Artificial Neural Network based Innovative Control Strategies for Active Power Filters

  • Mohammed Qasim

Student thesis: Master's Thesis

Abstract

Power electronic converters, such as diode bridge rectifiers, thyristor controlled rectifiers, and transistor based converters are essential to provide controlled electric power to electrical loads. These power electronic based nonlinear loads draw harmonics and reactive power components of current from the power system supply, which considerably degrade its overall efficiency. With the great advancement of the semiconductor technology, the solid state inverter known as the shunt Active Power Filters (APFs) have become one of the most dominant solutions to mitigate such nonlinear harmonic and reactive power burdens. The vital aspect for the success of the shunt APF operation lies in its control. Since the last decade, Artificial Neural Networks (ANN) have become very attractive in shunt APF control due to their powerful computation nature, fast learning capability, and remarkable achievement in many real-world applications. The most commonly used ANN structures in the shunt APF application are the feed-forward Multi Neural Network (MNN), and Adaptive Linear Neuron (ADALINE). While a lot of work has been done in the area of shunt APF control using MNN and ADALINE, a clear evaluation about each method can hardly be found. The first part of this thesis aims to provide a detailed comparison between the feed-forward MNN and ADALINE. It also provides a clear demonstration about the fundamental concepts of implementing and training the MNN and ADALINE shunt APF controllers. Most of the shunt APF control strategies necessitate the synchronization of the control signals with the supply voltage. For that purpose, researchers often use the Phase-Locked-Loop (PLL). However, the PLL has downsides, such as reduced control robustness, possible instability, and slow convergence at transient conditions. In this thesis, a new PLL-less approach for shunt APF control is proposed. The PLL-less operation is achieved by estimating the fundamental system frequency using the Nonlinear Least Square (NLS) technique, and by generating a phase-locking signal with the aid of ADALINE. The non-linear least squares (NLS) based approach is modified to provide accurate estimation of the supply frequency, irrespective of the distortion level in the supply voltage. To demonstrate the effectiveness of the proposed PLL-less approach, the synchronous reference frame (D Q theory) based PLL-less shunt APF controller is developed and validated experimentally at different supply and load conditions and under a sudden change in the supply frequency. In the final part of this thesis, a complete neural network based control scheme for the shunt APF is proposed to achieve a better dynamic performance. The control structure consists of three ADALINEs. Two ADALINEs are used for frequency estimation and supply voltage synchronization, while the third ADALINE is used to extract the fundamental active component of the load current. The main factor that affects the estimation speed and accuracy is the learning rate involved in the ADALINE training. Generally, this learning rate is selected by trial and error. In this work, the learning rate of each ADALINE is tuned using Particle Swarm Optimization (PSO) to achieve the best dynamic performance. Furthermore, an adaptive learning rate for the frequency-ADALINE is proposed to enhance the estimation speed. The proposed triple ADALINE control structure is validated with a detailed experimental study.
Date of AwardJun 2013
Original languageAmerican English
SupervisorVinod Khadkikar (Supervisor)

Keywords

  • Electric Power Systems; Electric Filters; Active Design and Construction.

Cite this

'