Πρόβλεψη περιβαλλοντικών χρονοσειρών με ακραίες τιμές με την χρήση μεθόδων βαθιάς μάθησης και κλασικών στατιστικών μοντέλων

Μαρίνος, Αντώνης

Master Thesis

Author

Μαρίνος, Αντώνης

Date

2025

Abstract

This thesis focuses on the forecasting of environmental time series using both traditional statistical models and modern deep learning techniques. The objective is the comparative evaluation of these methods in terms of their ability to capture the fundamental characteristics of time series (trend, seasonality, variability) as well as to predict extreme events, which are of critical importance for environmental applications. To this end, three different datasets were employed: maximum monthly temperatures in Hong Kong, monthly seismic intensities in California, and monthly rainfall in Sydney. Within the scope of this study, traditional models (ARIMA, SARIMA), neural networks (CNN, GRU), and more specialized approaches (DAN, Reweight-EVT, Reweight-META) were applied. The performance of the models was assessed using error metrics (RMSE, MAE) as well as through qualitative analysis of the predictions. The results showed that linear models perform adequately in series with strong seasonality but fall short in datasets with high variability and extreme values. Neural networks proved superior in capturing non-linear patterns, with CNNs providing better accuracy in peak values and GRUs offering more stable forecasts. The specialized models further improved performance, with Reweight-META demonstrating the most adaptive behavior in predicting extreme events. Overall, the study highlights that the integration of deep learning techniques, and particularly reweighting-based methods, can enhance the accuracy and practical utility of environmental time series forecasting models, offering valuable tools for the study and management of rare but critical phenomena.

Postgraduate Studies Programme

Πληροφοριακά Συστήματα και Υπηρεσίες

Department

Σχολή Τεχνολογιών Πληροφορικής και Επικοινωνιών. Τμήμα Ψηφιακών Συστημάτων

Number of pages

147

Language

Greek

URI

https://dione.lib.unipi.gr/xmlui/handle/unipi/18726

Collections

Τμήμα Ψηφιακών Συστημάτων

Show full item record

Except where otherwise noted, this item's license is described as
Αναφορά Δημιουργού-Μη Εμπορική Χρήση-Όχι Παράγωγα Έργα 3.0 Ελλάδα