Long short-term memory neural networks for forecasting sea level and seiche occurrences in the Maltese Islands

The ability to predict seiches can help prevent the damage and mitigate the risks associated with these natural phenomena. This paper presents a novel approach for seiche prediction in Marsaxlokk, Malta, using Long Short-Term Memory (LSTM) neural network models. The LSTM models are implemented on time series data obtained from two tide gauge stations installed in Marsaxlokk and Portomaso, Malta. Two recurrent neural networks, the Vanilla recurrent neural network (RNN) models and gated recurrent unit (GRU) models, along with time-lagged multiple linear regression models are also fitted as baseline models. Due to the presence of missing data, this paper also explores the use of gap filling methods to obtain complete datasets for the recurrent neural network model training and testing. Two variants of each recurrent neural network model are presented. These are trained on the Portomaso dataset and tested on the Marsaxlokk dataset with the second version of each model involving calibration. The significantly superior calibrated LSTM model was able to identify with good precision the seiches that occurred on November 28, 2021 and June 30, 2022 and to also model the long-term dependencies that are important for predicting sea levels with enough lead time. Conclusion: The calibrated LSTM model gave the overall best performance out of the tested RNN and time-lagged multiple regression models and was able to identify the seiches that occurred on November 28, 2021 and June 30, 2022. • Prediction of seiches with RNN and MLR models is largely unexplored in the literature. • Various models were trained and tested on different stations due to limited data. • To address prediction discrepancy, calibration on training set was further applied. • Calibration yields smaller RMSE, MAE and NSE over uncalibrated approach in all cases. • Calibrated LSTM gave overall superior predictions of sea level and seiche occurrence.