Exploring LSTM neural networks for sea level forecasting based on two-decadal tidal and meteorological data

Abstract The analysis of long-term sea level signal fluctuations and their systematic forecasting showcase the value and implications not only for oceanographic, climate monitoring and related studies but also, on socio-economic aspects of life. Such changes affect significantly the stability in natural processes functioning as well as the usage of nearshore infrastructures; and therefore, interrelate to human activities in coastal areas. This study contributes to the forecasting of sea level variation employing emerging Deep Learning techniques in the form of Recurrent Neural Networks, particularly Long Short-Term Memory (LSTM). Several combinations of data were tested exhaustively to identify the most suitable setup and examine the accuracy depending on available datasets. Univariate, multivariate and multivariate to univariate LSTMs were formed and hyperparameter tuning took place exhibiting RMSE values around 4–5 cm. Also, the effect of seasonality and trend to LSTM forecast was investigated thoroughly leading to performance improvement after eliminating it from the network. The combination of the above networks introduced the “fill-in and predict” technique leading to RMSE slightly over 1 cm for the multivariate to univariate LSTM network. The forecast results were validated against more than two years of historical data.