Transfer learning effectively utilizes knowledge from the source task to reduce the dependency of data-driven fuel cell health prediction methods on large datasets. However, in practical applications, the dynamic variations in stack operation hinder the knowledge transfer between the source and target domains. Additionally, the complex and variable operating conditions of PEMFC pose additional challenges for accurate prediction and diagnosis. Accurately characterizing the degradation characteristics of future fuel cell stacks under complex conditions using limited data under specific operating conditions is crucial for effective fuel cell health monitoring. To address this issue, we propose a transfer learning model based on LSTM, which successfully transfers knowledge from static condition cells to dynamic condition cells, effectively tracking the voltage variations of the stack under dynamic loads. Its generalization capability is thoroughly demonstrated across nine current levels in the activation and ohmic regions, providing a more comprehensive insight into PEMFC health management under dynamic load conditions.