目的 雷达回波外推是进行短临降水预测的一种重要方法，相较于传统的数值天气预报方法能够实现更快、更准确的预测。基于卷积长短期记忆网络（convolutional long short-term memory network，ConvLSTM）的回波外推算法的效果优于其他的深度学习外推算法，但是忽略了普通卷积运算在面对局部变化特征时的局限性，并且在外推过程中将损失函数简单定义为均方误差（mean squared error，MSE），忽略了外推图像与原始图像的分布相似性，容易导致信息丢失。为解决以上不足，提出了一种基于对抗型光流长短期记忆网络（deep convolutional generative adversarial flow based long short-term memory network，DCF-LSTM）的回波外推算法。方法 首先，采用光流追踪局部特征的方式改进ConvLSTM，突破了一般卷积核面对局部变化特征的限制。然后，以光流长短期记忆网络（flow based long short-term memory network， FLSTM）作为基本模块构建外推模型。最后，引入对抗网络，与外推模型组成端到端的博弈系统DCF-LSTM，两者交替训练实现外推图像分布向原图像分布的拟合。结果 在4种不同的反射率强度下进行了消融研究，并与3种主流的气象业务算法进行了对比。实验结果表明，DCF-LSTM在所有评价指标中表现最优，尤其在反射率为35 dBZ的条件下。结论 由实验结果可知，引入光流法能够使模型具有更好的抗畸变性，引入深度卷积生成对抗网络（deep convolutional generative adversarial network，DCGAN）判别模块能进一步增加结果的准确性。本文提出的DCF-LSTM回波外推算法相比于其他算法，雷达外推准确率获得了进一步提升。

Objective Radar echo extrapolation is an important method for short-term precipitation prediction. It can achieve faster and more accurate predictions compared with traditional methods, such as numerical weather forecast and optical flow method. Among them, numerical weather forecasting requires complex and meticulous simulations of physical equations in the atmosphere and then uses observation data as input to predict future weather conditions. The optical flow method is currently the mainstream method used by the meteorological department, but it has two inherent flaws. On the one hand, only two adjacent frames can be used to estimate the optical flow; on the other hand, the radar echo sequence cannot be fully used for prediction. Nevertheless, the radar echo extrapolation method based on deep learning can take full advantage of spatiotemporal sequence data to achieve faster and more accurate prediction. In addition, the echo extrapolation algorithm based on convolutional long short-term memory network (ConvLSTM) has been proved to be effective in real applications, and the effect is superior to other deep learning extrapolation algorithms. However, it ignores the limitations of ordinary convolution operations in the face of locally changing features, and in the extrapolation process, the loss function is simply defined as mean square error (MSE), ignoring the distribution similarity between the extrapolated image and the original image, which is easy to cause information loss. To solve the above problems, an improved echo extrapolation algorithm based on adversarial long short-term memory network (LSTM) is proposed. Method First, in view of the local-invariance limitations of the traditional convolution kernel, we borrowed the idea of the dense optical flow method and constructed a two-dimensional instantaneous velocity field for all pixels to extract the motion information of each part of the object. Based on this idea, ConvLSTM is improved to form flow long short-term memory network (FLSTM), which is an optical flow optimization extrapolation algorithm. The algorithm uses optical flow to track local features, breaking through the limitation of local invariance of general convolution kernels. Then, according to the characteristics of radar sequence data (high-dimensional spatiotemporal data), the convolutional layer is used to extract effective spatial features to reduce spatial redundancy in the encoder, and then deconvolution is used in the decoder to amplify the generated decoded features to the size of the original image to form an output sequence. The convolutional layer and FLSTM are cross-stacked in depth to encode the input spatiotemporal sequence data into a fixed-length vector. The deconvolution and FLSTM are cross-stacked to decode the output sequence from the encoded vector. Finally, in order to obtain extrapolated images with higher accuracy, an adversarial generation network is introduced, and an extrapolation model forms an end-to-end game system deep convolutional generative adversarial flow-based long short-term memory network (DCF-LSTM). In this system, the generation network is the extrapolation model that tends to be stable after pre-training. Then, the pre-trained generation network continue to be alternately trained with the discriminator to further fit the extrapolated image distribution to the real image distribution, thereby improving the accuracy of the extrapolated image. Result Experiments were carried out under four different reflectance intensities. The DCF-LSTM model is compared with the flow based ConvLSTM (FLSTM) and DC-LSTM, which is an optimized convolutional LSTM by integrating deep convolutional generative adversarial network (DCGAN), and three mainstream meteorological business algorithms. The experimental results show that DCF-LSTM had the best performance under all intensity thresholds. Its probability of detection (POD) and critical success index (CSI) are higher than the other two methods, and it has the lowest false alarm rate (FAR) and mean square error (MSE), especially when the reflectivity is 35 dBZ. The higher the value of POD and CSI, the better the model performance; the lower the FAR value, the more accurate the model. Compared with FLSTM, DCF-LSTM has a 0.012 higher POD, 0.02 lower FAR, 0.015 higher CSI, and 0.115 lower MSE. Compared with DC-LSTM, DCF-LSTM has 0.035 higher POD, 0.03 lower FAR, 0.034 higher CSI, and 0.274 lower MSE. In addition, compared with TrajGRU, ConvLSTM, and Flow methods, DCF-LSTM has a 0.018, 0.047, and 0.099 higher POD; 0.015, 0.036, and 0.083 higher CSI; and 0.012, 0.034, and 0.087 lower FAR, respectively. Conclusion The experimental results show that the optical flow method can enable the model to learn the dynamic changes of local features in the radar sequence, breaking through the limitation of local invariance of the convolution operation and making the model more resistant to distortion. In addition, the introduction of DCGAN module for further game training prediction model can further increase the accuracy of the results. Compared with the three mainstream meteorological business algorithms, the DCF-LSTM echo extrapolation algorithm proposed in this study has further improved the accuracy of radar extrapolation.