Towards efficient and accurate CO2 sequestration optimization: Integrating hierarchical spatio-temporal information into deep-learning-based surrogate models