Kernel recursive least-squares (KRLS) has shown better predictive energy efficiency in time series prediction. However, in complex and non-stationary environment, there are still some problems of low prediction efficiency and accuracy. In view of these problems, we propose adaptive sparse KRLS (RP-ASKRLS) with random projection. RP-ASKRLS introduces random projection into KRLS, which can sparse data and maintain manifold information. On this basis, sliding window sparse strategy and adaptive update standard are integrated, which can effectively restrain the dimension of kernel matrix, and track time-varying characteristic. Therefore, RP-ASKRLS can not only availably constrain testing time, but also reduce computational complexity, thus better prediction effect is obtained. The experimental results show that RP-ASKRLS online prediction has better forecast performance.