Forecasting the minimum and maximum temperature as accurately as possible is significant to reduce damage in the fields of agriculture, water, and livestock. In this research, daily maximum and minimum temperatures were forecasted using machine learning, with a training period of 100 days using the least squares of the pseudo-Fourier spectral error-time series. In this regard, the data from a two-year period were used for more than 540 synoptic stations in Iran. The forecasts of the machine learning method with the output of the WRF model were compared with each other and with the observations, and their abilities were evaluated with the skill score. The results showed that the implementation of the machine learning algorithm was successful in daily forecasts, and the average skill score was more than 0.6. While the average skill score for the WRF model is approximately 0.01. The main advantage of this method compared to the WRF model is that it is less computationally complex, and only the data from the last 100 days are used in the calculations. Another advantage of this method is the use of a machine learning algorithm with short-term memory, which reduces the adverse effect of the long-term lack of data in the training period's time series on the output of forecasts. In this method, short-term forecasts (24, 48, and 72 hours) of the minimum and maximum temperature for all Iran’s stations can be provided with appropriate accuracy, with less computational cost, and in a very short time