Continuous real-time prediction of carbon and temperature is the key to the end point control in the process of converter steelmaking. Aiming at the problem that the fluctuation of process data affects the similarity measurement of furnace samples, which causes difficulties in modeling and poor generality, the time series characteristics of steelmaking process data are also considered, a real-time learning method of automatic clustering and calculating the posterior probability of samples to be tested is proposed. Firstly, the fuzzy C clustering strategy weighted by grey relational degree is adopted to automatically cluster the historical database samples. Then, the mixed Gaussian model is used to calculate the posterior probability of samples to be tested to determine the sample set with the largest correlation degree. Finally, the best small sample is measured out from a subset of the samples under test to predict end point carbon temperature with the LSTM network. Through the method of data verification, steel converter steelmaking experimental results show that, in accordance with requirements of the steelmaking process, the temperature prediction error on the accuracy of $\pm 10^\circ$C is 93.33%, the accuracy of carbon content of the prediction error in $\pm $0.02% is 90.0%.