Abstract: Concentrate grade is an important indicator of flotation effectiveness and has been the focus of attention in the industry. The principles of three machine learning algorithms, namely, BP neural network, random forest and partial least squares （PLS）, are introduced respectively. Using the production data of the flotation process of a large domestic copper ore processing plant, the prediction results and errors of the above three machine learning algorithms on the copper concentrate grade in the copper flotation process are compared and analyzed. The results show that random forest is the algorithm with the smallest prediction error under the conditions of the data studied in the article, and the percentage of the number of samples with the absolute error of prediction in the range of ±1% is 91.78%, and the percentage of the number of samples in the range of ±2% is 99.43%, and the MAE is 0.462 6, and the MSE is 0.383 9. it can predict the grade of the copper concentrates more accurately, and provide real-time decision-making support for the operators. It helps to maintain the stability of concentrate grade in the production process.