Temperature prediction of molten steel always plays a significant role in the smelting process of the Consteel electric arc furnace (EAF), as obtaining the accurate temperature timely contributes to refining the whole process and reducing the production cost. However, limited by the extreme smelting conditions of the EAF and the rather sparse temperature measurements, neither the physical-based model nor the data-based machine learning one alone is able to achieve precise prediction results. To address this problem, we provide a temperature prediction model by integrating mechanism knowledge with a LSTM network. First, the model framework is constructed according to intrinsic characteristics of the Consteel smelting process. We leverage a LSTM network as the core of the model to accomplish the basic task of continuous temperature prediction. Then, a special constraint layer is developed to limit the output of the intermediate layer of the model to a reasonable range, which is determined by the energy balance of molten steel. In this way, valuable smelting knowledge can be embedded in the training process of the network, and the problem of label insufficiency is compensated resultantly. Finally, a series of experiments are conducted based on a practical dataset collected from a Consteel EAF production site. The proposed model is compared with several state-of-the-art models, where the results illustrate its superiority.