Abstract: Data-driven industrial process modeling needs to rely on a large number of well-labeled data sets, but the labeled data directly related to the target task is often limited. Therefore, the auxiliary training data with correlation can be used for modeling to achieve task transfer. However, the imbalance of samples affects the performance of transfer learning to a certain extent. Therefore, We propose a multi-source transfer learning method based on sample balance strategy. First, for the samples in the same source domain, using the least squares method to build a single source predictor by fusing multiple candidate predictors, so that the transferable information of different samples in the same domain can be synergized. In addition, for samples among different source domains, a multi-source prediction model is obtained by weighting multiple source predictors based on the error function. Finally, the case study of ethylene distillation column verifies the effectiveness of the proposed method.