In gas-solid two-phase flows, complex spatiotemporal mesoscale structures are present, making the development of accurate mesoscale drag models essential for precisely simulating fluidization dynamics. The filtered gas pressure gradient, as a key physical quantity in mesoscale drag modeling, has been extensively studied and applied in the simulation of mesoscale resistance. Typically, during the filtering process, the pressure gradient is derived at two discrete scales: the fine-grid scale and the filter scale. The pressure gradients calculated at these two scales often differ significantly, leading to considerable discrepancies when used in numerical simulations of fluidized beds. These discrepancies can affect the predictive accuracy of mesoscale drag models and hinder their practical application in simulating industrial-scale systems. To address these challenges, an artificial neural network (ANN) approach is employed to systematically analyze the influence of pressure gradients at different filtering scales on the construction of mesoscale drag models. Two drag models are developed based on the pressure gradients at the fine-grid and filter scales, referred to as model A and model B, respectively. Through extensive validation and analysis, the performance of these models is compared across a range of filtering scales and flow regimes. The results reveal that model B, which uses the pressure gradient at the filter scale, exhibits superior predictive capabilities compared to model A, particularly at larger filter scales. This superiority is further substantiated through posterior analysis. Simulations of fluidized beds under three typical flow regimes—bubbling, turbulent, and fast fluidization—demonstrate that model B provides axial solid volume fraction distributions that align more closely with resolved results than those predicted by model A. These findings confirm that constructing mesoscale drag models based on the pressure gradient at the filter scale enhances their accuracy and applicability, making them more suitable for simulating complex fluidized bed dynamics in industrial and research contexts.