부산대학교 도서관

Accurate and efficient prediction of the typhoon-induced surge plays an important role in coastal disaster reduction. In earlier works, efficient models based on the backward propagation neural network (BPNN) with local meteorology data have been developed for short-lead-time (i.e., 1 to 3 hours) storm surge predictions along the coast of Taiwan. Sufficient prediction lead-time is still needed to help preparedness, early warning, and response work for reducing life and property losses during typhoon events. The purpose of this study is to develop a long-lead-time storm surge prediction model by a novel deep learning method (i.e., long-short-term memory networks, LSTM) with effective typhoon parameters. The basic idea of LSTM is the use of memory functions to capture the relationship between the time series of data, significantly improving the prediction accuracy. Consistent with previous works, the 1-h-ahead surge predictions using LSTM are in excellent agreement with the observation data (e.g., correlation coefficient C.C. up to 0.95). In terms of a longer lead-time (i.e., 6 ahead of surge prediction), the LSTM model presents greatly improved learning and generalizing capability. The performance assessment for the training and validation phases indicates C.C = 0.85 and C.C. = 0.8, respectively.