In those oceans where measured wave data are not available, numerical wave models are usually adopted to hindcast wave parameters in order to define design waves for marine structures. To utilize these hindcating data, it is very important to perform error corrections of model results for accurate estimation of the appropriate wave parameters. In this paper, a new non-parametric correction model is established to improve wave model accuracy through modifying a previous approach released by Caires and Sterl in 2005. The new correction model introduces a kernel algorithm to learn error information from both value magnitude and series trend through training datasets, and utilizes the information to correct potential errors in model outputs. It is shown that the two-dimensional learning method is more effective than the previous one-dimensional which only learns error information from the value magnitude. Furthermore, an error constraint parameter is initially adopted in the new correction model to decrease the possibility of overcorrection. The new correction model performs better than its predecessor, especially when modeling wave period and altimeter synchronized wave height. Though this paper evaluates the model correcting performance with WAVEWATCH III outputs, the modified model can be adopted to correct other kinds of time-series data.