Blast furnace ironmaking is a kind of energy-intensive and complex industrial process. To carry on a steady operation of the blast furnace, reduce energy consumption, and improve product quality, it is important to predict silicon content in real time. However, in such a complex industrial process, we need to consider changes in operating conditions. To solve this problem, we propose an unsupervised feature alignment domain adaptation regression (UFADAR) based soft sensor modeling method. By reducing the feature subspace distance between the source domain and the target domain and introducing a cycle-consistent adversarial network to ensure the reversibility of domain features, the proposed model can reduce the distribution discrepancies between different operational conditions, achieve the prediction of quality variables, and improve the accuracy and robustness of the blast furnace soft sensor model. The effectiveness of the proposed UFADAR is verified by the case study of silicon content prediction in the blast furnace ironmaking process. Experimental results show that the proposed model is superior to other prediction models in prediction accuracy.