TYPHOON WAVE PREDICTION USING LONG SHORT-TERM MEMORY NETWORKS FOR OFFSHORE WINDFARM ON WESTERN COAST OF TAIWAN

An accurate and efficient typhoon wave prediction model is important for improving the efficiency of offshore wind farm management. In the earlier studies, short-lead-time (i.e., 1 to 3 hours) typhoon wave prediction models were developed for the Taiwan coastal area. These models were constructed by BPNN with local meteorological information. However, Sufficient prediction lead-time is essential for early warning and response to offshore wind farms during typhoon events. Furthermore, past research on typhoon waves along the western coast of Taiwan often presented an underestimated tendency due to the structure of the typhoon being destroyed by the Central Mountain Range. This study aims to establish a novel long-lead-time typhoon wave prediction model using Long Shor-Term Memory (LSTM) networks while carefully considering typhoon parameters. The basic concept of LSTM is to utilize the memory cell to capture the features or vectors of time-related data, significantly enhancing prediction accuracy. The results of LSTM demonstrate high consistency with in-situ data for 1-hour lead time (i.e., the correlation coefficient is up to 0.98). For longer lead time (e.g., 6 hours), the method significantly improves learning and generalizing capability more than shallow learning methods. The correlation coefficients for training and validation reach 0.91 and 0.86, respectively.