1 030 sets of concrete compressive strength test data were used to train the multilayer perceptron（MLP） through beetle antennae search algorithm（BAS） and combined with shuffled complex evolution（SCE）， multi-verse optimizer（MVO） and the results of two MLP coupling algorithms were compared and analyzed， and an algorithm model that can be used to predict the compressive strength was obtained. The results show that BAS can significantly improve the training accuracy and prediction accuracy of MLP， and this model is faster， more accurate and stable than the other two coupled model algorithms. Compared with artifical neural network（ANN） and support vector machine（SVM） individual learning algorithms， the prediction of concrete compressive strength shows the superiority of meta-heuristic algorithm. The factors related to the number of training data sets and the number of input variables in the BAS-MLP model are discussed， and it is found that using 80% of the 1 030 sets of data can obtain good prediction results.