In this paper, an estimation model of fuel cell life degradation and lithium battery life degradation is defined based on experimental data. A dynamic programming algorithm based on global optimization is proposed, which takes the total equivalent hydrogen consumption cost of power system and the cost of fuel cell power degradation as the target function, and weights the power degradation term of fuel cell with a penalty coefficient, so that the designed energy management strategy can take into account the economy and durability of the vehicle. Considering that the control strategy based on global optimization is sensitive to operating conditions, i.e. the reduction of hydrogen consumption and fuel cell degradation costs by dynamic programming algorithm is based on operating conditions. Once the input condition of operating conditions is changed, the control strategy will not guarantee the optimum. Therefore, this paper uses the optimal results of dynamic programming algorithm to provide the data set of off-line training. Transformer is applied to make time series prediction, which will help to solve the problem of balancing power distribution considering economy and durability under various conditions. The input characteristics of the transformer network are the current bus demand power, the battery state of charge and the fuel cell power of the previous time, and the output is the fuel cell power predicted at the current time. The results of Transformer model are compared with those of dynamic programming, and they are very close. The Transformer model was trained using standard UDDS and NEDC operating conditions and tested on HWFET. The test shows that the developed Transformer model has good generalization ability for different operating conditions.