Aiming at the prediction for coke oven gas holder level in steel enterprises, which is very difficult to be modeled using the mechanism modeling, a gasholder level prediction model combined with the analysis of the gas production-consumption and level variation is established based on the least square support vector machine. A gradient grid search algorithm for selecting the model’s parameters and an effective big samples selection approach to build the training samples are proposed to improve the prediction accuracy. The simulation results using the practical gas data in Shanghai Baosteel show that, the proposed method has shorter parameter optimization time and better performance, and can provide scientific guidance for the gas balance scheduling process.