Harmonic elimination using a parallel active power filter adapted with artificial neural networks in MATLAB/SIMULINK environment

Delivering electricity to consumers in a high-quality and uninterrupted manner has been a fundamental goal for energy producers. However, the widespread use of power electronics elements and their non-linear load characteristics lead to the formation of harmonics in electrical systems. Harmonics are defined as distortions in current and voltage waveforms. These distortions originate from elements such as transformers, generators, arc furnaces, uninterruptible power supplies, static VAr compensators, computers, and converters. The adverse effects of harmonics can be summarized as energy loss, overloading of system components, resonance events, and early failure of devices. These issues, particularly in industrial applications, result in significant costs and efficiency losses. Therefore, it is necessary to eliminate harmonics and bring them within permissible standards. The first step in suppressing harmonics is to take precautions during the design phase. However, these precautions are often insufficient due to unpredictable system conditions. As a result, harmonics that form later can be eliminated using filters. Harmonic elimination is performed through various filtering methods, primarily categorized into two types: passive and active filters. The primary principle of passive filters is not to eliminate harmonics but to provide a path for harmonics to flow to the ground line. The primary principle of active filters, on the other hand, is to detect harmonics in the system and generate a wave with the same frequency and amplitude but in the opposite direction to cancel out the harmonics. As can be understood from these principles, passive filters are only effective under the specific conditions for which they are designed, while active filters can dynamically respond to varying harmonic waves under changing conditions. Both methods have their areas of application, and the most appropriate type of filter is determined by evaluating criteria such as system conditions, ease of application, cost, and control mechanisms. Passive filters are created by combining basic circuit elements such as resistors (R), inductors (L), and capacitors (C) in various configurations. These filters do not eliminate harmonics but merely redirect harmonic currents, preventing them from reaching undesired locations. Their working principle involves determining L and C values that will resonate at the harmonic frequency to be suppressed. Passive filters are advantageous due to their ease of application, practicality, and low cost. However, their disadvantages include large size, lack of adaptability to changing load conditions, and potential resonance with the grid or load. Active filters generate a wave with the same frequency and amplitude but in the opposite direction (180 degrees) to the harmonic component in the current or voltage waveform using switching elements, thereby eliminating the harmonic component. In addition to harmonic elimination, they also provide reactive power compensation and neutral current compensation, addressing power quality problems. They are generally designed as three-phase three-wire or three-phase four-wire systems. However, their high cost and design complexity are significant disadvantages. It is also possible to use passive and active filters together in filtering. A method in which series active filters and parallel active filters are combined is called a unified power quality conditioner. This type of filter combines the voltage-side compensation and grid harmonic isolation features of series active filters with the harmonic current and reactive power compensation features of parallel active filters, maintaining the current and voltage waveforms as perfect sine waves. This comprehensive capability addresses almost all issues that may arise in power systems. However, it is highly costly and has a complex design. The combination of an active filter with a passive filter is referred to as a hybrid filter. This method is more commonly preferred in the market. Especially, the series active filter-parallel passive filter combination stands out with its low cost and high performance. Active filters respond dynamically to the system using switching elements. These switching elements require signals for their on-off states. These signals are generated using techniques such as PWM, sliding mode, fuzzy logic-based control, and hysteresis band control. These techniques require reference signals. At this point, generating reference signals becomes one of the most critical aspects of active filter design. Techniques such as frequency domain methods, time domain methods, and instantaneous reactive power theory are used to generate these signals. Each method has advantages in specific scenarios, and the appropriate method can be selected based on system requirements. In this study, a design based on the instantaneous reactive power theory was implemented. The fundamental principle of this theory involves converting three-phase voltages and currents into a two-phase stationary αβ structure for calculations, then obtaining reference signals through reverse transformation. Filtering methods developed to effectively suppress harmonics play a critical role in improving energy quality. Parallel active power filters, which stand out as a dynamic and effective method, have become a focus of researchers to explore whether they can be simplified with artificial neural network (ANN)-supported solutions. ANNs learn from examples and do not rely on specific algorithms. Their primary principle is to learn the relationship between inputs and outputs from examples provided to the network and produce outputs for new inputs. Their advantages include not requiring mathematical algorithms, possessing learning capabilities, solving problems that are difficult or impossible to model, needing only examples for modeling, effectively addressing non-linear problems, quickly reaching results, and being easily retrained after system changes. These features have led to the rapid development and attention of ANNs. This study examined the design of a parallel active power filter integrated with an ANN and presented it as an effective method. Different ANN models were designed and simulated in Matlab/Simulink. The aim of the study was to train an ANN to replace the filter. Thus, the filter inputs source voltages Vsa, Vsb, Vsc , load currents iya, iyb, iyc and ploss were also used as inputs for the ANN. The filter outputs, reference filter currents i*fa, i*fb, i*fc served as ANN outputs. The ANN was designed with seven inputs and three outputs. Data was collected from 60 harmonic waves, with 200,000 samples taken from a single harmonic wave. The output values in the dataset ranged between -50 and +50. Given the presence of negative values, normalizing the data between -1 and +1 was a logical approach for the ANN. Consequently, the activation function was chosen as tangent sigmoid. The training algorithms used were Levenberg-Marquardt (LM) and scaled conjugate gradient (SCG), primarily due to the dataset's size and the predictive nature of the problem. For ANN training, 80% of the dataset was allocated for training, and 20% for testing. The learning rate was set at 0.001. During training, both the number of hidden layers and neurons were adjusted through trial and error, and results were compared using MAE, MSE, and RMSE as performance criteria. The first scenario involved one hidden layer with five neurons and the SCG training algorithm. The second scenario used one hidden layer with five neurons and the LM training algorithm. The third scenario employed two hidden layers with 20 neurons each and the SCG algorithm. For the RNN model, an LSTM-based architecture was used, with 200 hidden layers, a learning rate of 0.001, and 100 epochs. For the unfiltered model, the THD value was 29.84%. Using the p-q method with the parallel active power filter, the THD value was reduced to 3.58%, indicating a significant improvement. However, ANN-based parallel active power filters yielded even better results. For the Elman network, the THD values were 3.03%, 3.32%, and 2.56% for the first, second, and third scenarios, respectively. For the MLP network, the THD values were 3.23%, 3.53%, and 2.88% for the respective scenarios. For the LSTM-based RNN, the THD value was 3.46%. The results demonstrated that the best performance was achieved with the Elman neural network model with two hidden layers. This model provided a dynamic control mechanism for effectively suppressing harmonics and improved the adaptability of energy systems to various conditions. Additionally, the SCG algorithm showed faster and more effective training performance compared to the LM algorithm. ANN-based systems are expected to be widely used in future energy system designs for harmonic elimination. This thesis has demonstrated the effectiveness of ANN-based harmonic control solutions and their applicability in energy systems. It is recommended that future studies combine different artificial intelligence models and optimization techniques to develop new approaches for harmonic control. These methods are expected to enhance energy efficiency and optimize costs.