Research On Fault Diagnosis Strategy Of Three-level Wind Power Converter

Currently, energy shortage and environmental pollution have become two majorglobal problems, and wind energy development and utilization is an effectiveapproach to relieve them. As the key equipment in doubly-fed wind power generationsystem, wind power converter assures the generator to emit electric energy meetingthe requirements of power grid. Due to hot and humid, oil and dirt, high voltage andlarge current, the converter is vulnerable to failure. Thus, it is necessary to strengthenthe fault diagnosis research on the wind power converter and to construct the efficientand reliable state monitoring and fault diagnosis system.The main research work in the dissertation includes:(1) By comparison of commonly used time-frequency analysis technology, thewavelet packet decomposition is applied to the fault diagnosis of converter. Theselected wavelet function and scale are adopted to transform the original fault signalsinto eigenvectors acting as training and testing sample for the fault classfier.(2) According to the characteristics of the converter fault data, a sparse LSSVMalgorithm based on boundary neighbor sample is presented. In order to expand themulti-class classification ability, a multi-fault diagnosis method based on Huffmantree is provided. The experimental results show the method advantages in testingspeed and accuracy.(3) The paper proposes an improved multi-population immune co-evolutionparticle swarm optimization algorithm. The algorithm takes full into account of bothcompetition and collaboration between the populations. At the same time, the negativeimmune operator is used to divide the whole population into sub-populations toincrease their diversity and improve the global search ability. Futhermore, twomutation operators are adopted, which can increase the early exploration ability andimprove later exploitation ability. The experiments show that the proposed algorithmcan not only effectively solve problem of lack of local search ability, but alsosignificantly speed up the convergence and improve the stability.(4) Based on the incremental study and region labeling, an improved progressivetransductive semi-supervised LSSVM algorithm is proposed. Experimental resultsdisplay that the algorithm has high generalization capability.