Research On Modeling Of Harmonic Source For Power Electronic Devices Based On Least Squares Support Vector Machine

Power electronic devices are the major sources of harmonics, but they are also an effective way for improving power supply reliability and transporting sinusoidal voltage and current to the terminal-users. It is necessary to study harmonic sources of power electronic devices and establishing appropriate models to express the non-linear characteristics. The established models are the important tool to estimate the harmonic level and develop harmonic suppression strategy.Through analysising the existing methods of modeling harmonic sources, this paper founds that the established model based on internal mechanism applies only to a particular topology of the device. When the device changes the internal structure, it must be re-anlysis the changed topology and then re-establish a new model for it. But neural network treats the harmonic source as a black-box to establish models to express the non-linear characteristics of current and voltage, regardless of their internal principles. Therefore the method of establishing model by neural network has general applicability. Taking into account that support vector machine can overcome the drawbacks that neural network is difficult to avoid, and support vector machine has the advantages of generalization ability and convergence to the global optimum. So this paper presents a method for modeling harmonic sources by least squares support vector machine.First of all, on the base of introducing the models of typical components in power system, this paper summarizes and compares the existing methods of harmonic source modeling, because of the shortcomings of nonlinear modeling for neural-network, the study identified the main modeling method in this paper. Then it introduces the basic principle of least squares support vector machine in nonlinear regression, the key research is the choice of kernel function and optimization methods for parameters of model. After selecting the radial basis function as kernel function, it analyses the impact of two parameters(γandσ2) on model accuracy and proposes methods for parameters optimization, such as grid search based on cross-validation and genetic algorithm. What's more, on the assumption that the supply voltage and the absorbed current are three-phase symmetrical periodic function and the supply voltage does not contain harmonics, it analyses the principles of typical power electronic devices and simulates them by MATLAB. Within a reasonable change in parameter values it obtained learning datas for modeling by MATLAB simulation. These datas include supply voltage size, characteristic parameters of power electronic devices, amplitude and phase angle of AC-side harmonic current. Moreover, learning datas are divided into training set and test set, and normalized to a certain context. on that basis, it establishes models of harmonic sources by radial basis function neural network and least squares support vector machine. In process of parameter optimization, it uses two methods of grid search based on cross-validation and genetic algorithm respectively. Finally, taking into account that power supply voltage is not a pure sine wave in actual power system, in order to make broader range of application for the harmonic model, it simulates the same power electronic devices for getting learning datas with considering the case that the supply volage contains harmonic, and then establishes models by least squares support vector machine based on genetic algorithm. It verifies least squares support vector machine is an effective method for harmonic source model.