Power quality has become a major concern of both electricity providers and their customers.
For customers, the economic
impact of power disturbances can range from several hundred dollars to repair or replace
home appliances to millions of dollars in product losses, production losses, and cleanup.
For utilities, system disturbances can lead to customer dissatisfaction and load and revenue losses.
Software and hardware for efficient, accurate, and economic power quality monitoring are needed
for both utilities and customers.
Identification and classification of voltage and current disturbances in power systems is
the key task for power quality monitoring and protection. Existing automatic recognition methods need much improvement in terms of their versatility, reliability, and accuracy.
This project develops state of the art signal classification algorithms for classifying different types of power quality disturbances, based on the recent advances in signal processing and pattern recognition techniques. The goals of this project are to enhance the real-time power system protection and accumulate statistics of real power quality data.
The contents of the thesis include an introduction to power quality, development of two new power quality classification algorithms, simulation results, a demonstration software, discussions of several post-processing techniques, explanation of data resources of this project, and discussion of a new classier under exploration - hidden Markov models (HMM).
The emphasis of this thesis is the development of two new classification algorithms. The first method,
which extracts features from disturbance signals by class-dependent TFRs, is a new class of power quality classification approach.
In the proposed algorithm, the concept of class-dependent time-frequency representations (TFR) is used for feature extraction
and a feedforward neural network is chosen as the classifier. The basic idea is to design a TFR
for our specific classification task. The TFR needs to be class-dependent, instead of signal-dependent. The
design of this TFR starts from the signal's time-frequency ambiguity plane. A class-dependent kernel is
designed by a large set of training examples based on the Fisher's Discriminant kernel. The desired TFR
is obtained by array multiplying the designed class-dependent kernel with the time-frequency ambiguity plane.
Extensive testing with simulated signals confirms the feasibility, high recognition rate, and
computation efficiency of the proposed algorithm. A six-class classification simulation
gives us the following recognition rates: 100% for harmonics, 100% for capacitor high
frequency switching transients, 100% for voltage swell, 94% for capacitor low frequency switching
transients, 92% for voltage sudden sag, and 93% for voltage sag decay.
This approach is a new class method for discriminating PQ events.
One of the major characteristics of this algorithm is the utilization the class-dependent
time-frequency representations, which are designed solely for our classification task.
Therefore, there is a high
potential of adapting this methology for discriminating more sophisticated types of PQ events.
An algorithm demonstration software
is developed using Matlab GUI (Graphic User Interface) tools.
This new method and the algorithm demonstration software show promise for
further development of a fully automated and highly efficient real-time power quality monitoring
system with a high recognition rate. Currently, the creation of a power quality database is in progress, through our own signal recording and simulations
and several developing industrial partnerships. Modification and explorations on the classification algorithms
are being done to adapt these advanced techniques onto real power system data and the cause-effect identifications
of power quality problems, which will bring up significant help to utility companies.
At the same time, thrust towards hardware
implementation of these new signal processing algorithms is being considered.