LEE- Elektrik Mühendisliği Lisansüstü Programı
Bu topluluk için Kalıcı Uri
Gözat
Konu "Digital signal processing" ile LEE- Elektrik Mühendisliği Lisansüstü Programı'a göz atma
Sayfa başına sonuç
Sıralama Seçenekleri
-
ÖgeAdaptive signal processing based intelligent method for fault detection and classification in microgrids(Lisansüstü Eğitim Enstitüsü, 2021) Azizi, Resul ; Şeker, Şahin Serhat ; 724566 ; Elektrik MühendisliğiThe ever-increasing energy demand, the environmental issue of fossil fuels and the high investment cost for the establishment of bulk power plants lead energy plans to more flexible and scattered small-scale energy sources. The main feature of these new topologies is that they consume renewable energy sources for electricity generation. It also requires less time to plan, build and operate. Moreover, they are close to energy sources and local loads. So, there are more efficient, with minimal environmental issues. However, besides their benefits and advantages, they pose a new challenge for traditional power systems. These challenges include protection issues, stability concerns, and complex control systems and so on. Traditional power systems include mass generation followed by transmission and distribution. In this topology, it is possible to plan generation because consumption at the transmission level of the power system is more predictable and fuel resources are always available for generation units. On the other hand, the transmission system and its conditions can be controlled by state estimators and SCADA system. Therefore, production and consumption uncertainties are minimal and conventional protection is sufficient to protect these systems. Also, distribution systems have no generating units, systems are mostly radial and overcurrent protection systems are sufficient to protect them. In these passive networks, it is not necessary to have fast and reliable protection systems as in transmission systems. The initial role of these new energy sources was to act as a backup for mass production and to eliminate the small generation and consumption mismatch during peak consumption. On the other side, huge demand growth and investment time of mass production units and environmental concerns make these distributed energy resources (DERs) (wind, solar, biomass, etc.) popular in the distribution system. However, the contribution of the early DER groups to the total production is low and the control systems are very sensitive to voltage disturbances such as faults. Thus, according to the grid codes, after any minor fault or disturbance in the system, the DERs are disconnected, synchronized manually and reconnected after the fault is cleared. With the increasing penetration of DERs in distribution systems, they play an important and rapidly increasing role in the total production of the system. Therefore, de-energizing all these DERs in an area in the distribution system after a fault has occurred can lead to stability problems due to generation and consumption imbalance. Accordingly, a new concept called microgrid emerged and mainly established in distribution systems. This topology is the microscale of the power system. It can operate autonomously and cover the total demand of this local distribution system. Like the SCADA power system, it has an equivalent centralized monitoring and control system. The total generation is almost sufficient for the total demand of the loads in distribution networks converted to microgrid. It can operate as a standalone ecosystem separated from the main grid and is self-sufficient. The basic requirement of this topology for connecting to the main grid through PCC (point of common coupling) is to increase the total inertia of the system and increase the post-fault stability region. In addition, this topology can transfer energy to the main system if it produces more power than the loads consume. This can reduce the stress of mass production units. Last but not least, if the main upper grid disturbed, the microgrid can continue to supply its loads by disconnecting from the grid. In this new concept, grid codes expect the micro grid to be able to ride through faults and disturbances thanks to low voltage ride through (LVRT) systems. In fact, as a micro-scale model of the power system, the voltage of the DERs at the time of fault occurance is controlled by the LVRT, and the DERs continue to operate without disconnection after the fault is cleared by circuit breakers or other elements). Therefore, more complex control systems are required for DERs. However, microgrids are distribution systems and unlike traditional power systems, there is a high amount of uncertainty in generation and consumption (loads). The distribution system has changed from a passive network to an active dynamic network. In this system, topology, generation and consumption are changed faster and faster than in conventional power systems. This situation constantly changes the fault current level and direction, and the conventional overcurrent protection is completely insufficient to protect them. Also, due to the high penetration of sensitive DERs, prolonged fault current is not allowed (stability concerns). Moreover, inverter-based DERs have a very small contribution to the fault current level. The current protection method of microgrids is adaptive protection. In this model, all operating conditions of the system are extracted and all components of the systems are continuously monitored by central or decentralized control system or even dynamic load estimation. This model cannot be applied to a central control system because it has to process large amounts of data at a high sampling rate and it is impossible to make real-time decisions. Based on these facts, a new intelligence-based method for fault detection and classification of microgrid is proposed in this thesis. In the proposed method, three different adaptive signal processing methods are used to extract the short-time transient component of the signal instead of the fault current level. It transfers data (feature extraction) into three different data spaces. The main feature of these signal processing methods is that they do not use a predefined basis to decompose a signal. The basis is adaptive to signal and extract components depend on the noise penetration level and frequency components of the signal. An intelligence-based method called Brwonboost is used to make decisions in these data spaces, and the total decision is taken by the majority of votes of these three intelligence-based methods in these three data spaces. The main unique feature of the proposed method compared to traditional machine learning methods is its adaptability and uses a non-convex optimization method for detection and classification. The proposed method is a set of weak classifiers and tries to learn the data space step by step and iteratively. It tries to adapt the data by classifying the data that was misclassified in previous iterations. On the other hand, the unique non-convex optimization feature of the proposed method gives it an intelligence to select or discard misclassified data. It can decide step-by-step removal of the algorithm's iteration data in the training process if there is an outlier or a violation in another class area. This feature provides evidence against overfitting and becomes as practical a method as it is for real-world measured data. Finally, a Brownboost decision is also made by a majority vote of the weak classifiers. An intelligence-based method called Brwonboost is used to make decisions in these data spaces, and the total decision is taken by the majority of votes of these three intelligence-based methods in these three data spaces. In this method the classifier works base on the margin. This means, instead of only finding a classifier that minimize the classification error, it selects a classifier that has maximum discrimination between data of every class. The unique feature of the proposed method compared to traditional machine learning methods is its adaptability and uses a non- convex optimization method for detection and classification. The proposed method is an ensemble of weak classifiers and tries to learn the data space step by step and iteratively. It tries to adapt to the data by classifying the data that was misclassified in previous iterations. On the other hand, the unique non-convex optimization feature of the proposed method gives it an intelligence to select or discard misclassified data. During this step-by-step process, the algorithm can detect outliers or misclassified data that intensely violated other class area and remove it. This feature makes it robust against overfitting and becomes as practical method for real-world measured data. In total, the proposed method tries to classify the data in three different data spaces. The data area that makes maximum distinction between the data of each class is less sensitive to noise. Thus, a classifier has are fewer generalization errors to unseen new data (higher margin). Therefore, its Brownboost has more voting power in decision making. The results are test in test benchmark microgrid. DERs are modeled with the detailed model to extract the true detail form of the signal. Various types of control model and fault ride thruogh feature of DERs are implemented.