An improved tidal prediction method using meteorological parameters and historical residual water levels

Abstract

Accurate water levels information is essential for ocean and coastal management. Water levels are categorized into astronomical tides and residual water levels. The temporal and spatial continuity of residual water levels is influenced by a diverse set of factors. These complex influences pose challenges to achieving precise water level predictions. The TidalMet-HR method is proposed. By feeding wind speed, wind direction, atmospheric pressure, and historical residual water level data into a bidirectional long short-term memory network (Bi-LSTM), we determine the optimal network structure using Bayesian hyper parameters. Subsequently, we integrate astronomical tides with the predicted residual water levels to produce highly accurate future water level forecasts. To evaluate the performance of TidalMet-HR, the water level data of four tidal stations were analyzed. The results indicate that the TidalMet-HR method achieves water level prediction accuracy values within 5 cm and maintains limit errors below 10 cm for the next 36 h. Comparatively, the water level prediction errors at the four stations are reduced by 24.0% (compared to LSTM) when utilizing other artificial intelligence-based forecasting methods (e.g., BP, an RBF, and LSTM). These results certify the TidalMet-HR method as a significant benchmark for achieving precise tidal forecasting.