Optimal scheduling of generating resources plays a significant role as a decision-making tool for power system operators in the liberalized and real-time electricity spot markets. The real-time scheduling of generating units will become a very complex task with respect to the instantaneous fluctuation of the load demand due to several demand response scenarios in the smart grid context. In this study, a hybrid mathematical method for the online scheduling of units based on the least square support vector machine (LSSVM) and the third version of cultural algorithm (CA3) has been presented, where the CA3 has been specifically employed to tune the adjusting parameters of LSSVM. For the training purpose of the proposed method, the optimal scheduling of the daily load curve for four different test systems and various physical and environmental constraints of generating units have been prepared by using a modified mixed integer quadratic programming (MIQP) to deal with non-convex behaviors of the test systems. A mean squared error (MSE) objective function has been used to reduce the prediction errors during the training process to enhance the precision and reliability of the results. A radial basis function (RBF) and the proposed LSSVM-CA3 were used to check the convergence process. A high accuracy of generator schedule predictions are demonstrated by comparing the results of the proposed method with those of artificial neural networks. From the results, it can be inferred that the method is highly compatible for real-time dispatching of generation resources in deregulated electricity markets. (C) 2016 Elsevier Ltd. All rights reserved.