Model-based numerical prediction is often affected by bias when compared to local observations. In this study, the European Center for Medium-Range Weather Forecasting (ECMWF) data were used to generate the analog ensemble (AnEn) prediction over the 15 national weather stations and 274 automatic stations of Beijing, with a focus on correcting ECMWF prediction of the daily maximum and minimum temperatures, 1-7 day ahead, twice a day. The analog of a forecast for a given location and time is defined as the observation that corresponds to a past prediction matching selected features of the current forecast. The best analogs form AnEn, which produces accurate predictions and a reliable quantification of their uncertainty with similar or superior skill compared to traditional ensemble methods while requiring considerably less real-time computational resources. An analysis of the performance of ECMWF and AnEn in space and time was presented. The results demonstrate that a short training period of 60 days may be a good compromise for the computational efficiency and the quality of deterministic predictions. Extending the training periods would further increase the prediction quality than optimizing the environmental parameters, no matter 1-month, 3-month or 6-month optimizations. AnEn correction results are better than the predictions generated by the forecasters, particularly for daily minimum temperatures. AnEn effectively reduces the bias of ECMWF predictions, resulting in a skilled downscaled prediction at the observation location, consistently over time and space. However, AnEn is not very effective in improving predictions of haze, precipitation, and strong winds, which may require a much longer training data set. Furthermore, this study tests the results over time and space to make sure the method’s reliability for the future smart grid forecast operation.