부산대학교 도서관

Because short-duration heavy rainfall (SDHR) occurs suddenly and locally, district-level warning of SDHR is very important for activities such as production planning, living arrangements and disaster prevention. In order to improve the accuracy, deep learning method is used to transform the district-level warning problem into a multi-label image classification problem. The self-attention convolutional gated recurrent unit (SA-ConvGRU) model with a self-attention memory (SAM) is proposed to implement the district-level warning of SDHR. We use the composite reflectivity data from the Beijing-Tianjin-Hebei radar network from June to August 2016 to 2021 as the input, and whether there is SDHR in each district of Beijing as the output. Accordingly, 80% of each month is used for training and the rest is used for testing. To explore the warning potential of the model, false alarm rate (FAR), probability of detection (POD), critical success index (CSI), etc. are selected as the performance indicators for comprehensive evaluation of the SA-ConvGRU model, and compared with the convolutional gated recurrent unit (ConvGRU) model. Experiments show that the proposed method has strong ability in district-level warning of SDHR, which is better than the compared model, and its FAR is decreased by at least 10% compared with ConvGRU model. Both POD and CSI are better than ConvGRU model, and it has certain application value for the warning of SDHR.