The construction of feature vector is a key issue for protein secondary structure prediction. In the present methods, only the BLOSUM62 matrix is taken into account, which neglects the amino acid mutation of protein in the evolutionary process. In this study, we propose to construct feature vector by combining PSSM matrices of different evolutionary times, which cannot only reflect the position information, but also reflect the interaction of amino acids. Based on the feature vector, logistics, randomforest and M-SVM_CS models are utilized to predict protein secondary structure on the public datasets (RS126, CB513, and 25PDB). The experimental result demonstrates that the method can achieve a better performance than traditional methods.