Abstract: Based on air temperature from ECMWF data, three groups of training period schemes that include original schemes, improved schemes and application schemes are designed to test and compare results of daily maximum and minimum temperature of national stations in Fujian Province twice per day from 2014 to 2015. For the quasi-symmetrical mixed running training period (QSRTP) method, several days of current year before forecasting day and equal numbers of days in last year after are adopted separately to compose an initial 1-year or multi-year dataset. When using multiple-year data, lengths of the optimal training period every year are different or the same in each scheme. From three original schemes, it is found that the methods of QSRTP are much better than other running training period methods and traditional fixed period classification, and the number of optimal training day is more stable. The QSRTP with 2-year data shows better performance than that with 1-year data for comparison of improvement schemes, and different training period lengths lead to better performance in 2-year evaluation schemes. Considering model version updating and the continuity of weather, lengths of the optimal training period before the first year are slightly longer than that in the second year. Similar to the traditional method, the optimal training period of original and improved schemes are obtained after verification rather than before forecast. Three application schemes with different terms of evaluation are designed to test the stability, the usability and seasonal patterns of the optimal training period. Based on 1-year evaluation, total samples of the training period are stable with the best forecasting score. In terms of the monthly evaluation, the best period has no significant patterns with a relatively low score. For 10-day evaluation, the best period varies greatly, but when the forecasting time becomes shorter, the forecast quality becomes better. When the weather changes suddenly, the optimal training days will be quite different. Hindcasting experiment of 2015 suggests that the MOS forecast for daily maximum and minimum temperature using the optimal training period of last year has a much higher score than the original ECMWF products, it is better than the subjective forecast, and the forecast absolute deviation is significantly reduced. With the accumulation of data in the future, the forecast quality would be improved greatly, indicating that the method of multi-year QSRTP has an important application prospect on the daily operation.