An Improved LSTM-based Method Capturing Temporal Correlations and Using Attention Mechanism for Radar Echo Extrapolation

Precipitation nowcasting has always been a hot topic in the field of meteorology, its main task is to forecast the possible future precipitation events of a certain region in the short term, which plays a key role in people's daily lives, such as work arrangements, travel planning. At present, precipitation nowcasting is mainly based on radar echo extrapolation. So, the accuracy of extrapolated radar echo maps is related to whether the weather forecaster can successfully predict the possible severe weather events. Recently, deep neural networks have been widely used in radar echo extrapolation tasks as an attempt to improve the accuracy of extrapolated echo maps. Compared with traditional radar echo extrapolation methods, which are data-driven and have less limitations compared with centroid tracking-based methods and other traditional radar echo extrapolation methods. However, these existing methods only use convolution or gated structure in LSTM to capture and store features, thus failing to fully utilize the information in the maps, which may lead to the underestimation of the echo intensity. To address the problem, this paper proposes a new designed recurrent unit called TA-ConvLSTM, which improves the model's ability to capture the variation features of radar echoes by adding a temporal correlation extraction module and using an attention mechanism. The comparative experiment results show the extrapolation performance of the proposed method, outperforming some representative models in previous work.