Since the energy and environmental problems have become increasingly prominent, wind energy as an important renewable energy resource has been paid more and more attention in recent years by virtue of its widely distributed, pollutionfree and renewable [1]. However, due to the characteristics of randomness and volatility of wind power generation, largescale wind power access to power grid will cause great impact on the power system is not conducive to the maintenance and smooth operation of the power system. Wind speed perturbation easily causes great changes of voltage and frequency of the power grid system, seriously the power system will be instable [2]. The study about short-term forecast on wind speed and power of wind farms is beneficial to stable power system operation, which prompting the relevant dispatching department to adjust the plan according to the forecast result, so as to reduce the influence of intermittent wind power [3]. At present, many experts at home and abroad have done a lot of research on short-term wind speed forecasting of wind farms. Wind speed forecast can be classified into three categories according to technology: digital weather forecast, statistical method and neural network forecasting method [4]. In recent years, artificial neural network (ANN) has made some achievements in wind speed forecast. However, ANN has some weakness. In fact, it is difficult to determine the network structure, besides over-learning and easily fall into the local minimum [5,6]. The support vector machine (SVM) method based on statistical theory can solve the problem of small sample and non-linearity well. It is proved to be better than artificial neural network and other methods in wind speed forecast. Hu Qian et al. proposed an integrated learning model based on Adaboost to forecast the short-term wind speed and obtain desirable results for solving the problem that the traditional single SVM model is not accurate [7,8]. The least squares support vector machine (LSSVM) replaces the inequality constraints with equality constraints on the basis of SVM, which avoids the time-consuming quadratic programming problem, so it is a powerful tool for nonlinear system modeling and prediction [9]. In the LSSVM model, the parameter penalty factor and kernel function have great influence on the forecast effect of the model. Therefore, it is the key to forecast the wind speed by determining the reasonable LSSVM penalty factor and kernel function [10]. Fang Biwu and others the parameters of LSSVM to achieve short-term wind speed forecast by improving firefly algorithm [11–۱۳]. Zhang et al. adopted genetic algorithm to optimize the LSSVM parameters. But the genetic algorithm is not suitable for the whole modeling because of the complex operation. Some scholars optimized LSSVM by using genetic algorithms and other bionic algorithm. Chappelle proposed gradient descent method to optimize LSSVM parameters. Despite the fact that the efficiency of genetic algorithm has been significantly improved, this approach is easy to fall into the demerit of local optimum.