الفهرس 
introductionOnly 14 pages are availabe for public view
 |  | from | 176 |  |  |
| | | from | 176 | | |
Abstract
In this thesis, a Long Short-Term Memory (LSTM) network is proposed to detect and identify the power quality problems such as sag, swell, interruption, surge, flicker, harmonic distortion, sag with harmonics and swell with harmonics. The LSTM technique is a special type of Recurrent Neural Networks (RNNs), which can detect and classify simple and complicated PQ problems depending on automatic feature extraction. Hidden units in RNNs receive feedback from the previous states to current states. These features are suitable to handle the temporal-spatial data.
LSTM network can deal with the correlation problems of time series in both short and long duration of time, using the hidden layer as a memory unit. LSTM network is used to develop a more complex and deeper neural network that is nonlinear in nature, which is capable of demonstrating the effect of long-term memory.The LSTM network system is designed and implemented to perform the detection and classification process with five parameters as inputs: amplitude, start time, end time, duration and THD% and two outputs: waveform class and harmonics indication.
Several simulations of PQ problems areused to test the proposed technique.Allsimulated signals are mixed with the random white noise of different values of the signal to noise ratio (SNR): 40 dB, 30 dB and 20dB.The simulated signals are for normal voltage, sag, swell, interruption, surge, flicker, harmonic distortion, sag with harmonics and swell with harmonics for noiseless and a noisy environment.The simulation results of the LSTM technique show that the mean classification accuracies of 100 signals without noise and withSNR 20 dB, 30 dB and 40 dB are 100%, 94.56%, 98.22% and 99.94%, respectively. These simulation studies show that the proposed LSTM network classifier is efficient for the classification and identification of power quality problems.
To confirm the efficiency of the LSTM technique, its classification results are compared with other methods such as ADALINE and FFNN, ST and fuzzy decision tree, ST and dynamics, DRST andDAG-SVMs, KFand FES and SSD on hybrid dictionaries. The comparisons show that the proposed technique can detect and classify all simulated PQ problems, but other methods did not classify all PQ problems.
In addition, many sets of real data: 3ph voltages and currents are used to ensure the capability of the proposed technique to identify and detect the PQ problems successfully. The real data are measured and collected from HV/MV substation by using data acquisition device. The practical results summaries the LSTM’s inputs: magnitude, start time, end time, duration and THD% and outputs: waveform class and harmonics indicationand PQ events in the measured waveform.
The simulation and the real results imply that the LSTM technique is a good choice for future PQ analysis without needing human experts if larger data is available. Finally, the results show that the proposed technique is capable to accurately detect and identify PQ problems. This makes a new probability of developing deep learning technology for PQ data identification and detection.
In this thesis, a Long Short-Term Memory (LSTM) network is proposed to detect and identify the power quality problems such as sag, swell, interruption, surge, flicker, harmonic distortion, sag with harmonics and swell with harmonics. The LSTM technique is a special type of Recurrent Neural Networks (RNNs), which can detect and classify simple and complicated PQ problems depending on automatic feature extraction. Hidden units in RNNs receive feedback from the previous states to current states. These features are suitable to handle the temporal-spatial data.
LSTM network can deal with the correlation problems of time series in both short and long duration of time, using the hidden layer as a memory unit. LSTM network is used to develop a more complex and deeper neural network that is nonlinear in nature, which is capable of demonstrating the effect of long-term memory.The LSTM network system is designed and implemented to perform the detection and classification process with five parameters as inputs: amplitude, start time, end time, duration and THD% and two outputs: waveform class and harmonics indication.
Several simulations of PQ problems areused to test the proposed technique.Allsimulated signals are mixed with the random white noise of different values of the signal to noise ratio (SNR): 40 dB, 30 dB and 20dB.The simulated signals are for normal voltage, sag, swell, interruption, surge, flicker, harmonic distortion, sag with harmonics and swell with harmonics for noiseless and a noisy environment.The simulation results of the LSTM technique show that the mean classification accuracies of 100 signals without noise and withSNR 20 dB, 30 dB and 40 dB are 100%, 94.56%, 98.22% and 99.94%, respectively. These simulation studies show that the proposed LSTM network classifier is efficient for the classification and identification of power quality problems.
To confirm the efficiency of the LSTM technique, its classification results are compared with other methods such as ADALINE and FFNN, ST and fuzzy decision tree, ST and dynamics, DRST andDAG-SVMs, KFand FES and SSD on hybrid dictionaries. The comparisons show that the proposed technique can detect and classify all simulated PQ problems, but other methods did not classify all PQ problems.
In addition, many sets of real data: 3ph voltages and currents are used to ensure the capability of the proposed technique to identify and detect the PQ problems successfully. The real data are measured and collected from HV/MV substation by using data acquisition device. The practical results summaries the LSTM’s inputs: magnitude, start time, end time, duration and THD% and outputs: waveform class and harmonics indicationand PQ events in the measured waveform.
The simulation and the real results imply that the LSTM technique is a good choice for future PQ analysis without needing human experts if larger data is available. Finally, the results show that the proposed technique is capable to accurately detect and identify PQ problems. This makes a new probability of developing deep learning technology for PQ data identification and detection.