Abstract: Numerical simulation is an essential technique for CO2 geological storage operations. However, high-resolution geological models typically consist of a large number of grid blocks, making numerical simulations computationally expensive and time-consuming. Upscaling methods are commonly used to coarsen the fine-scale geological model, with global flow-based upscaling methods generally demonstrating the highest accuracy. However, since these methods require solving flow equations over the global domain, which is still time-consuming, their applications are typically limited to cases where the coarse model is reused repeatedly (e.g., history matching or optimization). To overcome these limitations, this study develops a novel deep learning (DL)-based upscaling framework for the simulation of CO2 injection into saline aquifers. The framework incorporates convolutional neural networks (CNNs), Transformer encoders, and Fourier neural operators (FNOs) to construct surrogate models for upscaled well index, permeability, relative permeability, and capillary pressure. A preprocessing procedure is first applied to address the issue of inaccurate upscaled parameters, which are typically caused by weak flow conditions in traditional upscaling computations. Then the surrogate models are trained using relevant local information, and the trained surrogate models are used to replace traditional numerical upscaling computations, enabling instantaneous and parallel predictions of upscaled parameters. Two representative flow patterns (left-to-right and bottom-to-top) are considered to evaluate the framework's performance. The results demonstrate that the DL-based framework significantly improves computational efficiency, achieving a speedup factor of approximately 1133 times compared to traditional upscaling methods. Additionally, it maintains or even enhances simulation accuracy, as the surrogate models correct inaccurate upscaled parameters.