The degradation characteristics of fuel cell (PEMFC) vary significantly across different models and operating conditions. Accurately capturing the complex temporal relationships and long-term degradation trends with limited operational data is a challenging task. This paper proposes a multi-condition transfer learning prediction model based on a combination of Bayesian optimization

temporal convolutional networks (TCN)

and long short-term memory networks (LSTM). This model effectively utilizes prior knowledge from the source domain

reducing reliance on target domain data

and enables voltage degradation prediction under various fuel cell stacks and operating conditions. The method improves degradation prediction accuracy by an average of 34.3% under dynamic operating conditions. Additionally

excellent generalization performance of this model is validated through transfer learning experiments across static

quasi-dynamic

and dynamic condition datasets. Moreover

in comparison with several advanced transfer learning models

the proposed approach reduces average time costs by 41.56% and increases average accuracy by 48.29%，demonstrating its strong advancement.