Outlet moisture content in tobacco drying is affected by heat supply, heat transfer, moisture removal regulation, and material states, showing multi-scale temporal characteristics that include overall evolution, stage-wise variation, and short-term fluctuation. To address the insufficient representation of dynamic information over multiple temporal ranges and the weakening of local fluctuations in continuous production, a multi-scale temporal dependency method is proposed for online quality prediction. The method uses the effective material processing duration as the temporal observation scale, constructs a hierarchical multi-scale backbone to extract process features, applies a local-global split attention mechanism to coordinate local details and global dependencies, and combines cross-scale feature fusion with local dynamic compensation to enhance the representation of local peaks and valleys. The method is validated using online historical data from a cigarette factory. The mean squared error, mean absolute error, root mean squared error, and coefficient of determination on the test set are 0.0004, 0.0128, 0.0209, and 0.9850, respectively. The prediction errors are lower than those of typical time-series prediction models, indicating that the method can support processing-state perception and stable outlet moisture control.