DEVELOPMENT OF A HIGH-ACCURACY WAVE FORECASTING SYSTEM BY INTEGRATING LSTM AND SWAN

Abstract

 Using the third-generation wave model SWAN, wave predictions were conducted for high wave events caused by winter monsoon. The accuracy of predictions was validated against observed data at the Tanabe-Nakashima Storm Surge Observation Tower, located at the mouth of Tanabe Bay in southwestern Wakayama Prefecture. The model residuals, defined as the differences between the predicted and observed values, were used to train a Long Short-Term Memory (LSTM) network. This LSTM network was then employed to directly correct the output of SWAN. The study also investigated the effect of lead time on the accuracy by varying the lead time to 2 hours, 4 hours, and 6 hours. The results showed that the SWAN output could be effectively corrected, particularly with shorter lead times. The combination of physical models and neural networks demonstrated the potential for improving wave prediction accuracy, highlighting the usefulness of the developed wave prediction system.