Fuel　cells　are　a　promising　solution　for　increasing　driving　range　of　electric　vehicles.　To　guarantee　the　high　efficiency　and　stable　operation　of　fuel　cells,　the　effective　regulation　of　oxygen　and　hydrogen　reactants　is　needed.　During　varied　driving　conditions,　in　which　the　drastic　current　demand　changes　may　result　in　insufficient　reactant,　the　fuel　cell　can　even　be　damaged.　In　this　paper,　a　hierarchical　model　predictive　control　(HMPC)　strategy　is　proposed　based　on　the　deep　learning　for　vehicle　speed　predictions.　Speed　variation　predictions　are　considered　by　the　MPC　to　regulate　the　air　supply　system　and　preventing　oxygen　starvation　in　the　fuel　cell　stack.　The　problems　of　fuel　cell　oxygen　stoichiometry　control　are,　first,　stated　with　the　preliminary　energy　management　description　as　well　as　with　the　limitations　of　the　traditional　MPC.　The　deep　learning　Back　Propagation　(BP)　neural　network　was　then　designed　as　the　first-level　predictor　to　forecast　the　vehicle　speed　by　training　with　integrated　driving　cycles,　and　predict　the　fuel　cell　current　based　on　its　cathode　flow　model.　Subsequently,　the　second-level　MPC　used　the　current　disturbance　prediction　and　filling　is　introduced　to　regulate　the　oxygen　mass　flow.　The　simulation　results　for　the　MANHATTAN　drive　cycle　demonstrated　that　the　root　mean　square　error　(RMSE)　for　speed　predictions　was　less　than　1　km/h.　Compared　with　the　conventional　MPC,　HMPC　offers　better　robustness　in　the　face　of　influence　from　current　changes　induced　by　speed-variations,　and　the　RMSE　of　the　oxygen　stoichiometry　control　was　decreased　by　63.37%.　©　2020　Elsevier　Ltd