At present, wind forecast based on Numerical Weather Prediction is widely recognized and applied for a safer and more sufficient usage of wind sources. However, because of the unescapable inherent errors of numerical techniques, there are many negative cases of forecasts. Thus, aiming to quantize and evaluate the inherent errors of physical outcomes, this paper analyzes the characteristic of residuals between numerical results and actual measured data In statistical way, designs combined non-linear and non-parameter algorithms to correct original prediction values, and achieves probabilistic one-day-ahead 96-step wind speed forecasts. The concise process of the method can be described as followings. Firstly, this work utilizes autocorrelation analysis to verify the non noise attribute of error sequences. Based on the characteristic, adaptive and structured error correction models of nonlinear autoregressive with exogenous inputs network are established to acquire deterministic optimized outcomes. Then, aiming to calculate conditional error boundaries of different confidence levels, mixture kernel density estimation is adopted step by step to estimate joint probability density of corrected values and revised errors. The results on test set show the correction considering inherent errors of numerical techniques can integrate the physical with statistical information effectively and enhance the forecast accuracy indeed.