Large language model-based neural architecture search for efficient hydro-turbine fault detection

Hydropower stations, vital to renewable energy, face challenges from turbine faults that affect stability and efficiency. This paper introduces an innovative Large Language Model -based Neural Architecture Search framework to significantly improve hydro-turbine fault detection accuracy and efficiency. Our method synergistically combines an LLM with NAS, automating the optimization of the fault detection process after data preprocessing with clustering techniques. The LLM interactively guides the NAS by generating initial search spaces and design principles, and dynamically refines these based on feedback, reducing manual intervention. This LLM-guided evolutionary search is enhanced by zero-cost proxy evaluations for rapid filtering of suboptimal architectures and an integrated adversarial hyperparameter optimization for improved model robustness during architecture assessment. Validated on monitoring data from a Yangtze River Basin hydropower station, our LLM-NAS model achieves high precision (0.899), recall (0.930), and an F1 score (0.914). These results outperform traditional machine learning, standard deep learning models, and demonstrate competitive performance with significantly reduced search time (18.56 GPUh) compared to other mainstream NAS techniques. The framework thus offers a robust, efficient solution for hydro-turbine fault detection, presenting a novel paradigm for LLM-augmented automated model design.