Abstract: Aiming at the problems of sample accumulation, model inflation and slow online updating speed in engine condition online prediction process, an online prediction method based on incremental sparse kernel extreme learning machine(ISKELM)was proposed. Firstly,a sparse measurement matrix was defined for the kernel function matrix of KELM, and the operations of forward sparseness and backward deletion for large-scale samples were performed according to the principle of sample information measurement consisting of coherence minimization and self-information maximization. It improves the efficiency of sample sparseness. Then the sparse measurement matrix was expanded and pruned online by using the effective samples under the best dictionary order, which limited the model inflation. Lastly, the kernel weight matrix of the model was updated in a recursive way through the improved incremental modeling method. So an online learning model of ISKELM with a limited order and sparse structure was established to distinctly improve the online modeling speed. The online prediction experimental results with simulation data and engine condition parameters show that, compared with two existing online prediction methods, ISKELM has higher efficiency of sample sparseness and online modeling. When the engine exhaust temperature is predicted by 120 steps, the prediction speed is increased by 80.50% and 31.72% respectively, and the prediction accuracy is improved by 48.56% and 15.81% respectively.