Molten steel temperature prediction is a crucial step in ladle furnaces (LFs). Due to the complicated working conditions, process data usually suffers from various types of outliers. However, most of existing temperature models have not taken robustness to outliers into account. Hence, their accuracies usually cannot satisfy the industrial production. In this paper, we propose a comprehensive scheme that integrates temperature prediction with outlier detection. Of this scheme, we develop a three-level ensemble model where Gaussian process (GP) is used as the base learner, to accomplish the prediction task. Motivation for GP base leaner is two-fold. One is that GP models perform well on the nonlinear regression problem. The other is that GP is a Bayesian method and its output can be used in the outlier detection step. Motivation for our ensemble model is also two-fold. First two problems regarding GP, i.e. high computational complexity and model selection, can be alleviated. Second, the predictive accuracy can be further improved. As for the outlier detection task, we develop two types of detectors implemented for both training and testing data points. We proposed a novel detector based on one-class classification (OCC) and use it for training samples and inputs of testing data. And the detector for the output values of testing data is constructed from outputs of the prediction model. Finally, we verify the prediction performance on several real-world data sets and compare their performance with several competitors. The significance of the proposed outlier detection step is also validated. Experimental results approve the potential of our scheme. (C) 2018 Institution of Chemical Engineers. Published by Elsevier B.V. All rights reserved.