Abstract: A modified Learn++ algorithm adaptable to regression problems is applied to modeling soft sensors to rise above the defects in conventional method of soft sensor modeling. In the Learn+ + algorithm not only the conventional integrated algorithm can be kept up to improve the performance of single learner, but the drawbacks that the learnt information is easy to be forgotten by existing learning method of soft sensors and the waste of time/resource due to using repeatedly original training data can both be overcome. Learn+ + algorithm thus has incremental learnability. ELM algorithm is selected as a " weak learner " in modeling process because it learns faster and simpler and has higher generalizability than conventional neural network, thus getting rid of the limitation of local minimization and overfitting. The new modeling method based on Learn+ + has been used to model the soft sensors for molten steel temperature in LF, and the results revealed that its high accuracy can meet actual requirement in production.