Abstract: In view of the fact that the object in the dense medium separation process generally has complicated characteristics, such as multivariate coupling, nonlinearity, multiple working conditions and time-varying, and the ash content is difficult to detect online and cannot be fed back to the control system in time. Meanwhile, adopting a single data-driven model is difficult to deal with the object characteristic involving the complex industrial production process, as a result, the prediction accuracy and generalization performance of the established model cannot be guaranteed. A multi-model soft sensor modeling method for dense medium ash content based on robust stochastic configuration networks (RSCNs) and weighted K-means clustering is proposed. Firstly, the selected data set is divided by means of clustering algorithm to obtain a subset of samples, which considers the characteristics of each dimension of samples having different effects on clustering. Then, each sample subset is trained by RSCNs respectively, and to establish the dense medium ash content prediction sub-model. Finally, the final output is obtained by fusing the predicted values of each sub-model. The model prediction results are substantiated using actual industrial process data, the results demonstrate that the proposed method can predict ash content more accurately compared with other algorithms.