The　mismatch　between　the　supply　and　demand　in　the　ride-hailing　market　often　results　in　operational　inefficiencies,　such　as　low　vehicle　occupancy　rates　and　prolonged　passenger　waiting　times.　This　paper　proposes　a　novel　multi-step　spatio-temporal　prediction　model,　termed　the　ConvLSTM　+　model,　which　integrates　a　convolutional　long　short-term　memory　(ConvLSTM)　network　within　a　dual-layer　architecture　enhanced　by　a　residual　correction　mechanism.　The　proposed　model　is　specifically　designed　to　predict　passenger　pick-ups　and　drop-offs.　Using　taxi　datasets　from　New　York　City,　Chengdu,　and　Beijing,　the　experimental　results　demonstrate　that　the　ConvLSTM　+　model　significantly　outperforms　several　widely　used　passenger　flow　prediction　models　in　terms　of　multi-step　prediction　accuracy.　This　study　not　only　provides　valuable　decision-making　support　for　ride-hailing　drivers　but　also　offers　actionable　insights　for　improving　service　quality　and　operational　efficiency　within　the　ride-hailing　market.