Molten iron quality (MIQ) indices prediction based on data-driven models is an important way to monitor product quality and smelting status in the blast furnace ironmaking process. However, some challenges still place in the MIQ prediction: 1) limited nonlinear and dynamic description capabilities and interpretability of data-driven models; 2) high demand on the number of the labeled samples; and 3) insufficient exploration of the underlying relationship between MIQ indices. In this case, we propose a novel data-driven deep model for the online prediction of MIQ indices. First, we design an attention-wise module to self-learn the nonlinear and dynamic relationship between process variables and prediction targets and enhance interpretability. Then, the minute-level molten iron temperature (MIT) data detected by our previously developed equipment is used to pretrain the attention-wise deep network to obtain the improved weights and reduce dependence on labeled samples. Finally, the pretrained model is extended to a structure with a weight-shared attention-wise module and task-separated prediction networks to explore the relationship between multiple prediction tasks. The effectiveness of the proposed attention-wise deep network is verified in an industrial ironmaking plant, which shows a significant improvement in performance, i.e., high accuracy and interpretability.