Hyperparameter tuning design of performance indicators prediction for hydrogen fuel cells based on Bayesian optimisation with AIGC

Abstract

Hydrogen fuel cells often encounter unstable performance and anomalies during operation, highlighting the need for continuous condition monitoring and iterative design improvements. Efficient prediction models are essential for assessing the performance of deployed cells, providing valuable feedback for enhancing designs and developing new products. Machine learning (ML) algorithms are widely used to build such models. However, the vast operational parameter data and hyperparameters without clear physical significance hinder training efficiency and prediction accuracy. Nowadays, artificial intelligence-generated content (AIGC), such as ChatGPT, offers novel solutions to these challenges. This study introduces AIGC-based knowledge queries into ML hyperparameter design for hydrogen fuel cell performance prediction. Firstly, the complex nonlinear relationship between state parameters is accurately measured by integrated information theory with copula entropy to identify interpretable key features. Then, an improved XGBoost model for performance prediction is established by Bayesian optimisation with AIGC to achieve the fast optimisation design of hyperparameters. The method is validated using actual industrial data including 950,000 state parameters and performance indicators, achieving a 53.95% efficiency improvement over conventional Bayesian optimisation while ensuring accuracy. This study provides technical support for the design and monitoring of hydrogen fuel cells and a feasible idea for the industrial application of AIGC.