National Technical University of Athens (NTUA)
Ανάπτυξη υβριδικών μεθοδολογιών για την ανάλυση και πρόγνωση της στάθμης της θάλασσας από παλιρροϊκά δεδομένα με εφαρμογή στον ελλαδικό χώρο
Abstract
dc:descriptionMore than 70% of the Earth's surface is covered by seawater, while the remaining part comprises the continental landmasses. The sea, as a fluid, is characterized by highly sensitive kinetic behavior due to complex effects resulting from the gravitational attraction of celestial bodies, their movements, as well as a series of hydrodynamic and atmospheric influences and fluctuations. These factors result in the horizontal movement of water mass and the vertical variation of its surface. This dissertation discusses sea level variation as a result of all these influences, the development of a deterministic model for estimating this variation as a function of time, and a stochastic approach for forecasting its future values. The analysis and processing were conducted using real data from two tide gauges in Greece, located at Piraeus and Thessaloniki, along with data from the nearest meteorological stations for the period 1999–2019. The first axis of the dissertation interprets the theoretical background governing the phenomenon of sea level variation analyzing and modeling its main components: the astronomical tide, the meteorological tide (storm surges), and the residual signal of other influences. The astronomical tide, resulting from the Earth-Moon-Sun system's influence, is determined by individual tidal constituents that depend on the characteristics of each specific region. Their calculation is derived from a harmonic analysis of the sea level signal obtained from the tide gauges. The synthesis of these tidal constituents constitutes a function of time for its calculation. The meteorological tide is caused by atmospheric (meteorological) factors and is derived through the adoption of a two-dimensional barotropic model, primarily influenced by wind action and changes in atmospheric pressure. Calculating the storm surge effect on an hourly basis, the time series of the meteorological tide for the period under study is formed. The residual signal, obtained by subtracting the astronomical and meteorological tides from the total sea level, is transferred from the time domain to the frequency domain to be harmonically analyzed into its constituent components. Thus, a local function is created to calculate the residual signal for each time step. During this processing, characteristic parameters of each region (gulf depth, coastal azimuth, gravitational acceleration, water and air density) were used. The analytical sea level model is evaluated through statistical indices and achieves a calculation accuracy for sea level fluctuation in the range of 7-8 cm. Also, additional products characterizing each location were derived (storm surge index, skew surges, extremes analysis, etc.). The second pillar of the dissertation utilizes the generated time series from the analytical model and employs Deep Learning methodologies to formulate a special type of Recurrent Neural Networks (RNNs) known as Long Short-Term Memory (LSTM), with the aim of forecasting future sea level values. These networks are configured using various techniques, testing different input data to achieve the most accurate prediction of sea level at the examined locations. This includes forecasting over different time horizons with a one-hour temporal resolution and multiple combinations of network parameterization. Consequently, univariate, multivariate, and multivariate-to-univariate LSTM networks are structured. In the first two cases, the target variable (sea level) resides in both the input and output variables, while in the third case, it is included only in the output. Concerning forecasting techniques, both static (closed loop) and dynamic (open loop) methods have been implemented. Sea level forecasting, through all the aforementioned tests, extends up to four weeks of hourly values, with various combinations of input data, many of which lead to successful results. Furthermore, an effort is made to select hyperparameters by interpreting their physical significance. The results demonstrate that, on one hand, the sea level signal is better predicted byt its components indicated by the analytical approach and, on the other hand, that besides the astronomical tide, a dominant role is played by the meteorological tide, which is primarily determined by atmospheric pressure and the cross-shore component of wind speed. Finally, the accuracy of the sea level forecast is up to 0.06 m (multivariate technique), while enabling the multivariate to univariate technique can reach the impressive value of RMSE of 0.01 m for hourly forecasts of three weeks ahead.
Degree
thesis:*- Grantor dc:publisher
- National Technical University of Athens (NTUA)
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Papadopoulos, Nestoras
- Παπαδόπουλος, Νέστορας
Subjects
dc:subject × 17- Παλιρροιογράφος
- Νευρωνικά δίκτυα
- Αλγόριθμοι πρόγνωσης
- Αστρονομική παλίρροια
- Μετεωρολογική παλίρροια
- Tide gauge
- Long short term memory (LSTM)
- Forecasting
- Astronomical tide
- Meteorological tide
- Storm Surges
- Περιβαλλοντική Μηχανική
- Επιστήμες Μηχανικού και Τεχνολογία
- Μηχανική ωκεανών
- Environmental Engineering
- Engineering and Technology
- Ocean Engineering
Rights
- Language dc:language
- gre
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/62613