This study leverages the Weather Research and Forecasting (WRF) model, integrated with the three-dimensional variational (3DVAR) data assimilation system, to conduct a high-resolution simulation of a thunderstorm gale event in Beijing. By assimilating observations from Automatic Weather Stations (AWSs), the impact of different observation errors is systematically assessed and explored. This study demonstrates that varying observation errors significantly influence the performance of the wind analysis and forecast. Mitigating observation errors in cases with significant observation-minus-background (OMB) discrepancies enhances the reliability of observations with large departures from the background, resulting in greater analysis increments and improved agreement with the assimilated observations. In contrast, the Desroziers method computes observation error covariances through statistical analysis of OMB and observation-minus-analysis (OMA) residuals. This approach systematically assigns higher error estimates, enabling the assimilation of more observations with reduced weights while enhancing spatial continuity in the analysis. Both experiments employing modified observation error parameters demonstrated superior predictive accuracy compared to the default configuration, enabling more precise identification of strong wind regions while mitigating overestimation tendencies. Statistical metrics (BIAS, RMSE) and FSS scores confirm their superior performance, particularly in extreme wind speed forecasting, with Exp_D05 outperforming all other experiments. These findings highlight the importance of optimized observation error estimates in AWSs data assimilation for improved wind analysis and forecasting.