Optimizing Coastal Wave Forecasting with LSTM: A Univariate Time Series Approach

Coastal wave modeling and forecasting are vital in various fields, such as oceanography, sustainable marine energy, maritime operations, and ocean engineering. This research focuses on accurate wave parameter forecasting, which is crucial for marine-related activities. The primary objective of this experiment is to forecast wave conditions by introducing a time series forecasting model capable of forecasting wave parameters using a hybrid method that combines advanced, recurrent neural networks with hyperparameter tuning and optimization techniques. We proposed a novel univariate model that combines the Long Short-Term Memory architecture with KerasTuner hyperparameter tuning and optimization techniques to enhance time-series forecasting accuracy and efficiency. Seven independent multi-step forecast experiments were conducted for wave speed and propagation direction for 1, 3, 6, 8, 10, 12, and 24 hours into the future. The proposed model demonstrated the best results with minimal training and test errors. Comparative evaluations were performed using benchmark datasets to validate the model's superior performance over conventional LSTM, ARIMA, and Prophet Models. The proposed univariate time series forecasting model for wave speed and propagation direction exhibits significant value due to its accurate forecasts, low forecasting error, and potential implications in marine renewable energy assessment.