Abstract: In hot rolling, actual production data were not used to improve the thickness quality of products. For this phenomenon, an online data-driven modeling algorithm was proposed for strip thickness control based on subtractive clustering. Firstly, the input space is divided into several clusters by subtractive clustering, in each cluster a sub-model is built by a least square support vector machine (LS-SVM), and an offline model is obtained by weighting the outputs of these sub-models. Then, when the online data constantly increase, the clustering subsets are adjusted on-line by a subtractive clustering algorithm, and the parameters of the sub-models are determined using the recursive algorithm of the least squares support vector machine. The predictive outputs of the sub-models are the final outputs. Experimental results demonstrate that the method has good prediction accuracy and online learning ability.