Data feature selection of converter steelmaking process is the key step to realize the end point carbon content and temperature prediction. The highdimensional data of production process are not conducive to the rapid and accurate prediction of the end point carbon temperature. To address this problem, an improved genetic algorithm is proposed to select the data feature of converter steelmaking process. Firstly, Pearson correlation coefficient is used to measure the important contribution of different features. Then, the objective function is formulated to reflect the correlation between process data feature and terminal carbon temperature. The maximum, minimum, average fitness and random individual fitness of the population are defined by the objective function. In this way, an adaptive crossover mutation probability mechanism is established. This method not only makes the population distribution more reasonable during the iteration optimization, but also improves the late convergence speed to prevent the algorithm from falling into local optimization. Through verification and comparison experiments of data feature selection in actual steel mills, results show that the average time of feature selection is 025 s, the accuracy of temperature error within ±5℃ in terminal prediction is 8567%, and the accuracy of carbon content prediction error within ±001% is 8067%.