Short-Term Tidal Forecasting Based on a TPE-Optimized Hybrid BiGRU–ResNet Framework

Tidal level prediction is significant in maritime navigation, coastal development planning, and the anticipation of coastal hazards. This research proposes a hybrid deep learning approach that incorporates a Bidirectional Gated Recurrent Unit (BiGRU) network and a Residual Neural Network (ResNet) for short-term tidal level estimation. The BiGRU component captures temporal dependencies from forward and backward directions, while the ResNet module performs deep feature learning based on residual learning to capture sequential and spatial-temporal complicated patterns in tidal signals. Hyperparameters of both BiGRU and ResNet components were tuned by the Tree-structured Parzen Estimator (TPE) through optimizing hyperparameters of the model by probabilistic search procedure. Tidal level data collected from Ras Tanura during the time period from 2012 to 2021 were used in developing and testing the proposed hybrid BiGRU–ResNet framework. Forecasting performance was tested at short horizons of 5, 10, 15, and 20 days ahead. The proposed model outperformed the BiLSTM, BiGRU, and CNN models across all forecasting horizons, achieving lower prediction errors at each test interval. The outcomes indicate that the proposed hybrid model produces satisfactory short-term tidal level forecasts for coastal development and maritime activity.