策略将这3种基学习器进行有效组合，最终获得了一个性能更优的组合模型。试验结果表明，基于集成学习算法改进的GCA-ConvLSTM雷达回波外推方法与现有深度学习方法相比，提升了短临降水预报方法的准确性和时效；该方法在25 dBz、35 dBz和45 dBz反射率阈值下的评估试验中分别比对比的主流深度学习模型CSI值平均高出0.149、0.192、0.085；同时该方法的外推结果拥有更加清晰的边缘和细节性纹理，减轻了外推后期模糊问题。

The optical flow method and its application to nowcasting

Precipitation nowcasting is of great significance for severe convective weather warning. Radar echo extrapolation is a commonly used precipitation nowcasting method. However, the traditional radar echo extrapolation methods are encountered with the dilemma of such as low data utilization and of the extrapolation ambiguity. To solve the above problems, weutilize the radar data of Shaanxi Province and chooses the Encoder-Decoder as the overall structure of the prediction model. Besides, we choose the ConvLSTM (Convolutional Long Short-Term Memory) as the unit of the prediction model, and designs aglobal channel attention mechanism integrated into the model tobuild a prediction network called GCA-ConvLSTM (Global Channel Attention based ConvLSTM). In addition, to further improve the fitting ability of our prediction model, we use the ensemble learning method to achieve this goal. The algorithm first samples the dataset through the Bagging algorithm, and then we use these sampled data to train three GCA-ConvLSTM networks as the base learner. In the end, we obtain a better performance model which effectively combines the three base learners by using a weighted voting strategy. The experimental results show that the improved GCA-ConvLSTM radar echo extrapolation method based on the ensemble learning algorithm improves the accuracy and timeliness of the precipitation nowcasting compared with the existing deep learning methods.In the evaluation experiments under the reflectivity thresholds of 25 dBz, 35 dBz and 45 dBz, this method is 0.149, 0.192, and 0.085 higher than the average CSI values of the mainstream deep learning models.The extrapolation results of this method have clearer edges and detailed textures, which alleviates the blurring problem in the later stage of extrapolation.