MOTOR LOAD DETECTION FOR VOLTAGE TRANSIENT BASED NON-INTRUSIVE LOAD MONITORING. Paul D. Del Mar

Transcription

1 MOTOR LOAD DETECTION FOR VOLTAGE TRANSIENT BASED NON-INTRUSIVE LOAD MONITORING by Paul D. Del Mar A thesis submitted to the Faculty of the University of Delaware in partial fulfillment of the requirements for the degree of Master of Science in Electrical and Computer Engineering Summer 2013 Copyright 2013 Paul D. Del Mar All Rights Reserved

2 MOTOR LOAD DETECTION FOR VOLTAGE TRANSIENT BASED NON-INSTRUSIVE LOAD MONITORING by Paul D. Del Mar Approved: Keith W. Goossen, Ph.D. Professor in charge of thesis on behalf of the Advisory Committee Approved: Kenneth E. Barner, Ph.D. Chair of the Department of Electrical and Computer Engineering Approved: Babatunde A. Ogunnaike, Ph.D. Interim Dean of the College of Engineering Approved: James G. Richards, Ph.D. Vice Provost for Graduate and Professional Education

3 ACKNOWLEDGMENTS First of all, I would like to thank my adviser Dr. Keith Goossen for his guidance and enthusiasm, but most of all for his patience throughout the duration of this project. I would also like to thank Cesar Duarte for all of his help with this work, including sharing his code for the notch filter needed to analyze some of the data presented in this paper. Thanks to Dr. Cory Budischak for his mentoring, motivation, and for his friendship. Most of all I would like to thank my parents, Paul and Charyl Del Mar, and my sister Jennifer Del Mar-Haroian for all of their love and support over the years for which I am and will always be extremely grateful. iii

4 TABLE OF CONTENTS LIST OF FIGURES... vi ABSTRACT... ix Chapter 1 INTRODUCTION Motivation Impact of Energy Management Outline of Thesis NON-INTRUSIVE LOAD MONITORING METHODS Direct Sensing Methods Non-Intrusive Load Monitoring Techniques Aggregate Non-Intrusive Load Monitoring Disaggregated Non-Intrusive Load Monitoring Techniques Real/Reactive Power Methods Power Harmonics Methods Current Waveform Based Methods Voltage Signature Based Methods SWITCHING TRANSIENTS AND SIGNAL PROCESSING BACKGROUND Voltage Transient Background Transients due to Arcing During Commutation in Motors Signal Processing Background Short-Time Fourier Transform iv

5 4 EXPERIMENTAL RESULTS Instrumentation Signal Post-Processing Power Line Comparison Voltage Transient Analysis Using Short-Time Fourier Transform Voltage Transient Analysis Using Wavelet Transform Motor Transient Detection and Classification DISCUSSION AND CONCLUSION REFERENCES v

6 LIST OF FIGURES Figure 1.1 Estimated US Energy 2010 Usage Breakdown... 2 Figure 2.1 Aggregate power plot showing step changes related to individual appliances... 8 Figure 2.2 Real versus reactive power disaggregating certain loads Figure 2.3 Figure 2.4 Power profile of a heat pump including the presence of an initial startup spike and noticeable slope in real power over time rd harmonics can be used to differentiate between an incandescent light and a computer with similar complex power values Figure 2.5 Spectral envelop pattern matching Figure 2.6 Current waveforms and 5 level wavelet decompositions of a) Personal Computer, b) 200W Dimmer, c) Both PC and Dimmer Figure 2.7 V-I trajectories for two pairs of similar loads Figure 2.8 Baseline noise on an outlet (top), added noise from an SMPS device (middle), and that noise extracted (bottom) Figure 3.1 Switching Circuit Figure 3.2 Circuit for commutation in a dc or universal motor and arcing when the brush moves to the right, making and breaking contact with commutator segments Figure 3.3 Rectangular Window in Time and Frequency Domains Figure 3.4 Hamming Window in Time and Frequency Domains vi

7 Figure 4.1 Figure 4.2 Figure 4.3 Transient Detection Setup (PC, Digital Oscilloscope, Differential Voltage Probe, Test Outlet, Load) Mask Filter around 60 Hz voltage signal with offsets of 1ms, 100mV Mask Filter catching a voltage transient from a small desk fan Figure 4.4 Voltage signal of the two different powerlines Figure 4.5 Power Spectral Density of Powerlines Figure 4.6 Powerline Power Spectral Densities under 5kHz Figure 4.7 PSD Comparison of both Powerlines Figure 4.8 Figure 4.9 Figure 4.10 Figure 4.11 Figure 4.12 Figure 4.13 Figure 4.14 Figure 4.15 Voltage transient of small desk fan in the office before (top) and after notch filtering (bottom) Zoomed in graph of the same transient seen in previous figure (top) and the corresponding STFT (bottom) Another voltage transient from a small desk fan at the office (top) and corresponding STFT (bottom) Voltage transient of small desk fan at Blackwood location (top) and corresponding STFT (bottom) Voltage transient of small desk fan at Blackwood location (top) and corresponding STFT (bottom) Voltage transient (top) and resulting scalogram (bottom) of a lamp being turned on Another voltage transient (top) and scalogram (bottom) for a lamp being turned on Voltage transient (top) and scalogram (bottom) of a lamp being turned off vii

8 Figure 4.16 Figure 4.17 Figure 4.18 Figure 4.19 Figure 4.20 Another voltage transient (top) and scalogram (bottom) of a lamp being turned off Blender created Voltage Transients (top) and the first three cycles overlapped (bottom) Distribution and Means Comparison for Devices using a window length of 500 samples and mask threshold of 100mV Example of a notch filtered voltage transient from a fan with multiple contact bounces and long pulses Distribution and Means Comparison for Devices using a window length of 5000 samples and mask threshold of 100mV viii

9 ABSTRACT Investment in smart grid technology by the electric utilities is continually growing. Demand peak shifting, motivated by time-of-use pricing, is a way to reduce electric capacity requirements, as well as align loads with variable alternate sources such as wind or solar. These methods require feedback from the customers via a smart meter to see real time power usage and other data such as individual load usage. Smart meters allow the end-user to take advantage of the time-of-use pricing schemes to intelligently use their highest power loads during the low point of pricing, as well as gaining a better understanding of how the loads in their household use electricity. Individual appliance monitoring can be by a distributed sensing network, where each appliance or device has its own direct communication line with the smart meter. A much less costly alternative is non-intrusive load monitoring (NILM), which uses information seen in the power consumption profiles or actual current and voltage signals, to determine appliance usage. NILM does not require professional installation and voltage signal acquisition and can be measured right at a typical household outlet. Devices impart voltage transients or noise on the power line of the house when they are switched that can be used for NILM. The transients of particular loads can provide signatures specific to them. However, the load itself can introduce transients independent of switching, and these need to be distinguished. Particularly, we show here that motors having commutating switches introduce repetitive transients that resemble load switching. Time-frequency analysis provides insight into these ix

10 short-time non-stationary signals. An experimental setup was implemented in order to record voltage transients of a few different common household loads. It is shown that universal motor loads with a commutating switch can be detected separately than transients created by a single switching action typical of most devices. x

11 Chapter 1 INTRODUCTION 1.1 Motivation As the cost of electricity production continues to grow due to rising fuel costs and greater peak demand, the adoption of smart grid technology will become more prevalent. Time-of-use pricing and demand peak shifting will decrease utilization of backup power systems which have higher operating costs, as well as improve system reliability by reducing the chance of blackouts. Also, residential customers can use smart meters to take advantage of time-of-use pricing by running high power loads when the electricity cost is low. In order to take advantage of time-of-use pricing, the customer must also be aware of the power consumption of household appliances. Ideally, smart meters should be able to accurately profile appliance usage with minimal processing time and complexity. Rather than installing expensive sensory and communication equipment on individual appliances, which may not be accessible, it is possible to record usage profiles in a non-intrusive manner. The provision of accurate and convenient information by smart meters allows customers to make informed decisions regarding daily power usage. The impact of cost-effective smart metering will be extensive, since individuals will be able to more intelligently and efficiently manage their power consumption. One way 1

12 of achieving the affordable non-intrusive load monitoring through a smart metering device is through the use of voltage transients that are imparted on the line by the switching of household appliances, which will be discussed in further detail in following chapters. 1.2 Impact of Energy Management The overall estimated United States 2010 energy usage flowchart provided by the Lawrence Livermore National Laboratory and the US Department of Energy is shown below in Figure 1.1 [1]. Figure 1.1 Estimated US Energy 2010 Usage Breakdown [1] 2

13 Out of the approximate 98 Quads of energy used (or BTU) in the US in 2010, 39.5 Quads of energy (in the form of fuel) was used for electricity generation, or ~40% of the total energy. However out of the 39.5 Quads of energy used in electricity generation only Quads of electricity were actually generated (this is equivalent to 3.72 trillion kwh). This means that Quads of the total energy (over 25% of the total energy used) is wasted through inefficiencies in the power plants and the transmission system; overall the total electricity generation industry is operating at 32% efficiency. While the actually achievable maximum efficiency is limited by physical thermodynamic limits, the amount of wasted energy is quite alarming and shows that any, even small, measures taken to increase the efficiency of this sector will have great impact. According to the US Department of Energy s Buildings Energy Data Book for % of total electricity consumption is from buildings in both the residential sector (39% of total) and commercial sector (36% of total) [2]. Using the Quads of total electricity generated in 2010 and applying these percentages, the calculated values (4.96 Quads residential, 4.58 Quads commercial) are in very close agreement to that in the LLNL US energy flowchart, which not surprisingly means that all of the electricity used in these two sectors is also used in buildings. Therefore, (using numbers from the flowchart for convenience) 4.95 Quads of electricity are used in residential buildings and 4.54 Quads in commercial buildings, a total of 9.49 Quads of electricity used in buildings. It has been shown that up to 15% of building electricity consumption can be reduced by applying better energy management schemes with direct feedback [3]. If this savings can be achieved, a total of 1.42 Quads of electricity could be saved just 3

14 over the year, equating to 4.16 x kwh or 416 billion kwh! This number does not even take into account any efficiency increases that the electricity generation sector may achieve due to improved load balancing and possibly eliminating usage of inefficient backup generators. Obviously, smart meters can indeed have a large meaningful impact if they are eventually implemented nationwide, but they are still being developed. As introduced in the previous section, detailed information on how electricity is being consumed in buildings will be very useful for the utilities and policy makers. Therefore, developing an economically feasible and accurate method of load monitoring will have a significant impact on the ongoing attempt to decrease energy usage here in the United States and globally. 1.3 Outline of Thesis The following chapter of this document provides a background and comparison of non-intrusive load monitoring techniques that have previously been implemented, and describes the motivation for using voltage transients as a means of detection and classification of residential loads. Chapter 3 focuses on the theory of voltage transients caused by switching operations and provides an initial description of how different types or classes of residential loads can be differentiated. Universal motors are looked at with greater detail and the transients that are caused by commutation are investigated. Then a brief background of the signal processing techniques used to analyze the voltage transients is presented, specifically the shorttime Fourier transform and the continuous wavelet transform are used for timefrequency analysis. 4

15 Chapter 4 then reports the experimental setup and procedure used for the detection of these voltage transients. Actual signals showing the characteristics of various load types and analysis are then presented. The spectrogram resulting from the short-time Fourier transform is used as well as the scalogram which is the continuous wavelet transform equivalent. Also, motor created transients are identified and shown to be significantly different from other load types. Chapter 5 wraps up this document by examining the analysis of the experimental data and provides a means to identify universal motors and other typical load types. Future considerations and conclusions for this method of load classification are then discussed, which is followed by the references used in this work. 5

16 Chapter 2 NON-INTRUSIVE LOAD MONITORING METHODS Load monitoring will be instrumental for understanding typical electricity usage in the residential and commercial sectors. It is evident that detailed power consumption data will not only aid utilities to properly design and implement loadbalancing strategies and pricing changes, but also help policy makers intelligently develop energy efficiency incentives and measures [4]. This data will also be beneficial for residential end-users so they can further understand their own power consumption patterns and take advantage of any time-of-use pricing schemes implemented by the utilities. This chapter describes the various load monitoring methods that have been previously investigated and will serve as a basis for describing why voltage transient based non-intrusive load monitoring was used in this work including the advantages and limitations of this particular method. 2.1 Direct Sensing Methods One way load monitoring could be achieved is through a building-wide automation network, where a smart meter acts a central data collection point and sensors and communication equipment are installed on each individual appliance in the building. This distributing sensing technique would be extremely accurate since there would be no ambiguity in which appliance was consuming power, since each appliance would be directly reporting usage to the smart meter [5]. 6

17 Another advantage of the direct sensing method is that not only would individual appliance usage be documented you could also conceivably control the equipment. This would allow users to not only monitor the usage of the appliance but the smart meter could automatically program higher energy loads to only run when the electricity rate is low, for instance. Also, calibration of the system would be very easy since each sensor would be given its own unique label for full proof appliance identification [6]. However, this type of system would be extremely costly due to the necessary sophisticated communication equipment and a large number of sensors, resulting in time-consuming installation which can be further complicated by hardwired devices. Each new appliance would have to installed and integrated into the system, while all existing appliances would have to be modified [6]. This direct sensing method of load monitoring is quite invasive, so other non-intrusive load monitoring techniques have been developed instead and are reviewed in the following sections. 2.2 Non-Intrusive Load Monitoring Techniques Non-intrusive load monitoring methods have been developed because there is usually only one single point of sensing, typically located at the main building electrical panel, eliminating extensive distributed sensing equipment. All nonintrusive load monitoring methods have three things in common. First, specific appliance features or signatures need to be selected and characterized. Then, appropriate hardware for data acquisition and sensors to detect the signatures must be determined. Lastly, there has to be some sort of algorithm to extract signatures from 7

18 the data [5]. The next section shows the easiest but also most limited technique of non-intrusive load monitoring: aggregate power monitoring Aggregate Non-Intrusive Load Monitoring Early attempts to monitor load usage from a single sensing point used only the instantaneous aggregate power. The method used known appliance power ratings to identify any step changes in the overall power consumption. For instance, in Figure 2.1 below, at about Time = 2 minutes there is a jump in power by about 250W. This step change was compared to a database of power ratings for all the appliances in the household, finding that the refrigerator is rated for 250W, hence marking the time in which the refrigerator was turned on [7]. Figure 2.1 Aggregate power plot showing step changes related to individual appliances [7] 8

19 Other information such as typical duration of a connected device can also be used to identify appliances from aggregate power changes. However, if there are two appliances with similar power draw, it would be hard to differentiate between the two just looking at the aggregate power. Hence, methods to disaggregate the appliance usage had to be developed, and are reviewed in the next section Disaggregated Non-Intrusive Load Monitoring Techniques As seen in the previous section, looking just at the aggregate power changes in a building does not result in reliable data. In order to disaggregate the data so the actual consumption of each individual load is recorded, information other than just real power must be used. The following sections review the disaggregated nonintrusive load monitoring technologies previously developed and accentuate the differences between the particular data used, as well as examining the advantages/disadvantages of each method Real/Reactive Power Methods The same group that originally developed the aggregate power method at MIT also realized that they needed more information to truly disaggregate the data. They developed a method that records both real and reactive power with a low 1 Hz sampling rate [7]. They categorized appliances into three major classes. An algorithm uses the plot of the real power versus reactive power (see Figure 2.2 below) to automatically group the appliances into classes. Purely resistive elements are centered on the real power axis, while appliances with motors have a 9

20 reactive component. Clusters of appliances are then compared to known load power consumption data and used to correlate to the power versus time plot (as seen previously in Figure 2.1) to determine a load operating schedule. Figure 2.2 Real versus reactive power disaggregating certain loads [7] However, the database of individual appliance power consumption levels is obtained through an intrusive, albeit one-time, procedure. Furthermore, this method requires the capability of measuring VARs, so voltage and current must be sampled synchronously which results in expensive installation costs. More importantly, since 10

21 there are typically many similar small loads in a household, this load monitoring method makes it difficult to identify devices that consume little power [6]. This method was eventually extended to include typical real power spikes and slopes as detectable features. Often when a device is turned on, there will be a significant startup spike in real power and an associated slope in power draw afterward. These power profiles can have characteristic shape based on the type of load and can be used for further disaggregating of load. In Figure 2.3 below, the power profile of a typical heat pump is shown. Figure 2.3 Power profile of a heat pump including the presence of an initial startup spike and noticeable slope in real power over time [8] While just looking at real and reactive power makes it possible to identify some loads, many researchers started to look at more detailed microscopic features. 11

22 This includes looking at the information contained in the actual current and voltage waveforms as well as analyzing the frequency content, such as looking at power harmonics as discussed in the following section Power Harmonics Methods One of the biggest problems with using real and reactive power for disaggregating loads is when two devices have similar power complex plane profiles (they would overlap on a graph such as Figure 2.2). In order to differentiate similar loads, harmonic content can be analyzed. Below in Figure 2.4, a computer and an incandescent light have nearly identical complex power values for both turn-on and turn-off states [9]. However looking at the 3 rd power harmonic, shows a clear separation of loads, and therefore can be used to complement the complex power method. 12

23 Figure rd harmonics can be used to differentiate between an incandescent light and a computer with similar complex power values [9] Other methods using harmonics have also been investigated. One method uses the spectral envelope and identifies signature envelopes for a particular device. This method uses spectral envelopes contained in a database and matches a recorded spectral envelope using a least squares distance method. They define the typical spectral envelope as having two major sections with shape vectors (s1 and s2 seen in Figure 2.5 below) and compare the similarity of these sections against a previously created database. They also define an overall scale factor with a time shift and offset for each section (alpha, k, and bk respectively shown below), to aid with the pattern maching accuracy [10]. 13

24 Figure 2.5 Spectral envelop pattern matching [10] The biggest drawback of this method is that time consuming training is needed for each device in the household in order to develop a database of spectral envelopes. Also, if a new device is entered into the household, it is not know how this will affect this detection method and there will be no signature in the database for the new appliance. Hence, additional training is required if new appliances are to be monitored, resulting in non-robust method [5]. Another method using power harmonics uses both the transient and steady signals for non-intrusive load monitoring [11]. Their technique involves continuously calculating the signal harmonics and using a neural network for load detection. In an experiment they used eight selected appliances, and trained the network for all of the possible combinations of the appliance states (on of off). This results in 2 8 available combinations. While they were able to achieve decent detection accuracy in their experiments with eight loads, a typical household can have up to 30 appliances, 14

25 resulting in 2 30 possible combinations. This is over a billion combinations, so the use of continuously monitoring power harmonics with a neural network is not practical for household wide load detection [5] Current Waveform Based Methods Rather than just looking at the power profiles and harmonics, other ways of disaggregating appliances for use in non-intrusive load monitoring using the actual electrical current waveforms from the particular loads have been proposed. One such method uses the wavelet decomposition in order to analyze current waveforms with a high content of power harmonics [12]. Wavelet analysis will be discussed in greater detail in Chapter 3; however, it is essentially a way to see the time-dependent frequency content of a signal (see Figure 2.6 below). 15

26 Figure 2.6 Current waveforms and 5 level wavelet decompositions of a) Personal Computer, b) 200W Dimmer, c) Both PC and Dimmer [12] As shown in Figure 2.6a, the current waveform (top) of a personal computer is shown and is rich with harmonics. The 5 level wavelet decomposition (bottom) essentially shows the time-dependent frequency content at each level (or scale). Again, the concept of a wavelet transform will be further developed later, but 16

27 essentially each level corresponds to a particular frequency range. This offers more information on the frequency content of the harmonics present in the current waveforms measured for different non-linear loads. Another unique method that uses information in the current signal from appliances focuses on the geometrical features of normalized I-V curves [13]. Two dimensional I-V trajectories were studied for different load types to try and develop taxonomy of different load types. Steady state current and voltage were measured over only one cycle and magnitudes were normalized against all appliances. An example of two similar load pairs with similar I-V curves are shown below in Figure 2.7. Figure 2.7 V-I trajectories for two pairs of similar loads [13] 17

28 As shown in the top two I-V curves in Figure 2.7, the window-type air conditioner and microwave oven have similar geometries. The similarly of the two intersection points is related to 3 rd harmonic, which both motor-type loads exhibit. Similar geometries between similar loads need to be differentiated somehow and there are a few measures they use to do so. As seen in the bottom two I-V curves in Figure 2.7 for a radio and CD player, the normalized peak values is proposed as a way to differentiate between loads with similar I-V trajectories. While methods using power and current signals can provide some means of disaggregating appliances for non-intrusive load monitoring, they require a licensed electrician to install the sensors on the main electrical panel of the household which can be expensive. One way to eliminate the need of professional installation would be to measure the voltage in a typical outlet in the house and monitor the voltage noise that is imparted on the line by typical household loads. Methods using only voltage for load monitoring are discussed in the following section Voltage Signature Based Methods As mentioned before, instead of measuring current waveforms which requires costly professional installation of in-line sensing equipment, having a voltage probe that can be plugged into any outlet of the house would be much easier for installation and could make the non-intrusive load monitor more customer friendly and accessible, as well as more affordable. Electrical voltage noise is imparted on the power line by devices in the household and differences in this noise can be attributed to different load types. There are three main types of noise on the line: transient voltage noise (from switching), steady-state harmonic voltage noise, and continuous 18

29 voltage noise [14]. These types of voltage noise will be discussed in further detail in Chapter 3, but for now we will explore the methods currently using voltage noise as the main signal used for disaggregating load usage. The first method implemented using voltage noise looks at the transient voltage noise caused by the switching of appliances. Whenever a mechanically switched device is turned on or off it imparts a short burst of noise on the power line [15]. This mechanism will also be explained further in Chapter 3, but the transient noise is basically defined by the type or design of the actual switch, the load type, and the characteristics of the power line over which the transient travels. All of these features create a unique voltage signature for each appliance, which allows for a very good method of disaggregating loads in the household. In particular, since the pathway of the transient affects the recorded signal, two identical loads in different locations of the household could potentially be identified separately. The current method of detecting voltage transients uses a custom powerline interface and a data acquisition system that simply digitizes the incoming AC voltage signal and passes it to a PC for analysis [16]. They use a sliding window algorithm to try and identify any significant changes in the input line noise. Once an event is identified, a feature vector is created that includes the amplitudes of the frequency components from the Fast Fourier Transform of the sliding window segment data. They use a Euclidian distance measure to compare windows. Once a threshold value, which is previously defined, has been breached the start of the transient in determined. Then the window continues to progress until another significant change in Euclidian distance is detected, which results in the end of the transient. Once the transient has been segmented the feature vector is defined and then sent to a support 19

30 vector machine where the transient can be processed and hopefully classified. The support vector machine contains a database of previously learned transient voltage noise signatures (3-5) taken for each appliance in the home, and the inputted feature vector is compared against this database to find the most likely match for means of classification. The biggest disadvantage of this method is that there is an extensive training exercise needed for each appliance in each individual household. This also means that if appliances are moved around the household to different outlets, the system will need to be re-trained or will not be able to detect the load that was moved. Also, any new appliance entered into the home will not be in the support vector machine database and will not be able to be identified. Ideally to have a more robust system, analysis of the voltage transients themselves is needed. Since information about the load type is inherent in the voltage transient signature, this information could be used to have a more generic self-learning detection method which could be common across homes. Ideas for possible improvements of this type of system are discussed later. The second method of using voltage based noise for disaggregating appliances for non-intrusive load monitoring focuses on using continuous noise found on the power line. They found that typical consumer electronic based appliances found in the household use switch mode power supplies (SMPS) which have a specific resonant switching frequency per device [17]. In other words, the electromagnetic interference that is produced on the power line is unique for each SMPS powered appliance. This allows for some form of generalization across homes, where they 20

31 claim that there is wide range of resonant frequencies found in SMPS devices and minimal overlap in the frequency spectrum. Below in Figure 2.8, the top plot shows the baseline noise found at a particular outlet in a home. The middle plot shows the noise introduced after connecting an SMPS power device to the power line. The bottom plot shows the extraction of the added noise created by the SMPS device, where the noise is characterized by a Gaussian fit using amplitude, mean, and, variance as features. Figure 2.8 Baseline noise on an outlet (top), added noise from an SMPS device (middle), and that noise extracted (bottom) [17] 21

32 While looking at continuous voltage noise produced by switch mode power supply devices provides another method of disaggregating appliance usage, it has inherent issues and limitations. Most importantly is that they assume that different appliances all have different SMPS switching frequencies. If two different devices have a spectrum overlap of continuous noise produces by their SMPS, they will not be able to be differentiated. The work that follows in this document is similar to the first method of using voltage noise based transients. However, the authors of that method do not present any data about the transient signals themselves, which will be shown later in Chapter 4. Also, rather than just recording the transients in a database and comparing incoming data to the database, it would be beneficial to analyze the content found in these voltage transients to try and develop a more robust system. In order to this, understanding the voltage transient phenomena is paramount. A background on the creation of voltage transients is presented in the following chapter. Also in the following chapter, the techniques that will be used for analyzing the transients are also reviewed. 22

33 Chapter 3 SWITCHING TRANSIENTS AND SIGNAL PROCESSING BACKGROUND Since this work primarily focuses on classifying residential loads according to the voltage transients that are imparted on the power line due to switching operations, the first half of this chapter is dedicated to describing the sources and characteristics of these transients in order to further understand the information contained in the signals collected in the experimental section. The second half of the chapter is devoted to describing the signal processing techniques used to analyze the voltage transients shown in the following chapter, namely the short-time Fourier transform with resulting spectrogram and the continuous wavelet transform and resulting scalogram. Both methods analyze the time-dependent nature of the signals in the frequency domain (essentially although the scalogram shows the frequency content in relation to scales, which will be described later), which is necessary due to the fact that the transients are time-varying, not stationary, signals. 3.1 Voltage Transient Background The transients created by power electronics in electronic devices have been characterized by Patel, et al. as shown in the previous chapter. They look at the specific switching frequencies of such devices as switch-mode power supplies which have a characteristic high frequency switching noise that is present whenever the 23

34 device will be on. While this is useful in non-intrusive load monitoring for now, if power electronics because more universal (i.e. have the same switching frequencies and characteristics) this method could be problematic. However, still most devices are controlled by mechanical switches and this section describes the theory behind the transient noise produced by mechanically switched devices. Most mechanical switches have solid metal contacts which are in air. Typically when a switch is closed it can bounce, slide, or be influenced by contaminants on the surface, resulting in a series of closings and openings [15]. The real behavior affects the voltage signal and can impart huge spikes of voltages due to the arcing from the openings and closings. When a switch is opened it can also have some bouncing due to switching actions and well as external vibrations. Also, another possible reason for multiple openings could be due to the electrostatic force between the closely spaced contacts. Nevertheless, all switches are built and designed differently and their specific structure and material make up can dramatically influence the noise which is imparted on the power line. Due to this huge variability in switch design is near impossible to truly characterize voltage transient based only on load type and power. For instance, as will be seen later in the next chapter, a small desk fan can generate larger transient signals than higher power devices. However, there is some form of a circuit model in that the load and the line before the switch do influence the transient noise imparted on the line, although it is somewhat influenced by the non-ideal nature of the specific switch. 24

35 The classic circuit model for the switching circuit can be seen below in Figure 3.1. Notice that the line has a resistive, inductive, and capacitive element (R2, L2, C2), as well as the load branch (R1, L1, C1). Each branch of the circuit has its own related resonance, and since the parameters for the line are typically much smaller than in the load side, the resonant frequency f 2 = 1/2π (L 2 C 2 ) 1/2 would be higher than f 1 = 1/2π (L 1 C 1 ) 1/2. Also note, in the figure it shows a DC voltage source, whereas the power line is obviously an AC source, but due to the relative low frequency of the power line (60 Hz) the voltage is nearly constant over the time it takes to open or close a switch. Figure 3.1 Switching Circuit [15] Looking at this circuit model, it can be seen that opening and closing a switch would result in different equivalent circuits that are seen by the switch, and hence it could be hypothesized that the transient from turning on an appliance should indeed be different from turning off an appliance. This will be investigated further in the following chapter to see if this is indeed the case. 25

36 3.1.1 Transients due to Arcing During Commutation in Motors One specific case of mechanical switching that can possibly be differentiated from all other types of loads is found in a commutating motor. Universal motors are quite often found in the household in devices such as blenders, vacuums, etc. and have a commutating switch, much like dc motors. Commutation is the act of reversing the current in an armature coil in order to keep torque of the motor in the same direction present in the stator field of the motor. Since commutation is a constant switching action while the device is on, the transients created when the brushes contact each commutator segment can be considered as continuous or repetitive noise while the device is on. These devices and their continuous or repeating noise that is present while the device is on can be treated similarly to the SMPS devices that were introduced in Chapter 2 Figure 3.2 below shows the equivalent circuit; as the brush moves to the right it makes and breaks contact with the commutator segments related to the different armature coils creating an arc (and resulting voltage transient) at each gap. 26

37 Figure 3.2 Circuit for commutation in a dc or universal motor and arcing when the brush moves to the right, making and breaking contact with commutator segments [18] 3.2 Signal Processing Background As mentioned before, since the voltage transients examined in this work are not stationary signals, typical Fourier analysis would not reveal any information about the time-varying transients, mostly because since they have such short time durations the amount of energy contained in frequency content of the transients would be dominated by the unwanted high frequency line noise which is present throughout the entire voltage signal. In order to analyze these short voltage transients, time-frequency analysis must be used. The short-time Fourier transform results in the frequency content of a windowed section of the signal, however there is a tradeoff between time and frequency resolution as will be discussed later. The continuous wavelet transform is similar to the short-time Fourier transform. However, in the short-time Fourier 27

38 transform, at a particular frequency, changing the window width will result in a change in the number of data points for the frequency content related to the data selected in the window. Alternatively for the wavelet transform, at a particular frequency of interest, changing the window size only scales the carrier frequency, keeping the number of cycles in the window constant. This and other differences of these two time-frequency methods will be further explained in the following sections Short-Time Fourier Transform The short-time Fourier transform or time-dependant Fourier transform uses windows to compute the discrete Fourier transform (DFT) of only the data inside the window function. The window is then shifted to the next desired position and the DFT is calculated of the new window-enclosed data. This process is continued until the window reaches the end of the data set. Equation 3.1 below shows the short-time Fourier transform (STFT): X[ n, ) x[ n m] w[ m] e m j m where w[m] is the windowing function. Notice that this equation transforms the one dimensional signal x[n] into a two dimensional function with parameters n, which is discrete, and wavelength which is continuous [19]. The resulting graph from the STFT is called the spectrogram. The biggest advantage of viewing data this way is that it allows you see the frequency content of the signal as it changes with time, which is particularly useful when looking at non- 28

39 Amplitude Magnitude (db) stationary signals, such as the voltage transients that are later used in this work. The biggest limitation of this technique though is the trade-off between resolutions in the time and frequency domains. Stemming from the uncertainty principle, the more refined your resolution is in the time domain, the poorer the frequency resolution and vice versa. The window function itself also affects the resulting spectrogram and the time-frequency trade-off. One might think that a square window function would be ideal, since you would know exactly what samples of data you are selecting with no variable rating. But as the uncertainty principles dictates, the better resolution in the time domain, the worse the resolution in the frequency domain. Figure 3.3 below show a square window function in the time domain and its resulting spectrum. Time domain 40 Frequency domain Samples Normalized Frequency ( rad/sample) Figure 3.3 Rectangular Window in Time and Frequency Domains 29

40 Amplitude Magnitude (db) Notice that while all 100 samples in the time domain will be selected with the same amplitude, the frequency domain shows that it has poor resolution in the frequency domain. The sidelobes have a significant presence as they are above 0 db. Ideally you would want an impulse at the center frequency, but there are window functions that operate where the frequency resolution is sufficient with little tradeoff in the time domain. Below in Figure 3.4 the Hamming window is shown. Notice how all of the sidelobes in the frequency domain are below 0 db and has a well defined main lobe. The time domain shows that there is some significant weighting of the data samples, but with enough overlap you can still analyze all of your samples. Hamming windows were used on all spectrograms in the following chapters with a 50% overlap. 1 Time domain Frequency domain Samples Normalized Frequency ( rad/sample) Figure 3.4 Hamming Window in Time and Frequency Domains 30

41 Chapter 4 EXPERIMENTAL RESULTS This chapter first describes the experimental setup, including equipment and procedures used in obtaining voltage transient signals from various types of appliances. Then recorded data and resulting analysis are provided in order to show the characteristics inherent in the transient signals of particular types of devices. 4.1 Instrumentation An experimental setup was used in order to record voltage transients caused by switching in appliances. The detection system consists of a digital oscilloscope (PicoScope 4227) with 12-bit resolution and 1% accuracy and up to 100MHz bandwidth connected to a laptop which included the oscilloscope software interface. An active differential voltage probe with a 100:1 attenuation was used to measure the mains voltage. The measurement was taken between the line (or hot) and the neutral, since not all loads use a ground leg. A test outlet was put together in order to get a direct reading of the voltage transients right at the point of connection for the load being measured. The setup can be seen in Figure 4.1 below. 31

42 Figure 4.1 Transient Detection Setup (PC, Digital Oscilloscope, Differential Voltage Probe, Test Outlet, Load) Rather than just streaming the voltage signal and trying to timestamp the switching event a mask filter was used to detect transient signals. Essentially a mask filter is no more than setting an allowable range around the stationary 60 Hz voltage signal from the outlet. Figure 4.2 below shows an example of a mask filter set around the 60 Hz waveform with an X (or time) offset of 1 ms and a Y (or voltage) offset of 100mV (the light blue shaded region denotes areas outside of the allowable region). Typically the range should be as tight to the 60Hz as it can without triggering the mask filter without any known transient action (due to variations/fluctuations in the 60 Hz from harmonics on the line, etc.). 32

43 Figure 4.2 Mask Filter around 60 Hz voltage signal with offsets of 1ms, 100mV Whenever the voltage signal goes outside of the predetermined allowable range set by the mask filter, the current waveform stored in the buffer is automatically saved to the PC and labeled. Below in Figure 4.3, the voltage transient resulting from switching off a small desk fan is shown. The yellow portion of the transient is the part of the signal outside of the allowable region and triggered the software to save the signal for further processing. 33

44 Figure 4.3 Mask Filter catching a voltage transient from a small desk fan As the voltage transients are being captured, the transition of the load is recorded (i.e. On-to-Off or Off-to-On). As will be shown later, the transition information is important and can dictate the content and form of the voltage transient captured. Once multiple instances of like transients have been captured, they are ready for post-processing and analysis in MATLAB. 4.2 Signal Post-Processing After the voltage transients have been captured by the mask filter and saved by the oscilloscope software, some post-processing of the raw waveforms is needed to be done in MATLAB before the transients can be properly analyzed. The first step of post-processing is applying a notch filter to remove the 60 Hz component of the raw waveform. Since the 60 Hz component is stationary and present throughout 34

45 the entire waveform, it completely dominates the frequency domain and makes it virtually impossible to see any important information in higher frequencies which contain much less energy than the 60 Hz signal. After some initial frequency analysis of the power line, it was also necessary to filter out the odd harmonics from the 3 rd to the 11 th which had a significant presence in the signals. This allowed techniques of analyzing the time-frequency content to not be dominated by these always present harmonics which are shown in the following section. 4.3 Power Line Comparison Transient signals were captured at two different locations, one at a typical household in Blackwood, NJ and the other in Evans Hall in our office at the University of Delaware. As mentioned earlier in Chapter 3, the line does have some influence on the transient signals due from switching, although the values/impedances of the RLC circuit model are generally much smaller than the load s and may be dominated by the load. In order to see if the difference in the line conditions influences the voltage transient signals themselves, it is necessary to first compare the content of the different power lines themselves. Shown below in Figure 4.4 are the voltage signals of the power lines at the two locations. Notice the variations of the sinusoidal signals near the peaks, particularly in the office signal. Obviously, there is an influence of higher frequency harmonics which affect the power signal. Figure 4.5 shows the power spectral density of the two different signals, and Figure 4.6 shows a zoomed in version looking specifically at frequencies below 5 khz. 35

46 Figure 4.4 Voltage signal of the two different powerlines 36

47 Figure 4.5 Power Spectral Density of Powerlines 37

48 Figure 4.6 Powerline Power Spectral Densities under 5kHz 38

49 Power/Frequency The power spectral density graphs shown in Figure 4.5 show that all of the high frequency content on the MHz scale is under -100dB and does not have significant frequency content (mostly just from the noise). You can notice that it peaks at 60Hz as expected, but also have significant frequency content in the 0-5kHz range. By looking at the zoomed in version of the power spectral density of both of the powerlines it seems as if they are pretty similar, but Figure 4.7 below compares the two PSDs on a loglog scale. It shows that there is more harmonic content in the khz range for the office. This makes sense because the office powerline is on a system with far more machinery hooked up to it from adjacent labs, etc. and hence has more harmonic content PSD Comparison of Blackwood and Office Lines Blackwood Office Frequency (Hz) Figure 4.7 PSD Comparison of both Powerlines 39

50 4.4 Voltage Transient Analysis Using Short-Time Fourier Transform Since the goal of non-intrusive load monitoring using voltage transients is to try and develop a more general and simple way to detect load usage, it is necessary to analyze the signals and see how they change when they are recorded in a different location. However this is not so straightforward since there are variations and lack of repeatability of these voltage transients, but this will be discussed later in the chapter. First, looking at the time-frequency content of typical signals should be performed. Below in Figure 4.8, the top graph shows the recorded voltage transient of a small desk fan being turned off. The bottom portion is the same signal, but shown after notch filter processing of the 1-11 odd harmonics of 60Hz. The top portion of Figure 4.9 shows a zoomed in graph of the same voltage transient after notch filtering. Note: all transients shown after this will have been through notch filtering. The bottom half of Figure 4.9 shows the Short-Time Fourier Transform of this segmented voltage transient. The window used for the STFT was a Hamming window, with an overlap of half the window size, where blue represent low content and dark red is higher content. Notice the dark red veins of high frequency content during the long burst at the end of the signal. Now in order to see if these veins are a characteristic of this particular load, we need to examine multiple samples for repeatability. Figure 4.10 shows another transient of the same desk fan captured in the office and corresponding STFT. Notice that it still has the high frequency veins that are present during the long high frequency burst. It should also be noticed that these transients also contain frequencies higher than ~7.5 MHz, or half the sampling 40

51 frequency (which is MHz), since the frequency content is clipped. Unfortunately when the sampling rate is higher, our setup captures data at an even slower frame rate, so the odds of catching a transient become significantly worse. 41

52 Figure 4.8 Voltage transient of small desk fan in the office before (top) and after notch filtering (bottom) 42

53 Figure 4.9 Zoomed in graph of the same transient seen in previous figure (top) and the corresponding STFT (bottom) 43

54 Figure 4.10 Another voltage transient from a small desk fan at the office (top) and corresponding STFT (bottom) 44

55 In order to see if this characteristic of the high frequency veins in the long burst is always present when looking at a transient of this particular desk fan turning off, it should also be present when capturing a transient at a different location. Figure 4.11 and 4.12 show two different transients captured at the Blackwood location. The veins during the long burst are still present, but at different frequencies, a little less pronounced as before, and even have a slightly different shape. The difference in the power lines could be responsible for these differences, but could also just be due to variability in the signals themselves due to uncontrolled switching. Meaning that since this fan has a manual rotary switch the switching action can be different (i.e. faster, slower, harder, etc.) each time and could affect the resulting transient. Regardless of the differences, there does seem to be some characteristic nature seen in these voltage transients for the same load. They all have a few short bursts ending will the longest burst. Another similarity is that the veins all start at a higher frequency and decrease as time progresses. This could possibly be due to one of the circuit parameters of the load decreasing over time, such as the inductance of the motor. Also, while the transients different in length; they are all around a millisecond time scale. Looking more closely at the STFT graphs, while it is not obvious to see in all transients, the veins seem to repeat themselves (having the same characteristic shape) at higher frequencies, almost as if it is a harmonic of the characteristic resonant frequency. This can be seen really well in the office figures, but also in Figure Notice the darkest vein on the bottom and the one above it have very similar shapes as time progresses, almost as if the second harmonic has the same equation as the first but just scaled differently. 45

56 Figure 4.11 Voltage transient of small desk fan at Blackwood location (top) and corresponding STFT (bottom) 46

57 Figure 4.12 Voltage transient of small desk fan at Blackwood location (top) and corresponding STFT (bottom) 47

58 4.5 Voltage Transient Analysis Using Wavelet Transform As introduced in Chapter 3, there is another method for looking at the time-frequency content of a signal. The continuous wavelet transform, which is similar to the short-time Fourier transform, has the advantage of being a multiresolution method of analysis. Specifically, it has better resolution at high frequencies where high resolution is desired, and has a worse resolution at low frequencies, but high resolution at low frequencies is not typically needed. It is another way of looking at the time-frequency of our transients to see if characteristics of these transients will be revealed. As mentioned before, since the circuit model for the voltage transient changes when the switch is closing or opening (i.e. the line is part of the resonant circuit or not), the transient from the same load of turning it off and on should be different. If the contribution from the line is significant then major differences should be present in the frequency content of the transients for different operations. Figure 4.13 and 4.14 (top) show voltage transients from two instances of a lamp being turned on and the resulting scalograms (bottom) when using the cgau4 wavelet mother function over 32 scales. Notice the similarity in the signals and form in the scalogram. Figure 4.15 and 4.16 show two instances of the same lamp being turned off. The scalograms are noticeably of the same shape, but both are different than the other transition. Scalograms may be a way to unveil some inherent timefrequency characteristics of voltage transients, however it would be useful to compare the scalograms of the same transient using different mother functions, of which there is a myriad of possibilities. 48

59 Figure 4.13 Voltage transient (top) and resulting scalogram (bottom) of a lamp being turned on 49

60 Figure 4.14 Another voltage transient (top) and scalogram (bottom) for a lamp being turned on 50

61 Figure 4.15 Voltage transient (top) and scalogram (bottom) of a lamp being turned off 51

62 Figure 4.16 Another voltage transient (top) and scalogram (bottom) of a lamp being turned off 52

63 4.6 Motor Transient Detection and Classification As mentioned previously the physical properties and type of switch in a particular device can dictate the type of voltage transients created. This is especially true for devices that have motors with a commutating switch. In Chapter 3 the commutating switch was presented and shows how that an arc is created each time the brush goes over the gaps separating the commutator segments. This suggests that there should be a repetitive nature to the transients created by these commutating motors while they are in operation. In fact with universal motors you may think the transients would be periodic since the motor samples the 60Hz signal. In Figure 4.17 below, the top plot shows 6 cycles of voltage transients created when a simple household blender was in operation. One can see that the transients are present on every cycle of the voltage waveform while the device is under operation. This shows that universal motors like a blender have repetitive constantly present noise as suspected due to the commutating switch. However, in the bottom plot of Figure 4.17, the first three cycles are overlaid to see if the transients were indeed truly periodic, but they clearly are not. While they are not periodic, this shows that the voltage transients created by motors are clearly different than other devices with a one-time voltage transient created at the time of switching the device on or off, while the motors continually create transients while they are in operation. 53

64 Figure 4.17 Blender created Voltage Transients (top) and the first three cycles overlapped (bottom) 54

65 While the first attempt to differentiate motor transients from other devices by seeing if they repeated in a periodic way proved not to be true, the ever present nature of these transients can surely be used to show they are indeed significantly different than other types of devices that only create transients at the time they are turned on and off. In order to quantify this, a simple algorithm was written to calculate the number of mask failures over the six cycle length. In other words, a sliding window is used and at each segment the signal is analyzed and if there is any part in that segment where the magnitude is above the mask filter threshold (i.e. 100mV) then a failure is recorded. Ideally for the one-time transients this algorithm should always report a count of one, however some of the transients are very long in the time domain, and multiple windows would capture a mask failure of the same transient. On one hand you want the window size to be small enough to capture the separate small transients created from the motors, which should result in a higher count of mask failures, but if it is too small a very long one-time transient can report many failures. If the window is too long then the individual transients from the motor will not be counted. For example, if the window length was six cycles long, it would only record one mask failure, when in fact there are much more transients present that would trigger a mask failure. It should be noted that there is some variability with the signals captured and in fact sometimes the signal is triggered by random line noise which result in false data. Anyway, a number of signals were captured for a blender, fan, and a lamp. Originally, a window length of 500 samples with a mask filter threshold of 100mV was used to analyze these notch filtered transients. The distribution and comparison of means is shown below in Figure

66 Figure 4.18 Distribution and Means Comparison for Devices using a window length of 500 samples and mask threshold of 100mV This shows that with this small window length of 500 samples that the approximate mean failure count of the blender is 22, the fan is 19, and the lamp is 2. The comparison of means shows that the blender and fan are not significantly different. The lamp failure count mean makes sense in that it is a one-time device and should only have one transient in the six cycles (although the window is so small the length of the transients is longer, hence the mean of 2). For the fan you would still 56

67 expect the mean of failures to be near one, but the transients from the fan are very long and sometimes have a few pulses due to bouncing of the contacts, but still should be much lower than 19. Figure 4.19 below shows a voltage transient from a fan with multiple bounces x 10 5 Figure 4.19 Example of a notch filtered voltage transient from a fan with multiple contact bounces and long pulses In order to remedy this it is necessary to increase the size of the sliding window uses to take the mask failure count. Figure 4.20 below shows the distribution and comparison of means with the window length increased to 5000 samples. 57

68 Figure 4.20 Distribution and Means Comparison for Devices using a window length of 5000 samples and mask threshold of 100mV The results now show that the blender is in fact significantly different than the fan and the lamp. So this shows that by analyzing the number of mask failures in a number of 60Hz cycles (6 in this case) and comparing the distributions and more specifically the means over a number of data points, that motor operation can in fact be determined. For the overall system, one way this could be implemented would be to capture the transients as already mentioned before with the mask filter, but once the 6- cycle signals are triggered and stored, use this smaller sliding window technique to 58