Application of advanced deterministic, physically based as well as artificial intelligence mathematical models for the study of natural disasters (floods, fires) in Greece
Loading...
Date
2026-06-30
Authors
Μουσούλης, Ηλίας
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
This dissertation develops and applies an integrated modelling framework for analysing and predicting hydrological response under natural hazard conditions in Greece, with emphasis on floods, wildfires, sediment transport, and hydropower-related water management. The work combines deterministic, physically based modelling with Artificial Intelligence and machine learning approaches in order to examine how different modelling philosophies can support catchment-scale understanding, prediction, and decision-making under complex environmental conditions. The physically based component focuses on two representative applications. In East Attica, the distributed MIKE SHE model is used to investigate the hydrological effects of the August 2009 wildfire across multiple semi-arid Mediterranean catchments. The modelling results show that wildfire-induced land-cover change, reduced interception, altered soil infiltration, and increased surface runoff can substantially amplify post-fire hydrological response, particularly during the first years after disturbance. In the Upper Acheloos catchment, MIKE SHE, MIKE 11 rainfall-runoff, hydrodynamic, and sediment transport components are combined to simulate discharge, reservoir inflows, and sediment dynamics in a regulated mountainous river system. These applications demonstrate the value of process-based models for interpreting catchment behaviour, assessing land-use and hydraulic regulation effects, and supporting water and sediment management. The data-driven component develops a reproducible benchmarking workflow for hydrological time-series modelling across catchments in Central and Northern Greece. A structured scenario framework is introduced to evaluate alternative feature representations, including raw gauge inputs, dimensionality reduction, spatial aggregation through inverse distance weighting and kriging, lagged and rolling variables, and target-specific predictors such as potential and actual evapotranspiration, reservoir storage, and water level. Multiple AI/machine learning model families are compared, including multilayer perceptrons, support vector regression, random forest, LSTM, GRU, CNN1D, XGBoost, and LightGBM. Model performance is assessed using Nash-Sutcliffe efficiency, Kling-Gupta efficiency, root mean squared error, and a stability indicator designed to identify overfitting and support robust model selection. The results indicate that model performance depends strongly on temporal scale, feature design, and target variable. Daily discharge prediction benefits from explicit temporal-memory features and models capable of representing short-term dynamics, whereas monthly discharge and hydropower production are generally more stable and can be predicted satisfactorily using aggregated meteorological and reservoir-state information. Across applications, the best-performing model is not necessarily the most complex; rather, robust selection requires balancing predictive accuracy with generalization capacity and stability across training and testing periods. Overall, the dissertation contributes a transferable methodological framework that links data collection, preprocessing, feature engineering, physically based simulation, machine learning benchmarking, and multi-metric evaluation. By treating deterministic models and AI-based methods as complementary rather than competing tools, the thesis provides a practical basis for hydrological prediction, post-wildfire runoff assessment, sediment transport analysis, and hydropower forecasting in data-variable and hazard-prone catchments.
Description
Keywords
Hydrology, Hydrological models, Mathematical models, Artificial intelligence, Artificial neural networks, Natural disasters, Wildfires, Floods, Timeseries models