Edinburgh Napier University

  Precipitation forecasting has long been a prominent topic in meteorology, as accurate predictions of impending rainfall are crucial for daily life and travel planning. Currently, radar echo extrapolation serves as the primary method for precipitation proximity forecasting. The precision of extrapolated radar echo maps directly impacts the ability of weather forecasters to successfully anticipate severe weather events. Deep learning techniques have recently gained widespread adoption in this domain, enabling the identification of precipitation events with varying intensities. However, existing approaches primarily focus on capturing long-term motion dynamics through recurrent neural networks. As a result, these methods struggle to accurately capture the trends in radar echo motion as the prediction time horizon increases, leading to issues such as feature disappearance and blurring in the predicted radar echo maps. To address this challenge, this paper proposes a novel recurrent unit called RM-ConvLSTM. This unit incorporates a recurrent memory perception module and an additional memory component to enhance the model's performance in long-term radar echo map predictions. The real-world CIKM 2017 radar echo dataset is utilized to evaluate the proposed method and to conduct comparative experiments with representative models from previous studies. The results demonstrate that the proposed model outperforms existing methods in predicting radar image quality.