Abstract: Aiming at the problem that the sparsity of data in the high dimensional input space will lead to the biased estimation of support vector regression (SVR), an additive SVR is proposed in this paper. This algorithm is realized via the addition of the separated input spaces after kernelization. Thus, the curse of dimensionality can be overcome by the additive model such that the bias can be reduced for high dimensional regression problem. Moreover, a simplified quadratic programing (QP) formulation of SVR is proposed for easily constructing the additive SVR model. Finally, the proposed method is employed to predict the concentration of HAC in the bottom outlet. It is shown from the experimental results that the additive SVR model can attain better predication performance than traditional SVR and least quare support vector regression (LS SVR).