The superposition of aging characteristics in fuel cells is a major cause of inaccurate predictions. Unlike traditional methods that mix linear and nonlinear aging characteristics, this paper develops a prediction method based on Symplectic Geometry Mode Decomposition and Divide-and-Conquer Gated Recurrent Units (SGMD-DCGRU). The SGMD uses symplectic geometric transformations to decompose the aging characteristics of fuel cells into three types: linear, periodic fluctuations, and nonlinear aging characteristics. Leveraging this foundation, the DCGRU network provides distinct predictions for each aging sub-characteristic through an integrated approach that includes environmental variable feature extraction, a periodic time node attention mechanism, and bidirectional gated recurrent units. This approach ensures compatibility between aging characteristics and data-driven algorithms, thereby improving prediction accuracy. Furthermore, the Kepler optimization algorithm (KOA) is designed to optimize the hyperparameters of the DCGRU network and embedded in the multi-step prediction strategy. Finally, static and dynamic aging data are used to verify the performance of the proposed algorithm in multi-step short-term prediction and long-term remaining useful life prediction. In this case, the proposed method can improve the reliability and efficiency of the fuel cell system in various industrial applications, thus improving the maintenance strategy and reducing the operating cost.