Effective prediction of calcium addition in molten steel treatment is important for inclusion modification and clean steel manufacturing. This study presents an application-oriented framework that integrates multimodal feature fusion, a CNN-BiLSTM-Attention prediction network, adaptive differential evolution, and SHapley Additive exPlanations (SHAP) for calcium-addition prediction. The fusion module uses modality-level attention to combine process parameters, chemical composition measurements, and sensor signals. The prediction network extracts local patterns, models bidirectional sequence dependencies, and weights informative sequence positions. An improved adaptive differential evolution algorithm with chaotic initialization and self-adaptive parameter control jointly searches network hyperparameters and input feature subsets. On the reported split of 2 846 industrial heats, the framework achieved an RMSE of 0.1038 kg/t and an R² of 0.9843 and produced the lowest reported error among the six compared baselines. The ablation comparisons indicate complementary contributions from the tested components, while SHAP analysis yields feature rankings that are broadly consistent with metallurgical knowledge. These findings support the feasibility of the integrated framework under the reported experimental configuration, although broader validation across steel grades, plants, and repeated random splits remains necessary.