Automated generation of forecasting texts in aviation weather for thunderstorm activity: a hybrid workflow combining rule-based processing and large language models

Summarise with VIRAT
Authors: Alexander Poscher
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

This master's thesis examines the generation of forecasting texts in aviation weather using rule-based preprocessing and a large language model (LLM). Forecasting texts in aviation weather aim to provide a general overview of the weather situation in Austria. The specific focus of this thesis is on generating forecasting texts in aviation weather for thunderstorm activity. Toward the goal of generating forecasting texts in aviation weather, an AI-based workflow was developed as a proof-of-concept prototype for generating forecasting texts in aviation weather, including preprocessing, summarization, verification and comparison. Therefore, the AI-based workflow additionally serves as a tool for assessing the correctness of generated forecasting texts in aviation weather. An experiment was conducted using 145 cases of historical weather data containing thunderstorm activity to assess the correctness of the generated forecasting texts in aviation weather based on the preprocessed data. Additionally, 96 cases were used to assess the completeness of generated forecasting texts in aviation weather, comparing them to human-written forecasting texts in aviation weather in German. The correctness scores show that the format of the preprocessed data serving as input for the LLM has an impact on the generated forecasting texts in aviation weather. Using preprocessed data formatted as natural language as input for the LLM leads to better results than using preprocessed data formatted as tabular format. The completeness scores show that the generated forecasting texts in aviation weather differ from the human-written reference forecasting texts in aviation weather. An expert evaluation revealed shortcomings in preprocessing with missing details of thunderstorms, such as shape, direction of movement and size.

Fields of Study & Subjects

Topic Modeling Natural Language Processing Techniques Meteorological Phenomena and Simulations

Publication details

Journal / source
University Library Linz repository (Johannes Kepler Universitat Linz)
Publication date
1 January 2026
Language
EN
Document type
Thesis