Short- and long-term tidal level forecasting: A novel hybrid TCN + LSTM framework

Abstract

Tidal level forecasting is essential for maritime safety, coastal management, and infrastructure planning. This study proposes a hybrid framework combining Temporal Convolutional Network (TCN) and Long Short-Term Memory (LSTM) model to predict tidal levels across short- and long-term horizons. The TCN excels at capturing temporal patterns, while the LSTM effectively models sequential dependencies, facilitating accurate forecasting of tidal fluctuations. Covariance Matrix Adaptation Evolution Strategy (CMA-ES) was used for hyperparameter tuning of both TCN and LSTM component of hybrid framework. Historical tidal data from Ras Tanura (2012−2021) was utilized for training and evaluation. The analysis revealed that the hybrid TCN + LSTM framework optimized via CMA-ES outperformed other deep learning models, including standalone LSTM, GRU, and CNN, demonstrating enhanced accuracy and reliability across various forecasting horizons. For short-term predictions (T + 5 and T + 10 days), it achieved MAE values of 0.073 and 0.081, with MAPE values of 7.43 % and 9.15 %, respectively. For longer-term horizons (T + 30 and T + 60 days), it maintained accuracy with MAE values of 0.050 and 0.054 and corresponding MAPE values of 5.39 % and 4.93 %. The study demonstrates the potential of the hybrid TCN + LSTM framework for reliable tidal level forecasting, supporting better planning and decision-making in coastal and maritime applications.