AN ABSTRACT OF THE THESIS OF

Transcription

1

2 AN ABSTRACT OF THE THESIS OF Kelly Tray for the degree of Master of Science in Electrical and Computer Engineering presented on June 6, Title: Dynamic Composite Load Signature Detection and Classification using Supervised Learning over Disturbance Data Abstract approved: Ted K.A. Brekken Load modeling that can accurately represent the dynamic behavior of generators and loads is important in the operation and planning of transmission and distribution systems. Yet, it is a complex subject in power system research communities and electric utilities. The composition of the end-use loads is changing continually based on climate zone, season, and time. The WECC composite load model has been developed recently to better represent Fault Induced Delayed Voltage Recovery (FIDVR) events, which is caused by air-conditioning stalling phenomena. The approach is based on using the information of the load class at the substation level and composition of air-conditioning, induction machines, power electronics, and static loads associated with the load class. Therefore, it is important to be able to identify and classify the load class. This can be accomplished by using machine learning based signature detection since each load class has a unique signature response due to a particular disturbance in the system. The objective of this project is to implement a supervised learning, Artificial Neural Network (ANN), algorithm to detect and classify the composite load signatures in terms

3 of residential, commercial, agriculture, and mixed load class. Furthermore, the process of creating WECC composite load model data, using the Load Model Data Tool (LMDT), to be used in time-domain dynamic simulation (PSS/E) is demonstrated. The One-Area Reliability Test System is used for the purpose of demonstration and validation of our proposed methodology. In term of classification accuracy, the classifier gives about 87 percent using standardization for data normalization and Principle Component Analysis (PCA) for feature reduction.

4 c Copyright by Kelly Tray June 6, 2017 All Rights Reserved

5 Dynamic Composite Load Signature Detection and Classification using Supervised Learning over Disturbance Data by Kelly Tray A THESIS submitted to Oregon State University in partial fulfillment of the requirements for the degree of Master of Science Presented June 6, 2017 Commencement June 2017

6 Master of Science thesis of Kelly Tray presented on June 6, APPROVED: Major Professor, representing Electrical and Computer Engineering Director of the School of Electrical Engineering and Computer Science Dean of the Graduate School I understand that my thesis will become part of the permanent collection of Oregon State University libraries. My signature below authorizes release of my thesis to any reader upon request. Kelly Tray, Author

7 ACKNOWLEDGEMENTS I would like to express my sincere appreciation to a number of people who have helped and supported me during these past two years in Graduate School. First and foremost, I would like to express my sincere gratitude to my major advisor, Dr. Ted K.A. Brekken, for providing me the tremendous opportunity to be part of the energy system group. Also, I would like to thank him for his unconditional belief in me, his patience and kind guidance to make this research successful. His passion, technical expertise, and optimistic personality are always inspiring. I also would like to thank Dr. Eduardo Cottilla-Sanchez, Dr. Julia Zhang, and Dr. Annet von Jouanne for their inspirational teaching, continued supporting, and belief in me since undergraduate school. I really appreciate the time and effort they have taken to answer my questions and to give me advice. I also would like to express my appreciation to Dr. Yun-Shik Lee for taking time to serve as my GCR. Also, I would like to thank my fellow graduate students in energy systems for their friendship, help, and support. It has been a great privilege to work with and learn from all of them. Lastly, I would like to convey special thank you to my wonderfully supportive parents, brothers, and fiance who always believe in me and continue encourage me both inside and outside academia. More importantly, I would like to thank my better half, Alex Dziggel, for his unconditional love, constant motivation, and continual encouragement over the course of my study and research.

8 TABLE OF CONTENTS Page 1 Introduction Research Motivation Research Contribution Organization of this Thesis Background Fundamentals of Machine Learning Techniques Power System Model Data Composite Load Model Overview Structure Methodology Composite Load Model Development Composite Load Model Data Acquisition Load Model Data Tool Time-Domain Dynamic Simulation Setup Dynamic Power System Models Case Scenarios Automated Disturbance Data Collection Python based PSS/E Automation Python based Data Collection Detection and Classification Algorithms Multi-Layer Perceptron Theoretical Framework Load Signature Classification using Multi-Layer Perceptron Results and Discussion Dynamic Simulation Results Classification Results Conclusion and Future Work 50 Bibliography 50

9 TABLE OF CONTENTS (Continued) Page Appendices 54 A Default WECC Composite Load Model Data B Load Model Creation Process using LMDT 1.0 for PSS/E Dynamic Record 61 C Time-Domain Dynamic Simulation

10 LIST OF FIGURES Figure Page 2.1 A complete dynamic power system dataset The observed voltage recovery in Southern California, Simulation result using air-conditioning model data and a static model Simulation result using the existing composite load model, which consists of 20 percent of induction motor and 80 percent static load Composite Load Model Structure Equivalent circuit for three-phase induction motor model Methodology Structure LID Region WECC Climate Area Feeder types Load profile for residential load class at 4 PM peak load in the Northwest Valley climate zone Load fraction for residential load class at 4 PM peak load in the Northwest Valley climate zone Load profiles for agricultural load class at 4 PM peak load in the Northwest Valley climate zone Load fraction for agricultural load class at 4 PM peak load in the Northwest Valley climate zone Load Model Data Tool Bus Power System - IEEE One-Area RTS Single-line-ground and line-to-line fault locations Automation scheme for dynamic simulation Automated data collection. This automated scheme writes the output files generated from the dynamic simulation to a CSV file

11 LIST OF FIGURES (Continued) Figure Page 3.14 Single-layer neural network structure The gradient-based learning machine Multi-layer Perceptron Neural Network Structure Voltage signatures measured at Bus2 for single-line-to-ground fault at Bus Voltage signatures measured at Bus15 for single-line-to-ground fault at Bus Power signatures measured at Bus2 for single-line-to-ground fault at Bus Power signatures measured at Bus15 for single-line-to-ground fault at Bus Frequency signatures measured at Bus2 for single-line-to-ground fault at Bus Frequency signatures measured at Bus15 for single-line-to-ground fault at Bus Voltage signatures measured at Bus2 for line-to-line fault on the transmission line connected between Bus2 and Bus Voltage signatures measured at Bus15 for line-to-line fault on the transmission line connected between Bus2 and Bus Real power signatures measured at Bus2 for line-to-line fault on the transmission line connected between Bus2 and Bus Real power signatures measured at Bus15 for line-to-line fault on the transmission line connected between Bus2 and Bus Frequency signatures measured at Bus2 for line-to-line fault on the transmission line connected between Bus2 and Bus Frequency signatures measured at Bus15 for line-to-line fault on the transmission line connected between Bus2 and Bus A close-up of the voltage recovery (voltage signatures) due to line-to-line fault on the transmission line connected between Bus2 and Bus

12 Figure LIST OF FIGURES (Continued) Page 4.14 Normalized confusion matrix for all four load classes. The overall classification accuracy is if both standardization and PCA techniques are using for data preprocessing Normalized confusion matrix for all four load classes. The overall classification accuracy is if both scaling (0-1) and PCA techniques are using for data preprocessing Normalized confusion matrix for all four load classes. The overall classification accuracy is if only standardization technique is using for data preprocessing Normalized confusion matrix for all four load classes. The overall classification accuracy is if only scaling (0-1) technique is using for data preprocessing Normalized confusion matrix for all four load classes. The overall classification accuracy is if 80% of raw data is used for training and 20% of raw data is used for testing

13 LIST OF TABLES Table Page 3.1 Load composition data for Northwest Valley climate zone Load composition data for Northwest Valley MLP Classification Accuracy

14 Chapter 1: Introduction 1.1 Research Motivation The mission of power systems is to have sufficient and uninterruptible power. However, interruptions in power system operations are not always avoidable. In order to achieve better operating efficiency and avoid service interruptions, the utilities find it increasingly necessary to better monitoring, analysis, and control of their power grids [1]. There has been a lot of research and development in load modeling, energy management, as well as the deployment of intelligent electronic devices such as Phasor Measurement Units (PMUs), Supervisory Control And Data Acquisition (SCADA), and smart meters for real-time monitoring of the systems. While those technologies are important, obtaining accurate information of the load classes at the substation level is also important for grid planning, load modeling, and energy management in power systems. The grid planner needs detailed information for the load class in order to determine the composite load model data appropriately. The load modeling developer needs to estimate the load composition based on the load class for implementing a more detailed dynamic load model. For the energy management system, on the other hand, having up-to-date load information can help for more effective load management strategies to alleviate the system at peak demand. The conventional methods used to determine the load class information are energy consumption surveys, smart meter data, or consumer billing data for clustering the load classes [2 4]. However, obtaining billing data or energy consumption data is a time-

15 2 consuming process and requires human-written programs. Additionally, installing smart meters at every customer site to ensure observability of the distribution system is not technically or economically possible [1, 3]. The current deployment of synchrophasor technology, by installing PMUs at the substations, provides adequate and nearly realtime measurements of the power grid. This measurement data is called synchrophasor data. Thanks to PMUs, the load classification is more feasible within an acceptable level of accuracy without measuring real-power consumption from all meters. Utilizing synchrophasor data for load classification can provide up-to-date load class information and can give more accurate data for load modeling purposes. Moreover, the load class mix data also varies from substation to substation and is dependent on weather and time. It is necessary to often determine and update the load class mix data for each substation of the system. A new composite load model has been developed by the Western Electricity Coordinate Council (WECC) Modeling and Validation Work Group (MVWG) that include data for 12 climates zones in the WECC area, five seasons, four types of substations, and 24 operating hours. Also, it has been recognized that the new composite load model can better represent fault-induced delayed voltage recovery (FIDVR) phenomena, of which, air-conditioning stalling is a major factor. This helps to make power system simulation one-step closer to reality. In recent years, the advancement of simulation tools such as General Electric s PSLF, Siemen s PTI PSS/E, and PowerWorld Simulator, have implemented the WECC composite load model to better represent FIDVR events. Because of this, the synthesis of dynamic load signatures can be generated using prior information of the model parameters obtained from the WECC composite load model spreadsheet. This allows us to create a large amount of dynamic load signatures corresponding to residential, commercial, agricultural, or mixed load classes.

16 3 Machine learning has been well-known as a promising tool for pattern recognition and classification in many applications, including face detection, speech recognition, image classification, and power quality events detection. With the complexity of the dynamic changing of the load composition and the emergence of artificial intelligence in power systems, there is a need of automated load monitoring to classify the load classes as well as to estimate the load composition of each load class for improving load modeling and energy management. 1.2 Research Contribution The key contribution of this research includes: 1) To demonstrate automated disturbance data creation and collection using the latest composite load model developed by WECC. 2) Introduce a new automated solution to determine the load classes by leveraging synchrophasor data and machine learning techniques. 1.3 Organization of this Thesis Chapter 2 introduces the fundamentals of machine learning techniques, a summary of power system model data, as well as an overview of the latest dynamic composite load model. Chapter 3 explains the methodologies used for this research work. There are four main sections for this chapter. The first section demonstrates composite load model data development using the production-level tool and the open-source composite load model spreadsheet to create various dynamic records corresponding to the Northwest Valley climate zone for five different seasons with on-peak and off-peak operating hours. The

17 4 second section presents the static and dynamic power system models used for PSS/E simulation and determines the case scenarios for designing the automation scheme. The third section illustrates the automated PSS/E simulation and data collection schemes. Then, the data processing techniques and classification algorithm is presented theoretically and practically in the last section. Chapter 4 presents real-time simulation and classification results as well as the discussion and evaluation over the proposed methodologies. Finally, the main conclusions and suggestions for future work will be given in chapter 5.

18 5 Chapter 2: Background 2.1 Fundamentals of Machine Learning Techniques Machine learning is a subset of artificial intelligence that has been well known and widely used for many applications such as facial recognition, image classification, natural language understanding, robotics, and power quality monitoring. According to Authur Semuel s definition, a machine learning algorithm gives computers the ability to learn without being explicitly programmed [5]. The input to a machine learning algorithm is a set of instances, also known as examples. For this thesis, we will use term examples to refer to the input for simplicity. These examples are the characteristics used for classification. The output of the machine learning is a set of targeted classes, each element in the output vector is called class label. In general, a training data set is composed of a number of features (similar to measurements) and class labels. For example, a given training set may contain n number of examples x i (i = 1, 2,..., n), and each example is a set of {x 1,i, x 2,i,..., x d,i, y i }. d is denoted as dimension number of features (d-dimensional data) and y i represents the class label associated to example i. The training data with known class labels is called labelled data. Otherwise, it is called unlabelled data. [6]. Machine learning techniques are classified into three categories: supervised learning, unsupervised learning, and semi-supervised learning. Supervised learning requires labelled data for the training stage, whereas unsupervised learning can be trained with unlabelled data. Semi-supervised learning requires a small set of labelled data and large

19 6 amount of unlabelled data for training the model. Supervised learning techniques such as Artificial Neural Network (ANN) and Support Vector Machine (SVM) have been widely used for fault detection, classification, and location problems in power systems. Many researchers have been improving the classification task by combining ANN with some techniques such as Wavelet Transformation (WT) and Principle Component Analysis (PCA) for feature reduction [7 9]. Furthermore, ANN has the capability to map complex and highly nonlinear input-output behavior, which makes it more suitable for pattern recognition in power systems [9]. 2.2 Power System Model Data In general, a power system is composed of generation, transmission, and loads. For power system grid studies and planning, these components can be represented by algebraic and differential equations in dynamic simulations. In practice, the grid planning and operating decisions rely on simulations of the dynamic response of the power system [10]. Therefore, it is very important to carefully select the dynamic models that can minimize the margin of error of the simulations. Recently, the power system simulations range from relatively simple and approximate approaches to more advanced time-domain approaches. Also, power system modeling is arguably acceptable for analysis, system operation, planning, and contingency studies, as well as validating new algorithms and approaches. Figure 2.1 illustrates the structure of a completed data set for performing a dynamic simulation. Static data consists of generator, load, and transmission network information. This type of data is available in many of the IEEE standard test systems. The dynamic data, on the other hand, contains generator, exciter, governor, stabilizer, and

20 7 load information. This can change based on testing purposes. Protection data is often associated with the generator and dynamic load model data. Power Flow Data Static Data Line Limits Generator Exciter A Complete Data Set Dynamic Data Governor Stabilizer Load Protection Data Line Protection Generator Protection Load Protection Figure 2.1: A complete dynamic power system dataset.[11] Obtaining a complete set of data for dynamic simulation to evaluate the proposed methodologies is a challenging task. The generator and end-use load data may not be accessible or not available from the utilities. Fortunately, there have been many developments and implementations of standard test systems that are provided by IEEE, textbooks, and universities. The Reliability Test System (RTS-96) is one of the standard test systems that is widely used by several IEEE journals and international journals for

21 8 evaluating and reporting their research results [12]. Thus, we consider using one area of the reliability test system as the benchmark test for our research. The generator models range from the simplest to highly detailed representations of the synchronous machines. All models share certain common features, however, all generator models ultimately present the electric transmission network with a positive sequence source voltage where instantaneous amplitude and phase are known. More details for all the generator models can be found in the PSS/E model library and the program application guideline [13, 14]. Similarly, the load models range from the simplest to a more complex representation of the loads. For the very basic consideration, the load can be represented as constant power, constant current, and/or constant impedance. These are called static loads. The choice of load models for the simulation is dependant on the purpose of the study. For a more detailed study of the dynamic response of the system, a complex load model such CLOD models, induction machine models, or composite load models should be used [13]. 2.3 Composite Load Model Generally, there are two types of load models: static load models and dynamic load models (composite load models). However, the static load model has limited capacity to represent the induction motor transient behavior under disturbance conditions. Therefore, the composite load model is chosen for this research.

22 Overview It has been recognized that load modeling is a very important component in simulation for predicting the dynamic behavior of systems. As a result, there has been an increased amount of research and development in load modeling for more than two decades [15]. Nevertheless, load modeling is still a complex aspect in research communities and electric utilities for two main reasons. Firstly, the load demand changes constantly not only in magnitude, but also in composition, varied by temperature conditions and patterns of human activities. Secondly, load modeling is practically representing the aggregate of large numbers of load components connected at the distribution substation, which consists of the complex distribution components [15]. In today s energy system, there are significant increases in electronic loads such as electronic drives and electric vehicle charging systems. Plus, the incentive of replacing incandescent lighting with fluorescent lights for energy conservation. Decreasing of resistive loads and increasing of electronic loads adds more stress on the system when the voltage declines. Furthermore, the residential and commercial air-conditioners have been recognized as a major impact of fault-induced delayed voltage recovery (FIDVR) events in Florida and Southern California [16, 17]. According to the Western Electricity Coordinating Council (WECC), FIDVR is a phenomenon when the power system voltages remains at substantially reduced levels several seconds after a transmission fault is cleared as shown in Figure 2.2.

23 10 Figure 2.2: The observed voltage recovery in Southern California, [16] Simulation results in Figure 2.3 show that including an air-conditioning model into the composite load can better represent FIDVR events as compared the previous composite load model in Figure 2.4. In order to better model and understand the reliability risks in the event of a FIDVR, a new composite load model has been developed by the WECC Modeling and Validation Work Group (MVWG) that includes data for 12 climates zones in the WECC, five seasons, four types of substations, and 24 operating hours. Moreover, the production-level simulators such as GE PSLF, Siemen PTI PSS/E, Power World Simulator, and PowerTech TSAT have implemented new composite load models which are currently available in the model libraries. The new composite loads have a much more realistic model structure and have been shown to capture a wide range of dynamic load responses observed in reality. The model has been proven to be numerically stable and robust, as thousands of simulation runs have been made with the model to date [10].

24 11 Figure 2.3: Simulation result using air-conditioning model data and a static model [16] Figure 2.4: Simulation result using the previous composite load model, which consists of 20 percent of induction motor and 80 percent static load. [16]

25 Structure Figure 2.5 shows the structure of the WECC composite load model, which consists of two mains features: a) the model that represents the electrical distance between the transmission bus and the end-use components, and b) the model that represents load composition and dynamic characteristic of various electric end-uses. Figure 2.5: Composite Load Model Structure [14]. The WECC composite load model consist of four motors, an electronic load, and a static load. Motor A represents three-phase induction motors with high inertia driving constant torque loads. Those motors represent either small or large commercial cooling systems and refrigeration systems such as the rooftop air-conditioners on the shopping malls, grocery stores, etc. The model data represents 5-15 HP and/or HP for

26 13 small commercial and large commercial cooling systems, respectively. Typically, they trip at 0.65 p.u and they require manual restart when they are tripped. Motor B represents three-phase induction motor with high inertia driving loads whose torque is proportional to speed squared. Those motors are the fans motors in the residential and commercial buildings. Motor C represents three-phase motors with low inertia driving load whose torque is proportional to speed squared. Typically, they represent direct-connected pump motors used in commercial buildings such as water circulating pumps in central cooling systems with 5-25 HP [18]. The default model data uses two-cage representation for three-phase motors. The equivalent circuit for three-phase induction motor model is shown in Figure 2.6 [14]. It is worth considering that these motor models are intended to represent the aggregation of many motors in distribution system where no detailed information on the actual characteristics of the individual motors is known [18]. Figure 2.6: Equivalent circuit for three-phase induction motor model [14].

27 14 Motor D is a special design motor model that represents single-phase compressor motors found in cooling and refrigeration systems for residential and small commercial loads. The single-phase compressor motor protection includes thermal time over-current relay with two trip levels. The electronic model represents power electronics loads such as electric vehicle changing system. The composite load model also includes system and feeder protection, such as under-frequency and under-voltage load shedding. The static load model represents resistive lighting, electronic, and small appliances. The model is represented by the following equations [19]: P = P 0 (P 1c ( V V 0 ) P1e + P 2c ( V V 0 ) P2e + P 3 )(1 + (P frq f)) Q = Q 0 (Q 1c ( V V 0 ) Q1e + Q 2c ( V V 0 ) Q2e + Q 3 )(1 + (Q frq f)) P 0 = P load (1 F ma F mb F mc F md F el ) (2.1) Q 0 = P 0 tan(acos(p F s )) P 3 = 1 P 1c P 2c Q 3 = 1 Q 1c Q 2c Where P 0 and Q 0 are real and reactive power at pre-disturbance conditions. V 0 and V are the operating bus and rated voltage, respectively. And, the paramenter f is the variation of the bus frequency. The description of the remaining parameters can be found in Table 1 in Appendix A.

28 15 Chapter 3: Methodology Figure 3.1 summarizes the development approach for load classification. In machine learning, data collection and data preprocessing are the most challenging and timeconsuming tasks. Furthermore, the accuracy of the classification is heavily dependent on the quality and amount of training data. Once the training data has been scrubbed, it can be used for training any classifier available in Matlab or Python. The Python based machine learning library provided by the scikit-learn software is used for classification in this thesis. The following sections will provide a step-by-step process for building a dynamic signature library as well as providing a data processing methodology for improving classifier performance. The last section will discuss the theoretical framework of multi-layer perceptron and its application for dynamic load signature classification. 3.1 Composite Load Model Development The WECC composite load model data consists of 131 parameters as listed in Table 1 in Appendix A. This data is categorized into the following set: substation and feeder, transmission and distribution transformer, load composition fraction, electronic load, static load, motor type, motor A, motor B, motor C and motor D parameters. The composition of these parameters makes up the composite load model. The Siemens PTI PSS/E implementation of WECC composite load model (CMLDBU1) is used for this study. The following sections provide more details for WECC composite load model data acquisition and instructions for how to use the production-level Load Model Data Tool

29 16 Figure 3.1: Methodology Structure. to generate dynamic model record data in PSS/E format Composite Load Model Data Acquisition The WECC Modeling and Validation Work Group (MVWG) have developed a composite load model that includes data from 12 climate zones in WECC, five seasons, and four types of substations. The data is also diversified for 24 hours. Figures 3.2 and 3.3 show the geography boundaries of the 12 climate zones and the representative city for each climate zone.

30 17 Figure 3.2: LID Region. [10] Figure 3.3: WECC climate Areas. [10] WECC uses the long identifier (LID) load to classify the load type at the substation level. The LID is formatted as such <3 character climate zone> <3 character load class >. For example, the residential load in the Northwest Valley would have LID

31 18 as NWV RES. There are four types of load class mixes (also known as Feeder Type) - Residential (RES), Commercial (COM), Rural and Agricultural (RAG), and Mixed (MIX). Each feeder type is the aggregation of four load classes. Figure 3.4 shows the percentage of each load class made up for each feeder type. Figure 3.4: Feeder types. [10] The composite load model spreadsheet developed by WECC allows the user to create the dynamic composite load data for all 12 climate zone and four different load class mixes by selecting season and the operating hour. The 24-hour load profile and load fraction of motor A, motor B, motor C, motor D, electronic loads, and static loads for residential and agricultural loads in the Northwest Valley climate zone (NWV RES and NWV RAG) are shown in Figure 3.5 to 3.8.

32 19 Figure 3.5: Load profile for residential load class at 4 PM peak load in the Northwest Valley climate zone. Figure 3.6: Load fraction for residential load class at 4 PM peak load in the Northwest Valley climate zone.

33 20 Figure 3.7: Load profiles for agricultural load class at 4 PM peak load in the Northwest Valley climate zone. Figure 3.8: Load fraction for agricultural load class at 4 PM peak load in the Northwest Valley climate zone.

34 21 By specifying the operating hour, the composite load model spreadsheet generates the composite load data to be saved in a CSV format. This file will be needed for the Load Model Data Tool, which will be discussed in the next section. In order to obtain more training data for signature detection and classification task, we chose the maximum peak-load times (9 AM, 4 PM, and 6 PM pacific time zone) and the minimum peak-load times (2 AM and 9 PM in pacific time zone). This allows us to create five dynamic records for each load class. Table 3.1 shows the composite load data for the residential load in hot summer at 6 PM and 2 AM in the Northwest Valley climate zone.

35 Table 3.1: Load composition data for Northwest Valley climate zone. 22

36 Load Model Data Tool The Load Model Data Tool (LMDT) is an open-source licensed tool developed by the Pacific Northwest National Laboratory (PNNL) in collaboration with Bonneville Power Administration (BPA) and WECC MVWG[18]. Figure 3.9 shows the first version of the tool (LMDT 1.0), which was released in It requires machine characteristic, load composition, and power flow data in three separate spreadsheets to generate a dynamic record in GE PSLF or Siemens PTI PSS/E format. The user needs to provide input to LMDT that describes the load composition data. This can be obtained using the WECC composite load model spreadsheet as described in section In addition, the user needs to provide the power flow information and determine the feeder type for each load bus. The induction machine data is available on WECC. After importing all three files, the LMDT reads in the necessary long identifier (LID) information, and the supplemental base case power flow information to generate dynamic records. These dynamic records will be added to the existing dynamic files that contain the dynamic model data for the generators, which will be used in the time-domain simulation.

37 24 Figure 3.9: Load Model Data Tool. [10] 3.2 Time-Domain Dynamic Simulation Setup Dynamic Power System Models The One-Area Reliability Test System (RTS-96) is used for the study. The system has 10 generators and 14 loads that are connected by 34 branches. Three transformers are used to step up the voltage from 138 kv to 230 kv.

38 25 Figure 3.10: IEEE One-Area RTS-96. The following is the description for each model used for the generators in the reliability test system [13]: GENROU - is a round rotor generator model that represents the principle response characteristics of a normal generator, and is used for each generator in the system. IEEEX1 - is an excitation system that uses dc generator as its main exciter. IEEGO - is a linear and general-purpose hydro governor model that is widely used to represent the effect of power plants on power system stability. PSS2A- is used to represent a variety of stabilizers with inputs of power, speed, or frequency.

39 Case Scenarios The composite load is connected at bus 2. At time 2 seconds, a single-line-to-ground fault is applied at the bus connected to the composite load bus for 15 cycles, which is about 0.25 seconds. At time 2.25 seconds the fault is cleared. The simulation ends at a final time of 20 seconds. This time window is chosen to allow sufficient collection of measurement data since the default value of air conditioner restart time parameter (Tth) is set to 15. This means the voltage recovery will take up to 15 seconds. This process is repeated for all buses connected to the composite load bus and for dynamic records created in section Additionally, all of these processes are repeated for a line-to-line fault. Figure 3.11 highlights the locations where faults are applied. For dynamic signatures, we only consider the bus voltage, bus frequency, and bus real power measurements because a majority of the power quality disturbances are found in voltage waveform and current waveform [20]. Moreover, according to M. Biswal [21], the frequency and voltage data have the strongest disturbance signals for event diagnosis using synchrophasor data. To make data collection similar to real-world measurements using Phasor Measurement Units (PMUs) in power systems, this dynamic signature data is measured at the composite load bus (Bus2) and at the 230-kV level bus (Bus15) The user has the option to select the number of new initial conditions to create based on the given base case (power flow network). Afterword, the automation performs dynamic simulations for all the dynamic files (e.g 100 dynamic files are created) using the composite load model spreadsheet by varying the seasons and the time of day. For the purpose of this study, the automation uses the original power flow system to create two new base cases by randomly increasing or decreasing the real and/or reactive power by 2 percent and verifying the system stability by performing steady state simulation

40 27 CLM Figure 3.11: Single-line-ground and line-to-line fault locations.

41 28 before saving the new save case files to the folder. The automation generates up to 1200 output files to be saved in the dynamic signature library. 3.3 Automated Disturbance Data Collection In machine learning, the performance of the classifier is heavily reliant on the quantity and quality of the training data. However, performing dynamic simulations to get disturbance data is a time-consuming task. Fortunately, PSS/E provides a comprehensive Application Programming Interface (API) in Python for automating PSS/E simulations. By developing an automated simulation, one can reduce the time required to perform all the necessary static and dynamic simulations. Automation can also help reduce the chance of errors as all the processes are performed exactly the same every time whereas humans could miss steps or enter incorrect data. Once automation is developed, it can be used multiple times and for many different power flow systems.

42 Python based PSS/E Automation Obtain base case (power flow) Generate new cases for different initial conditions Save the new case (power flow) Get the new case Solve the power flow using Newton Raphson Method No Is the system solvable? Yes Convert generator and load Load dynamic file Yes Perform steady state simulation for 2 seconds Apply disturbance for 0.25 seconds (15 cycles) Clear fault and run the simulation for 20 seconds End No Are there any more dynamic files? Save the output file Figure 3.12: Automation scheme for dynamic simulation.

43 Python based Data Collection Get output file Assign class label based on output file name Search for missing data (NAN) Is there any NAN? Yes Delete output file No Count the number of output files for each class Balance the output files for each class Write all output files to a CSV file Figure 3.13: Automated data collection. This automated scheme write the output files generated from the dynamic simulation to a CSV file.

44 Detection and Classification Algorithms The multi-layer perceptron (MLP) is a supervised learning algorithm, and it is by far the most well-known and most widely used model among the existing Artificial Neural Network (ANN) paradigms for pattern recognition problems [22]. It can learn a nonlinear function f(.) : R m R o by training a dataset, where m is the dimension of the input vector and o is the dimension of the output vector Multi-Layer Perceptron Theoretical Framework Mathematical formulation To describe neural network, consider a simple single-layer neural network shown in figure 3.11 below: Figure 3.14: Single-layer neural network structure. The leftmost layer is called input layer, which consists of two input elements x1, x2, and a biased element b = +1, whereas the center layer is called hidden-layer and the rightmost layer is called output layer. The output y = f(h w,b (X)) = f( 2 i=1 W ix i + b)

45 32 and the function f() is called the activation function. In general, there are four different types of activation functions: step, sigmoid, tanh and logistic functions. The sigmoid function is most widely used in ANN, and it is written as follows: f(z) = e z (3.1) Assume binary classification, y = {0, 1}. Then, h w,b (X) is passed through the sigmoid function f(z) = 1 1+e z to obtain output values between zero and one. A threshold set to 0.5 would assign samples of the outputs greater than or equal to 0.5 to class one and the rest to class zero. For the multi-class classification, that have more than two classes, the weight summation h w,b (X) would be a vector of size n classes is passing through the Softmax function, which is given as: Softmax(z) i = e z i k l=1 ez l (3.2) where z i represents the i th element of the input to Softmax, which corresponds to class i. The constant k is the number of classes. The results would be a vector containing the probabilities that sample x belong to each class. The output is the class with the highest probability. Basic Learning Theory Gradient-based learning approaches are commonly used in machine learning. Considering the gradient-based learning machine shown in Figure 3.12, the learning machine takes the input X p, the p th input pattern, and the collection of adjustable parameters in the system (W) to compute a function M(X p, W ). The cost function, E p = 1 2 (Dp M(X p, W )) 2, measures the discrepancy between the desired output D p and

46 33 the output produced by the learning machine. The average cost function E train (W ) is the average of the errors E p over a given set of training data (X 1, D 1 ), (X 2, D 2 ),..., (X p, D p ). In the simplest setting the machine learning problem consists in finding the value of W that minimize E train (W ). Figure 3.15: The gradient-based learning machine.[23] Basic Backpropagation The backpropagation approach is used for maximizing the ability to predict the target pattern that the system has not previously seen. To better understand how backpropagation works, let s consider a simple multi-layer neural network using the sigmoid function for activation [23]. The weighted sums and output vector are computed as follows: Y n = W n X n 1 (3.3) X n = F (Y n ) (3.4) where X n is a vector representing the output at layer n. X n 1 is input vector to layer n. It is worth noting that the initial input vector is represented by X 0. W n is a matrix whose number of rows is the dimension of X n, and the number of columns is the dimension of

47 34 X n 1. F () is a sigmoid vector function applied to each of its input components. Y n is the vector of weighted sums. Assume the partial derivative of the cost function with respect to the output vector, Ep X n, is known. Then, by applying chain rule to equation 3.3 and 3.4, the backpropagation equations are obtained as shown below: E p Y n = F (Y n ) Ep X n (3.5) E p E p = X n 1 (3.6) W n Y n E p X n 1 = W T n E p X n (3.7) The gradient decent algorithm adjusts the parameters W using the following iteration: W (t) = W (t 1) η E W (3.8) where η is a learning rate parameter, and it is a constant for the simplest case Load Signature Classification using Multi-Layer Perceptron Figure 3.12 shows the structure of the MLP model designed for classification algorithms. Given a set of the training data in the form of (X 1, Y 1 ), (X 2, Y 2 ),..., (X m, Y n ), where m is the number of input examples. Also, m is corresponding to the number of the dynamic signatures generated by the simulation. n is the number of classes, where n = 4 for this case. The input features have 6 dimensions since there are two sets of voltage measurements, two sets of real power measurements, and two sets of bus frequency measurements.

48 35 Figure 3.16: Multi-layer Perceptron Neural Network Structure. The scikit-learn MLP classifier library is used for classification implementation. The MLP learns a non-linear function for classification using backprogation [24], which is a very popular neural network learning algorithm because of its simplicity, efficiency, and its high success rate [23] Data Preprocessing Class Labelling for Supervised Learning The class label is assigned based on the name of the output file which is properly named based on season, operating hour, and load class mix (feeder type). For example, the dynamic signature corresponding for single-line-to-ground fault at bus 1 occurred in hot

49 36 summer at 4pm (pacific time zone) at the residential load feeder, the output name would be Bus1Fault HS 16 RES.out. This output file represents one of the input examples for training machine learning. The automated data collection will read the three letter acronym before.out and assign a class label to each example accordingly. Data Cleaning and Normalization Most datasets encountered in practice contain missing values. Likewise, the synthesis data obtained from the simulation contains missing data, which are typically represented by NAN (Not A Number). This missing data can be handled by replacing it with the mean, zeroes, or replacing with something else based on the domain expert. For this specific case, the missing data may occur when the dynamic simulation fails to converge due to the unbalancing of the reactive power drawn by the load and the generator capacity limit. Therefore, we chose to remove the entire example set that contains NAN. For data normalization, it is highly recommended to scale the data since the Multilayer Perceptron (MLP) is sensitive to feature scaling. This must also be applied to the test data [24]. There are two common approaches for data normalization: The first approach is scaling all the numeric values in the dataset to lie between 0 and 1. The formula is given in equation 3.9. The second approach uses standardization to transform data to have zero mean and unit variance using equation 3.10 [25]. x new = x x min x max x min (3.9) x new = x µ σ (3.10)

50 37 Feature Reduction using Principle Component Analysis Due to a large amount of data in power systems, the detection and classification performance may degrade by the curse of dimensionality. [26] There has been significant research showing that the classification performance in terms of training speed and accuracy can be improved using Principle Component Analysis (PCA)[26, 27]. PCA is known as a simple and traditional technique, yet a powerful tool for feature reduction. The main purposes of PCA are to analyze the data to identify patterns and reduce the dimensions of the dataset with minimal loss of information. PCA toolbox and user-interface application are available in MATLAB which can be used easily with any well-formatted raw data. However, we are using a Python based PCA package that is available on the open-source scikit-learn library.

51 38 Chapter 4: Results and Discussion Dynamic Simulation Results The simulation results for dynamic responses of the system due to two different disturbances are shown in Figures 4.1 to Figures 4.1 to 4.6 show the dynamic responses due to a single-line-to-ground fault occurred at bus1, which is connected to the composite load bus (Bus2). As can be seen in the figures, the residential load has a longer time delay compared to the other three load classes. This is due to the amount of the residential air-conditioning loads connected to the feeder (or bus) in hot summer. According to Table 4.1, 35 percent of the load components are air-conditioning which is almost three times more than the other load components for the residential feeder. Comparing to the other feeder types, the percentage of air-conditioner load components dominates the load aggregation. Therefore, we can see a significant impact of the air-conditioning stalling on the dynamic response. On the other hand, the dynamic responses, shown in Figures 4.7 to 4.12, are the results from the line-to-line fault occurred on the line connected from bus2 to bus1. There is no significant delay due to air-conditioning stalling. Based on Table 4.1, the percentage of the air-conditioning loads are low for all feeder types in winter compared to the other load fractions. As has been noticed, some simulation results may have similar dynamic behavior. For example, the voltage responses due to line-to-line fault shown in Figure 4.8, look almost identical for all feeder types (load class mix). However, Figure 4.13, which is a

52 39 close-up of Figure 4.8, shows that there is slight difference in magnitude for all four load classes. In short, there are some discrepancies in each class because of their respective load compositions. This variance indicates that there are patterns that can be further applied for data processing to train the machine learning algorithm to classify those load classes. Figure 4.1: Voltage signatures measured at Bus2 for single-line-to-ground fault at Bus1 Table 4.1: Load composition data for Northwest Valley.

53 40 Figure 4.2: Voltage signatures measured at Bus15 for single-line-to-ground fault at Bus1. Figure 4.3: Power signatures measured at Bus2 for single-line-to-ground fault at Bus1

54 41 Figure 4.4: Power signatures measured at Bus15 for single-line-to-ground fault at Bus1. Figure 4.5: Frequency signatures measured at Bus2 for single-line-to-ground fault at Bus1

55 42 Figure 4.6: Frequency signatures measured at Bus15 for single-line-to-ground fault at Bus1. Figure 4.7: Voltage signatures measured at Bus2 for line-to-line fault on the transmission line connected between Bus2 and Bus1

56 43 Figure 4.8: Voltage signatures measured at Bus15 for line-to-line fault on the transmission line connected between Bus2 and Bus1. Figure 4.9: Real power signatures measured at Bus2 for line-to-line fault on the transmission line connected between Bus2 and Bus1

57 44 Figure 4.10: Real power signatures measured at Bus15 for line-to-line fault on the transmission line connected between Bus2 and Bus1. Figure 4.11: Frequency signatures measured at Bus2 for line-to-line fault on the transmission line connected between Bus2 and Bus1

58 45 Figure 4.12: Frequency signatures measured at Bus15 for line-to-line fault on the transmission line connected between Bus2 and Bus1. Figure 4.13: A close-up of the voltage recovery (voltage signatures) due to line-to-line fault on the transmission line connected between Bus2 and Bus1.

59 Classification Results The normalized confusion matrix, shown in Figure 4.14, evaluates the accuracy of MLP classification for each class. The diagonal elements represent the percentage for which the predicted label is equal to the true label, while the off-diagonal elements are those that are mislabeled by the classifier. The overall accuracy of the classifier is approximately 87% using PCA for feature reduction. The MLP neural network contains 15 hidden layers. 80% of the data is used for training, while 20% is used for testing the model. It is worth considering that the accuracy of the classification drops to 63.9% if 0-1 scaling method is used for data normalization. Also, the accuracy classification is decreased approximately to 47.7% if feature reduction using PCA is not considered. This result is lower than the classification performance using raw data, which has about 58% accuracy. Table 4.2 and Figures 4.15 to 4.18 summarize the accuracy of MLP classifier using the same dataset for training and testing and varying the data processing techniques. Table 4.2: MLP Classification Accuracy. Case Accuracy 1. With standard normalization and PCA 86.76% 2. With scaling normalization and PCA 63.97% 3. With standard normalization only 47.79% 4. With scaling normalization only 40.44% 5. With raw data 58.09%

60 47 Figure 4.14: Normalized confusion matrix for all four load classes. The overall classification accuracy is if both standardization and PCA techniques are using for data preprocessing Figure 4.15: Normalized confusion matrix for all four load classes. The overall classification accuracy is if both scaling (0-1) and PCA techniques are using for data preprocessing

61 48 Figure 4.16: Normalized confusion matrix for all four load classes. The overall classification accuracy is if only standardization technique is using for data preprocessing Figure 4.17: Normalized confusion matrix for all four load classes. The overall classification accuracy is if only scaling (0-1) technique is using for data preprocessing

62 Figure 4.18: Normalized confusion matrix for all four load classes. The overall classification accuracy is if 80% of raw data is used for training and 20% of raw data is used for testing 49