Πρόβλεψη χρονοσειρών κλιματικών μεταβλητών βασισμένη σε Transformers

Σταυρόπουλος, Γεώργιος

Transformers-based time series prediction of environmental variables

Master Thesis

Author

Σταυρόπουλος, Γεώργιος

Date

2026-01

Abstract

This study focuses on the analysis and prediction of environmental variables by leveraging advanced Artificial Intelligence and Deep Learning techniques. The primary objective is to develop a modern and reliable forecasting model based on the Transformer architecture, capable of predicting the long-term evolution of climate indicators such as temperature and rainfall statistics. The study relies on historical time-series data spanning more than a century, serving as the foundation for understanding environmental trends and climate change. Special emphasis is placed on data collection and preprocessing, as their quality and completeness are critical to training and evaluating an accurate and effective forecasting model. The work includes detailed visualizations of statistical indicators extracted from the time-series data, offering valuable insights into the evolution of climate conditions. By analyzing these datasets, the study first aims to present a clear and tangible picture of Athens’ environmental history over an extended time horizon. The innovative aspect of this work lies in the use of the Transformers architecture, a revolutionary method in sequential data processing, which, until now, has been applied only to a limited extent in problems related to climate time series. The model is developed using an attention mechanism that processes the entire input sequence simultaneously, enabling more efficient learning of long-term dependencies, which is essential for achieving accurate long-range forecasting. While experimenting with the forecasting model, we will implement alternative approaches to the Transformers architecture, as well as variations in the parameterization during model training. Each approach will be tested in forecasting with a specific time horizon, varying the input data each time. Using appropriate accuracy metrics, we highlight the strengths and limitations of each approach and present the algorithm’s best-performing configurations. Finally, by thoroughly examining the results, this work draws conclusions on the potential of Transformer architectures for predicting the evolution of environmental variables over time and offers recommendations for future improvements and applications.

Postgraduate Studies Programme

Προηγμένα Συστήματα Πληροφορικής - Ανάπτυξη Λογισμικού και Τεχνητής Νοημοσύνης

Department

Σχολή Τεχνολογιών Πληροφορικής και Επικοινωνιών. Τμήμα Πληροφορικής

Number of pages

Language

Greek

URI

https://dione.lib.unipi.gr/xmlui/handle/unipi/18896

Collections

Τμήμα Πληροφορικής

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