This paper proposes a design method of power system stabilizer by a least squares support vector machine (LS-SVMPSS) for wide area stability control. Both local and inter-area data based on synchronized phasor measurements considering time delay are considered as the input features of the LS-SVMPSS. A large number of the training data sets of a multi-machine power system are reduced by the measurement of similarity between samples. Removing the redundant data in the training set not only improves the LS-SVMPSS performance but also decreases computation expense during the operation of LS-SVMPSS. The LS-SVMPSS parameters and the similarity threshold are optimized by a genetic algorithm. As a result, the redundant data in the training set can be discarded while the reduced data are the optimal support vectors in the LS-SVMPSS model. The LS-SVMPSS control signals can be adapted in real time by various operating conditions and different disturbances. The performance of the LS-SVMPSS is compared with the conventional PSS and the neural network-based PSS. Simulation results in a two-area four-machine power system demonstrate that the proposed LS-SVMPSS is very robust to various disturbances under wide range of operating conditions in comparison to other PSSs.