Sea level modelling and forecasting combining analytical and deep learning methods, case study: Piraeus, Greece

Abstract This study contributes to sea level modelling and prediction through the development of a framework proposal consisting of both an analytical approach and Deep Learning methodologies. Analyses have focused on the strategically important harbour of Piraeus located near the metropolitan city of Athens, Greece. The research leverages a database comprising tide gauge recordings and meteorological variables, spanning the period from 1999 to 2019. Using these datasets an analytical model was designed. This model adopts a component-based approach, wherein the total sea level estimate is decomposed into its constituent parts. The analytical approach yields a sea level model achieving an accuracy of about 75 mm in terms of Root Mean Square Error (RMSE). Furthermore, recognizing the increasing tolerance of data-driven computational approaches, this study explores AI-based, Deep Learning methodologies through the implementation and evaluation of alternative Long Short-Term Memory (LSTM) recurrent neural network architectures. The proposed LSTM models were specifically designed to predict high resolution (hourly) sea level estimates at scalable forecasting times ranging from 7 to 28 consecutive days. The results obtained proved to be highly promising, demonstrating a strong fit between the sea level hindcast and the corresponding data recordings. Quantitative evaluation of the model performance revealed RMSE values ranging from 13 mm (multivariate to univariate forecast) to 61 mm (multivariate forecast) and 64 mm (univariate forecast) for a 7-day ahead prediction.