基于大语言模型的飞行轨迹预测方法

打开文本图片集

关键词：飞行轨迹预测；大语言模型；深度学习；指令微调；思维链中图分类号：TP391 文献标志码：A 文章编号：1001-3695（2025）12-015-3644-07doi： 10.19734/j . issn. 1001-3695.2025.04.0120

Flight trajectory prediction method based on large language models

Luo Kaiwei，Zhou Jiliu† （School ofComputerScience，SichuanUniversity，Chengdu61oo65，China）

Abstract：Flighttrajectorypredictionisacriticaltask inairtraffc managementsystems，wheredeeplearning methods have driven significant progress.However，theexisting approachesoften treattrainingdatainablack-box manner，limiting model interpretabilityLLMsexcelintextcomprehensionandgeneration，possessingpowerfulreasonngandthinkingabilities.However，fewstudieshaveexploredtheapplicationofLLMsinflighttrajectoryprediction.Thispaperproposedanovelapproach， FTP-LLM，which reformulatedtrajectorypredictionasalanguage modeling taskandpioneered the potential of LLMsinthis field.Theapproachextractedspatiotemporalfeaturesfromreal-worldflightdataandintegratedthemwithdomain-specific prompts toconstructaninstructiondatasetforfine-tuning.Toenhance interpretabilityand transparency，theprompts ncorpo ratedachain-of-thought（CoT）reasoning proces.Theexperiments fine-tuned various LLMs toevaluate their performance in trajectorypredictiontasks，whilefurtherinvestigatingtheirgeneralizationabilityinfew-shotscenarios.Experimentalresults showthatLLMsachievecertainperformance improvementsoverdeep learning methods in trajectory predictions.The Llama3.1 model achieves the highest prediction accuracy，reducing average erors by 7.16% for single-step predictions，10.71% for 4-step predictions，and 10.15% for 8-step predictions.

Key words：flighttrajectoryprediction；large language models（LLMs）；deep learning；instructiontuning；chain-of-though

0 引言

全球经济的快速发展显著提高了各行业对航空运输的需求，空域复杂性不断增加，空管系统面临愈加严峻的运行压力[1]。（剩余17711字）