Efficient matching of the preheat-numbers of the molten iron ladle with yield of molten iron in a single iron time is conducive to reducing energy waste, so it is of great economic significance to iron and steel enterprises to accurately predict yield of molten iron according to the characteristics of the per furnace. In this article, considering the variation of yield of molten iron and corresponding synergistic parameters depending on furnace conditions, we use the multi-head attention mechanism to capture the Attention-score of the feature parameters, and the weight matrix of the predictor is dynamically adjusted, so that it can assign more training weights to high-value feature parameters in real time according to the change of furnace conditions, efficiently complete the prediction of yield of molten iron. First, twenty characteristics are selected by mutual information method. Then, the influence degree of selected features on yield of molten iron is calculated in real time. Finally, stacked denoising auto encoder (SDAE) is used to train the deep neural network, and we constructed the predictor after adjusting the parameters to optimal. The test results show that the prediction accuracy of proposed model is 95% under the error range of ±50 t, which is higher than traditional SVM, DAE, and SDAE model without the multi-head attention mechanism. Ultimately, we analyze and quantify the influence process of 20 factors on yield of molten iron, and develop a dynamic prediction system of yield of molten iron based on proposed model, which has good applicability for the prediction of yield of molten iron and effectively guides the transfer scheduling of molten iron ladles.