Performance Analysis of Significant Wave Height Prediction under Different LSTM Model Input Conditions

In this study, a Long Short-Term Memory (LSTM) model was employed to evaluate the prediction performance of significant wave height at eight coastal stations along the Korean coast, and to analyze how prediction performance varies with input conditions. Input length, input variable configuration, and prediction lead time were defined as key design variables, and model performance was compared under different conditions. The results show that the optimal input length varies by station, while stable performance is generally achieved at 24 hours. In terms of input variables, the use of multiple variables improves prediction performance compared to a single variable, and the best performance is consistently obtained when nine variables are used across all stations. Model performance decreases with increasing lead time. High accuracy with low error is observed for short-term predictions (1~4 hours). Performance remains within RSR ≤ 0.60 and R2 ≥ 0.65 up to approximately 6 hours, whereas noticeable degradation occurs beyond 10 hours. Prediction reliability is significantly reduced at the 24-hour forecast horizon. These results indicate that input conditions have a strong influence on model performance in LSTM-based significant wave height prediction and confirm that the model is well suited for short-term forecasting.