Using GRAPESGFS forecast data and temperature observation data of Guangxi regional automatic weather stations during 2017-2018, errors of the 2 m temperature forecast of the GRAPESGFS model over Guangxi are analyzed. It is found that the 2 m temperature forecast of the GRAPESGFS model is lower than the observation in Guangxi. Forecast errors increase with the forecast time and regularly appear in the mountain areas in the northern Guangxi, Zuojiang, Youjiang river valley, and coastal areas. The temperature forecast error at noon is systematic in spring, summer and autumn but the errors at noon in winter and that at night in all seasons are random. To develop the optimal timewindow of the moving average method, we compare different moving average solutions with the unfixed timewindows and verify its improvement with the trial correction products of the optimal timewindow moving average method during 2020. Results show that the moving average solutions of fixed timewindow, optimal seasonal timewindow, and monthly optimal time window are all effective in spring, summer, and autumn. The correction skill is higher at noon than that at night. Among all the solutions, fixed long timewindow (15 to 60 d) solution, seasonal optimal timewindow solution and monthly optimal timewindow solution are more effective. Running optimal timewindow method based on different moving average solutions can steadily improve the 2 m temperature forecast.