Summary: | 碩士 === 中原大學 === 電機工程研究所 === 91 === With rapid developments of technology and wide uses of precise equipment as well as delicate electronic devices, much higher power quality (PQ) is required nowadays, especially the influence of the power transient events. To improve the power quality and find out the causes of the power transients, it needs to monitor the power signals extensively for a long period of time. Results of analyzing and recognizing the monitored signals can therefore be used as references to ameliorate the power quality.
There is many a method to analyze the power signals. Among these methods, Wavelet Transform (WT) approach has the abilities of multi-resolution analysis for both the time and the frequency domains. We may obtain the frequency information for different time points through the time-frequency diagrams using the WT. However, features of plenty of the WT coefficients may vary with occurring points of the PQ events. To reduce the amount of the features representing the power transients and solve the recognizing problems caused by different occurring points of the PQ events, spectrum energies of different scales of WT coefficients are calculated in the thesis. Through the proposed approaches, features of the original power signals can be reserved and not influenced by occurring points of the PQ events.
In order to test the recognizing abilities of the proposed feature extraction method in the classification systems, diverse patterns of PQ events are simulated via the Matlab software tools; besides, practical field data are collected as testing data. The artificial neural networks and fuzzy neural classification systems are used for signal recognition and fuzzy rules construction, respectively. In the mean time, to test noise tolerance of the proposed systems, signals with different degrees of signal-to-noise ratios are also simulated. Success rates of recognizing the PQ events from the noise-riding signals are investigated for feasibility evaluation in the practical applications. The testing results show that the classification systems proposed in the thesis have promising performance in recognizing the PQ events and the tolerance to noise.
|