Distributed Generation System: Loss Detection of Grid Events Via Pattern Identification
The purpose of this paper is to discuss the different types of pattern identification methods that are commonly used for loss detection of grid events in distributed generation (DG) systems. The research paper is divided into four parts; (1) introduction, (2) background, (3) literature review, and (4) conclusion. The introduction provides an overall overview of the topic, identifying reasons why pattern identification methods are important in recognizing islanding events. The second section of the paper highlights detailed analysis of distributed generation systems and the risks that might arise when islanding is not detected. The literature review analyzes three major pattern identification artificial neural networks, decision tree classifier, and adaptive neuro fuzzy inference system. These three systems use machine learning to train the systems through algorithms to identify islanding and non-islanding system. The fourth section is a generalized summary of the entire paper.
Keywords: loss detection, grid events, distributed generation (DG) systems, pattern identification methods, islanding, artificial neural networks, decision tree classifier, and adaptive neuro fuzzy inference system.
Traditional concept for the electric power grid relies on large, centralized power facilities to transmit electricity across great distances to individual homes and businesses. Due to its reliance on a small number of key nodes, the grid is easily disrupted by events beyond its control, such as transmission line or plant outages, severe weather, and consequent cascade failures and widespread blackouts (Guha 1). The global growth in energy demand has put an extreme pressure on the power system. This is due to the high costs associated with upgrading transmission infrastructure and the economic and environmental restrictions placed on the development of new power plants. Because of the drawbacks of the centralized approach, such as high transmission losses and high installation costs, the distributed power generating model is increasingly being pushed as a viable alternative. Distributed generation (DG) units (independent generators that provide electricity into the distribution network) are the building blocks of this architecture. This paper discusses the three main types of pattern identification methods including artificial neural networks, decision tree classifier, and adaptive neuro fuzzy inference system; which are used in the detection of grid events in a distributed generation system.
Decentralized electricity grids today are likely to include Distributed Generation (DG) sources. However, there may be a number of operational difficulties associated with connecting DG systems to the electricity grid. Islanding detection is one such challenges. Laghari et al. argue that islanding occurs when a DG system loses contact with the rest of the power grid (140). Due to the potentially catastrophic consequences of islanding, DG interconnection standards like IEEE 1547 and UL 1741 require the use of a reliable and rapid islanding detection method (Cebollero et al. 2). Passive monitoring has mainly been used in the monitoring of energy frequency and voltage in conventional islanding techniques.
Figure 1: A typical islanding scenario of a DG system.
Numerous risks arise as a result of islanding for an extended period of time. If islanding is not recognized and remedial action is not performed by the DG protection system, it is likely that power is being mistakenly given back by these DG sources. When power lines that have been disconnected from the grid are still electrified by a nearby DG source, for instance, this poses a risk to field engineers and maintenance employees. Most current DG systems are privately held; therefore, they are mostly out of the hands of power companies. When electricity is cut off to a specific location, known as “islanding,” the DG voltage in that region will no longer be in phase with the rest of the power grid. In this state, reclosing might cause strong currents and torques to enter the electrical system, which could cause serious damage (Noradin & Sobhani 1425). If the DG source continues to feed the fault in islanding mode, automatic reclosing may not work after faults. A lengthy period of grid disconnection would occur if the problem had to be cleared manually by reclosing the circuit. Thus, islanding detection methods such as pattern identification are required to ensure that a DG source is automatically detected and disconnected from the primary grid.
Types of Pattern Identification Techniques
Machine learning strategies employ learnable, adaptive algorithms to determine if an event is islanding or not. These techniques use input and output samples to train and refine classification models. In order to spot islands, algorithms have been described that take into account both the target class and a massive collection of input features (islanding or non-islanding) (Haddad et al. 19). The three main types of pattern identification methods that are commonly used include; (1) Artificial Neural Networks, (2) Decision Tree Classifier, and (3) Adaptive Neuro Fuzzy Inference System.
Artificial Neural Networks. It is common practice to use Artificial Neural Networks (ANNs) for function fitting, pattern identification, signal forecasting. They are classified as a subset of machine learning algorithms whose reasoning is heavily influenced by the concept of the human brain. These characteristics have found use in the field of power systems, where they have been used for fault classification, power system stabilizer design, voltage stability analysis. Guha (19) postulate that the building blocks of an ANN are a series of layers—an input layer, one or more hidden layers, and an output layer—that are all intricately coupled to one another. Hidden layers are called neurons, and the number of neurons required for a certain task is determined by the designer. As the number of neurons in a system grows, so does its complexity and the length of time it takes to train.
Multiple inverter-based DG and hybrid inverter-based DG have both been proposed to benefit from ANN-based islanding detection algorithms. Voltage and current transients in three-phase systems are used to identify islanding, and both methods fall under the umbrella term “passive islanding detection approach” because they do not degrade power quality in any way. Islanding detection data is gleaned using a discrete wavelet transform in this method. This data is then used to educate an ANN to recognize the difference between an islanding event and other disruptions. According to Haddad (20), discriminative aspects of voltage signals on the DG side are recorded and simulated islanding and non-islanding events are used to train the ANN. Common non-islanding cases in an ANN pattern identification system include line-to-line fault, load increase and decrease, and single line to ground fault among others.
Decision Tree Classifier. The islanding detection capabilities of the Decision Tree (DT) method have been thoroughly examined over the years. One form of pattern identification tool is the Decision Tree classifier, which attempts to find a good answer for all potential inputs by factoring in their statistical variances. A decision tree (DT) is a type of sequential flowchart in which each node performs a threshold comparison on an input variable (Guha 19). After comparing the current values of the system parameters to the thresholds, the tree eventually settles on an event classification. The information you provide is used to build the DT and choose the threshold. The leaves, or terminal nodes, of the tree reflect the categorization of the event, while the trunk represents the result of comparing the input parameter with the threshold.
In comparison to other pattern identification methods, DT’s key benefit is its rapid training time. The decision tree method may combine several smaller judgments into a single one, simplifying a previously incomprehensible decision-making process (Guha 21). The initial branch of the decision tree treats all of space as if it were a single node. As a first step, we use a predictor variable to divide the initial node into two branches. These child nodes are selected from the whole set of child nodes and only include the most pristine information. The offspring nodes can be used to create forks. No further branching occurs at a leaf (terminal) node. The structure of leaf nodes is used to create predictions. In order to use a decision tree for prediction, one must follow the branches down to a leaf node. Whenever an island is detected, the node linked to the island is automatically disconnected from the rest of the tree (power grid).
Figure 2: The basic structure of a Decision Tree Classifier.
Adaptive Neuro Fuzzy Inference System (ANFIS). Many nonlinear classification issues have been resolved with the help of artificial intelligence techniques like neural networks and Fuzzy logic (FL) inference. According to Laghari et al. (147), a Fuzzy logic system (FLS) principal benefits come from its ability to represent nonlinear input/output interactions using a collection of qualitative if-then rules. The ANFIS is an effective method for modeling nonlinear and complicated systems with little input and output training data, while nevertheless achieving rapid learning and high accuracy (Laghari et al. 147). ANFIS is a dynamic network that integrates FL and neural network architecture. It incorporates the best features of both approaches while removing those that only apply when just one is employed.
Typically, ANFIS training may make use of several algorithms to lessen the training error. The data set is split into a training set and a test set before a FIS is trained. The test set is used to identify when training should be ended to avoid over fitting, while the training set is used to train a fuzzy mode (Noradin & Sobhani 1428). With this combination, an ANFIS system is able to get around the difficulty of determining appropriate detection thresholds for islanding detection. The islanding condition is the expected outcome of the ANFIS model, which uses voltage as inputs. If islanding is found, the ANFIS value at output will be greater than the threshold value that has been used. On the other hand, if islanding is not identified, the result ANFIS will be in the range below the threshold value. This makes the ANFIS method the most accurate pattern identification technique for identifying islanding.
Islanding can have severe consequences on a distributed generation system, hence the need for having reliable and accurate pattern identification systems for identifying grid events. Some of the potential risks of islanding events include disconnection of the entire grid system as well as risking the lives of maintenance employees and field engineers. There commonly used types of pattern identification methods include artificial neural networks, decision tree classifier, and adaptive neuro fuzzy inference system. Each of these methods uses machine learning to train the system to be able to detect islanding and non-islanding events in a grid system. Out of the three, the adaptive neuro fuzzy inference system is the most effective method.
Cebollero, José Antonio, et al. “A Survey of Islanding Detection Methods for Microgrids and Assessment of Non-Detection Zones in Comparison with Grid Codes.” Energies 15.2 (2022): 460.
Guha, Bikiran. “Smart Distributed Generation Systems Using Improved Islanding Detection and Event Classification.” Electronic Theses and Dissertations 1262 (2015).
Haddad, Rami J., et al. “Smart distributed generation systems using artificial neural network-based event classification.” IEEE Power and energy technology systems journal 5.2 (2018): 18-26.
Laghari, J. A., et al. “Computational Intelligence based techniques for islanding detection of distributed generation in distribution network: A review.” Energy conversion and Management 88 (2014): 139-152.
Noradin, Ghadimi, and Sobhani Behrooz. “Adaptive neuro-fuzzy inference system (ANFIS) islanding detection based on wind turbine simulator.” International journal of physical sciences 8.27 (2013): 1424-1436.