Evaluating Large Language Models for Turboshaft Engine Torque Prediction

Recent advancements in deep learning (DL) have introduced transformative opportunities for time series forecasting, particularly through the use of transformer-based architectures. Among these, Large Language Models (LLMs), originally developed for natural language processing (NLP) tasks, have demonstrated strong capabilities in modeling sequences, learning from limited data, and integrating heterogeneous inputs. Their scalability and generalization abilities suggest that they could be leveraged beyond textual applications, offering new perspectives in industrial forecasting tasks. This study investigates the potential of LLMs for time series forecasting in the aerospace sector, with a focus on turboshaft engine torque prediction. Accurate prediction of engine torque is vital for maintaining the reliability, efficiency, and safety of helicopter operations. Although statistical models and DL architectures have achieved notable success in this domain, the application of LLMs remains relatively unexplored. To address this gap, this research evaluated and compared the performance of four different models: GPT-2 and ChatGPT (general-purpose LLMs), TimeGPT (a specialized transformer-based model for time series data), and a conventional time series transformer model. The models are assessed on a real-world dataset derived from a Bell 407 helicopter’s Health and Usage Monitoring System (HUMS), incorporating multiple exogenous variables relevant to engine performance. The evaluation includes forecasting accuracy, robustness to reduced training data (few-shot learning), and classification performance with respect to operational torque thresholds. Our findings indicate that GPT-2 and TimeGPT achieve a strong predictive accuracy. ChatGPT demonstrates potential as a prompt-based LLM. This work offers new insights into the applicability of LLMs to time series tasks in aerospace. It also highlights the importance of prompt design, data modality adaptation, and architectural specialization for achieving competitive performance. Overall, LLMs represent a promising direction for developing scalable and adaptable forecasting solutions in industrial domains.

Author dc:creator

• Alessandro Tronconi (23291317)

• Time series forecasting
• Large Language Models

Statement dc:rights

• In Copyright

DOI dc:identifier

https://doi.org/10.25417/uic.31451059.v1

OAI identifier oai:identifier

oai:figshare.com:article/31451059