Load Recognition for Different Loads with the Same Real Power and Reactive Power in a Non-intrusive Load-monitoring System Hsueh-Hsien Chang1, Ching-Lung Lin2, Hong-Tzer Yang3 1 Dept. of Electronic Engineering, Jin Wen University of Science and Technology, Taipei, Taiwan.
[email protected] 2 Dept. of Electrical Engineering, Ming Hsin University of Science and Technology, Hsinchu, Taiwan.
[email protected] 3 Dept. of Electrical Engineering, National Cheng Kung University, Tainan, Taiwan.
[email protected] Abstract This paper proposes the use of power signature to recognize different loads with the same real power and reactive power in a non-intrusive load-monitoring (NILM) system. To test the performance of the proposed approach, the data sets for electrical loads were analyzed and established using an electromagnetic transient program (EMTP) and onsite load measurement. Load recognition techniques were applied in a neural network. The effectiveness of load recognition and the time requirement were analyzed and compared using a back propagation classifier method. The experiments revealed that analyzing the turn-on transient energy signatures can enhance the efficiency of load recognition, particularly for different loads with the same real power and reactive power in a NILM system, and improve ability of computational speed. Keywords: Load Recognition, Neural Network, Non-Intrusive Load Monitoring, Electromagnetic Transient Program.
1. Introduction Traditional load-monitoring instrumentation systems employ meters for each load to be monitored. These meters may incur significant time and cost to install and maintain. Plus, with the number of meters increase, reliability of the measuring system may decrease. Therefore, a method for minimizing the number of instruments using non-intrusive load monitoring (NILM) is expected. To develop such a system, a number of load recognition techniques have been proposed [1-13]. The present study proposes a feature selection method for a NILM system, using the turn-on transient energy feature (UT) of analyzing the physical load signatures of the appliance, to recognize different loads
_____________________________________ 978 -1-4244-1651-6/08/$25.00 © 2008 IEEE 1122
with the same real power (P) and reactive power (Q). Many common loads are identifiable by matching their load profiles with known transient and steady-state shapes or templates of common loads. These templates can be collected on-site during a learning or installation phase or generated off-side by a computer-aided tool such as an electromagnetic transient program (EMTP). To maximize recognition accuracy, analysis of turn-on transient energy signatures uses a window of samples, Δt, to separate adaptively a transient representative class of loads. The effectiveness of load recognition and the time requirement were analyzed and compared using a back propagation classifier method. The experiments reveal that analyzing the turn-on transient energy signatures can enhance the efficiency of load recognition, particularly for different loads with the same real power and reactive power in a NILM system, and improve ability of computational speed. This paper is organized as follows. The power signature problems and turn-on transient energy algorithm are addressed and described in Section 2. The load recognition by neural network is described in Section 3. The turn-on transient energy repeatability is described in Section 4. Based on the algorithm and load recognition technique, a series of load recognition experiments for the feature selection of power signals are conducted in Section 5, which also includes comparisons of recognition accuracy and computational performance using selected features by turn-on transient energy algorithm and traditional power signal features, for example, real power and reactive power. Advantages of load recognition for different loads with the same real power and reactive power, using the turn-on transient energy algorithm, are concluded in Section 6.
2. Power signature problems and turn-on transient energy algorithm 2.1. Problems using power signatures for different loads with the same real power and reactive power In general, an appliance may have many load representations and a load may involve many physical components. For example, a dryer has two loads, a motor and a heater. A refrigerator has only one load, a compressor, but has different physical components for defrosting and freezing. Most appliances have certain unique power signatures that enable them to be distinguished from each other. These unique power signatures can be observed from voltage and current waveforms supplied to the appliance, or from processed reproductions of these signals such as the delivered real power and reactive power or harmonics [2]. According to the switch continuity principle, steady-state signatures, for example, real power and reactive power, are additive when two signatures occur simultaneously. In contrast to steady-state properties, transient properties are not addition [1]. Distinguish different loads may be problematic when they have equivalent real power and reactive power but no harmonic components, and/or when the sums of real power and reactive power of two load types are equal to that of another load during multiple load operations.
lighting, respectively [14]. During training phases, a window of samples of time length Δt is examined to separate transients representative of a class of loads. This separation process delineates a set of transient energy values representing a particular transient shape in each of the input envelopes. ∧ To maximize the recognition accuracy ( γ ) during the test phase, Δt is adaptively changed based on factor δ. Figure 1 shows the adaptive algorithm for Δt by factor δ. As demonstrated in Eq. 1, the search for a precise time pattern for instantaneous power turn-on identifies a complete transient. ts + Δt
∫ v ⋅ i ⋅ dt
=
W transient
Pins tan tan eous ( t ) =
dWtransient ( t ) = v( t ) × i( t ) dt
(2)
∧
v, i,Δt, γ initial Turn-on Transient Event Detector for ts
W
∫
=
transient
t s + Δ t
v ⋅ i ⋅ dt
t s
Energy Measurement
Δt = Δt + δ
δ < Δt
Neural Network Algorithm
2.2. Turn-on transient energy algorithm ∧
The transient properties of a typical electrical load are mainly determined by the physical task that the load performs [4-8]. Transient energy may assume different forms in consumer appliances, depending on the generating mechanism [1]. Estimating current waveform envelopes at the utility service entry of a building, for examples, enables accurate transient event detection in the NILM [8]. Load classes performing physically different tasks are therefore distinguishable by their transient behavior [4-8]. Since the envelopes of turn-on transient instantaneous power are closely linked to telltale physical quantities, they serve as reliable metrics for load recognition. Two different appliances consuming identical levels of real power and reactive power may have very different turn-on transient currents. Analysis of these transient currents can determine which of the two is actually present in the load. In general, the transient behavior of many important loads is sufficiently distinct reliably to recognize load type. The long characteristic switching-on transient, the less substantial switching-on transient, the short but very high-amplitude switching-on transient and the long two-step switching-on transient are the principal values measured in pump-operated appliances, motor-driven appliances, electronically fed appliances and fluorescent
(1)
ts
No
γ >γ
Yes ∧
γ ⇐ γ
Figure 1. Adaptive algorithm for Δt by factor δ at the maximum recognition accuracy for load recognition.
3. Load recognition by neural network Pattern classifiers partition multidimensional space into decision regions indicating to which class any input belongs [15]. Many classification techniques have been developed for load recognition. The non-parametric and learning-supervised classifier is adopted for electrical patterns of commercial or industrial appliances because the distribution of these patterns is quite complicated without any formulation, and the larger loads can be easily and clearly labeled. Most back-propagation neural network applications employ single- or multi-layer perceptron networks using gradient-descent training techniques, with learning by back propagation. These multi-layer perceptrons can be trained with supervision using analytical functions to
1123
activate network nodes (“neurons”) and by applying a backward error-propagation algorithm to update interconnecting weights and thresholds until proper recognition capability is attained [15]. In the present study, the back-propagation classifier is generally used as a trainable classifier. “Classification” in this context denotes a mapping from a feature space to the set of class labels – the names of commercial or industrial load combinations.
4. Turn-on transient energy repeatability
Δt
Figure 2. Current waveform in one phase at voltage phase 0o for the turn-on transient of a three-phase 300-hp induction motor. Δt
Most loads observed in the field have repeatable transient profiles or at least repeatable sections of transient profiles [5]. The load survey indicates that non-linearity in the constitutive relationships of the elements comprising a load model and/or in the state equation describing a load tends to create interesting and repeatable observable turn-on transient profiles suitable for use in identifying specific load classes [4-8]. Because of the varying transients (which often depend on the exact point in the voltage cycle at which the switch opens or closes), data sets for load recognition must provide accurate repeatability of the turn-on transient energy signatures. Pre-training has proven effective for highly repeatable loads that show up in large quantities, such as commercial and industrial loads. As Figures 2 and 3 demonstrate, the turn-on transient profiles exhibit repeatable measured current waveforms in one phase at voltage phase 00 and 900 for the turn-on transient of a three-phase 300-hp induction motor. The turn-on characteristics of a load clearly increase in complexity over time. Closer investigation of the load turn-on is thus required before the characteristics can be used as a distinguishing feature of a load. This information, collected via non-intrusive monitoring, can be used to answer important questions about the statistical validity of power measurements. Determination of whether or not turn-on transient energy content is repeatable would be useful in developing a turn-on transient energy signature. As Eq. (3) shows, the average value of the sample data ( xi ) for the turn-on transient energy of each load is x . The standard deviation (S) of the turn-on transient energy for each load is computed according to Eq. (4) for all loads monitored in isolation. An experiment [3] was used to demonstrate that the statistical validity of the turn-on transient energy for each load is repeatable in terms of the coefficient of variation (C.V.) according to Eq. (5). 1 n (3) x = ∑ xi n i=1 n 1 (4) 2 s =
n − 1 s C .V . = x
1124
∑ (x i − x ) i=1
(5)
Figure 3. Current waveform in one phase at voltage phase 90o for the turn-on transient of a three-phase 300-hp induction motor.
5. Experimental results 5.1. Case study environment 5.1.1. Case study 1: EMTP simulation. In case study 1, the NILM system monitors voltage and current waveforms in a three-phase electrical service entry powering a collection of loads representative of the major load classes in a commercial building. The neural network algorithm in the NILM system identifies three loads with transient and steady-state signatures operating at the 220-V common bus. These loads include a 2.6-hp induction motor, a 4.7-hp induction motor and an R-L linear load with real power and reactive power equivalent to that of a 4.7-hp induction motor. Figure 4 schematically illustrates the test stand used in case study 1. Three-phase 220-V electricity powers the loads, which are representative of important load classes in a commercial building. To compile data for training purposes, either every site of interest or a representative sample of the sites should be monitored. The sample rate is approximately 15 kHz. 5.1.2. Case study 2: Experiment. In case study 2, the NILM system is used to monitor voltage and current waveforms in a one-phase electrical service entry powering loads representative of important load classes in the laboratory. The neural network algorithm in the NILM system identifies three actual loads with transient and steady-state signatures at the 110-V common bus. These loads include a 119-W dehumidifier, a 590-W vacuum cleaner and an R-L linear load with real power and reactive power equivalent to that of a 590-W vacuum cleaner.
Host PC
1
3φ Voltage Measurements
R-L Linear Lad
3493W
3φ220V
3φ220V
4.7Hp I/M
3
2
1
Substation
Non-intrusive Load Monitoring System
2.6Hp I/M
was executed on an IBM PC with an Intel 1.5-GHz Pentium M CPU for load recognition. 3φ220V
Figure 5 schematically illustrates the test stand used in case study 2. The loads are powered by one-phase 110-V electricity and are representative of important load classes for laboratory test purposes. A dedicated computer connected to the circuit breaker panel controls the operation of each load. The computer can also be programmed to stimulate various end-use scenarios.
Output
Hidden
6
3φ Current Measurements 3
The Reactive Power
The Turn-on Transient Energy
Input
2
1
Host PC
Substation
2
1
R-L Linear Load
Figure 4. Electrical schematic of NILM in a commercial building in case study 1.
Figure 6. Neural network model designed for multiple load operations in case study 1. 1φ110V 590W
3φ220V 3493W 2221VAR Y-Connection R =9.7421Ω L=16.474mH
Vacuum Cleaner
3φ 220V 3493W 2221VAR 4.7HP I/M 4poles
3φ 220V 2.6HP I/M 2poles
1φ110V 590W
3
2
Dehumidifier
1
1φ110V 119W
Local PC
The Real Power
220V Bus
3
Output
1φ Voltage Measurements
Non-intrusive Load Monitoring System
1φ Current Measurements 1
Hidden
6
110V Bus
1φ110V 119W 43VAR Dehumidifier
2 1φ110V 590W 233VAR Vacuum Cleaner
3 1φ110V 590W 233VAR Y-Connection R =18.16Ω L=19.037mH
Figure 5. Electrical schematic of NILM in the laboratory in case study 2.
5.2. Case study results Changing circuit breaker number 1 to number 3 in multiple operations yields data for multiple loads. Seven combinations are possible. Each data set includes a voltage variation from –5% to +5% at 1% intervals. The total number of data sets is 77 (7×11). To confirm the inferential power of the neural network, the data are categorized as 39 learning and 38 test data sets. Notably, learning data and test data are selected randomly from all data. For this study, a neural network simulation program was designed using MATLAB. This program
3
Input
The Turn-on Transient Energy
1
2
The Reactive Power
Local PC
The Real Power
1
Figure 7. Neural network model designed for multiple load operations in case study 2. Figures 6 and 7 demonstrate the structure of the neural network for multiple load operations in case study 1 and in case study 2, respectively. The neural network consists of three layers. The number of input neurons is three, to which the real power, reactive power and the turn-on transient energy have been assigned. The number of output neurons is three, to which the identified individual appliances have been assigned. The number of cells in the hidden neural network layer is determined by the complexity of input–output mapping; for instance, the number of hidden neurons is the total number of input neurons and output neurons. The hidden neurons are sufficient for inputs and outputs of the neural network.
1125
5.2.1. Case study 1: EMTP simulation. Table 1 shows that values for the training and test recognition accuracy of load recognition in multiple operations are all 100% for features with real power and reactive power, as well as the turn-on transient energy (PQUT). However, the training and test recognition accuracy of load recognition in multiple operations are only 58.97% and 39.47%, respectively, for features with real power and reactive power (PQ). Those loads cannot be identified by real power and reactive power features because the second load and the third load are different loads with the same real power and reactive power, as are combinations of the first and second loads and combinations of the first and third loads. In other words, test recognition for those loads in multiple operations is quite low when using only real power and reactive power features. Table 1 Recognition accuracy and the number of hidden neurons of load recognition in case study 1 Loads Features PQ PQUT Numbers Numbers Training Test Training Test of of test training 1 2 3 1+2 1+3 2+3 1+2+3 Summary Recognition accuracy (%) The number of hidden neurons
5.2.2. Case study 2: Experiment. Table 3 shows that values for the training and test recognition accuracy of load recognition in multiple operations are also all 100% for features with real power and reactive power, as well as the turn-on transient energy (PQUT). However, the accuracy of training and test recognition of load recognition in multiple operations are only 51.28% and 39.47%, respectively, for features with real power and reactive power (PQ). Those loads can not be identified using only real power and reactive power features because the second load and the third load are different loads with the same real power and reactive power, as are combinations of the first and second loads and combinations of the first and third loads. In other words, the presence of different loads with the same real power and reactive power can be confirmed in two ways. First, test recognition in multiple operations is quite low when only using features of real power and reactive power. Second, the turn-on transient energy for one of the features can improve load recognition, especially for different loads with the same real power and reactive power.
5
6
5
5
5
6
6
5
2
0
6
5
6
5
3
0
6
5
5
6
1
0
5
6
5
6
0
0
5
6
6
5
6
5
6
5
6
5
6
5
6
5
39
38
23
15
39
38
1
58.97
39.47
100
100
2
Table 3 Recognition accuracy and the number of hidden neurons of load recognition in case study 2 Loads Features PQ PQUT Numbers Numbers Training Test Training Test of test of training
3 2+3
3+3
Table 2 Computation time and the number of iterations of load recognition in case study 1 Features PQ PQUT Training time 83.65 8.9030 (Sec.) Test time 0.4010 0.4110 (Sec.) The number of a 10000 973 iterations a The maximum number of iterations is set at 10000 by the neural network program.
Because of the algorithm can not be converged for the features only with real power and reactive power. Table
1126
2 shows that values for the training time of load recognition in multiple operations are much more for the features only with real power and reactive power than that for the features with real power and reactive power, as well as turn-on transient energy.
1+2 1+3 2+3 1+2+3 Summary Recognition accuracy (%) The number of hidden neurons
5
6
5
5
5
6
6
5
1
0
6
5
6
5
1
0
6
5
5
6
1
0
5
6
5
6
0
0
5
6
6
5
6
5
6
5
6
5
6
5
6
5
39
38
20
15
39
38
51.28
39.47
100
100
2+3
3+3
Because of the algorithm can not be converged for the features only with real power and reactive power. Table 4 shows that values for the training time of load recognition in multiple operations are much more for the
features only with real power and reactive power than that for the features with real power and reactive power, as well as turn-on transient energy. Table 4 Computation time and the number of iterations of load recognition in case study 2 Features PQ PQUT Training time 80.916 3.765 (Sec.) Test time 0.4110 0.4010 (Sec.) The number of 10000a 75 iterations a The maximum number of iterations is set at 10000 by the neural network program.
6. Conclusions and future work In multiple operations, a class shows that the representation can be one or many loads. In other words, a class may be a combination of more than one load. Therefore, classifications are more complicated, especially when identifying different loads with the same real power and reactive power. The EMTP simulation is invaluable for testing pattern recognition samples and enables the rapid development and implementation of successful prototypes. To improve the recognition accuracy within multiple operations for different loads with the same real power and reactive power but no harmonic components, features cannot be adequately measured only from steady-state parameters, i.e., real power and reactive power. In contrast to steady-state properties, transient properties such as the turn-on transient energy can play an important role. Combining transient and steady-state signatures is necessary in identifying different loads with the same real power and reactive power but no harmonic components. The recognition accuracy for these features is almost 100%. We do not assess the performance of the Neural Network for this classification task because of absence of benchmarking in the text. In the future, some previous works from other published papers have been mentioned in the text, it will be worthwhile if comparison with past work done is showed.
[4]
[5]
[6] [7]
[8]
[9]
[10] [11]
[12]
[13]
[14] [15]
International Conference on CSCWD 2007, pp. 1022-1027, 26-28 April 2007. K. D. Lee, S. B. Leeb, L. K. Norford, P. R. Armstrong, J. Holloway, S. R. Shaw, "Estimation of variable-speed-drive power consumption from harmonic content," IEEE Trans. Energy Conversion, vol. 20, No. 3, pp. 566-574, 2005. C. Laughman, K. Lee, R. Cox, S. Shaw, S. B. Leeb, L. Norford, and P. Armstrong, "Power Signature Analysis," IEEE Power & Energy Magazine, pp. 56-63, March/April 2003. U. A. Khan, S. B. Leeb, and M. C. Lee, "A multiprocessor for transient event detection," IEEE Trans. Power Delivery, vol. 12, No. 1, pp. 51-60, January 1997. L. K. Norford, S. B. Leeb, "Non-intrusive electrical load monitoring in commercial buildings based on steady-state and transient load-detection algorithm," Energy and Buildings, 24, pp. 51-64, 1996. S. B. Leeb, S. R. Shaw, and J. L. Kirtly, Jr., "Transient event detection in spectral envelop estimates for nonintrusive load monitoring," IEEE Trans. Power Delivery, vol. 10, No. 3, pp. 1200-10, July 1995. D. Srinivasan, W. S. Ng, A. C. Liew, "Neural-network-based signature recognition for harmonic source identification," IEEE Trans. Power Delivery, vol. 21, No. 1, pp. 398-405, 2006. M. Baranski, J. Voss, "Genetic algorithm for pattern detection in NIALM systems," IEEE international conference on Syst Man Cy., pp. 3462-3468, 2004. D. C. Robertson, O. I. Camps, J. S. Mayer, and W. B. Gish, "Wavelets and electromagnetic power system transients," IEEE Trans. Power Delivery, vol. 11, No. 2, pp. 1050-1057, April 1996. A. I. Cole, A. Albicki, "Nonintrusive identification of electrical loads in a three-phase environment based on harmonic content," In Proc. IEEE Instrumentation and Measurement Technology Conf., pp. 24-29, 2000. M. L. Marceau, R. Zmeureanu., "Nonintrusive load disaggregation computer program to estimate the energy consumption of major end users in residential buildings," Energy Conversion and Management, 41, pp.1389-1403, 2000. F. Sultanem, "Using appliance signatures for monitoring residential loads at meter panel level," IEEE Trans. Power Delivery, vol. 6, No. 4, pp. 1380-1385, Oct. 1991. R. P. Lippmann, "Pattern Classification using Neural Networks," IEEE Communications Magazine, pp. 47-64, November 1989.
References [1] G. W. Hart, "Nonintrusive appliance load monitoring," in Proc. 1992 IEEE Conf., vol. 80, pp. 1870-1891. [2] J. G. Roos, I. E. Lane, E. C. Botha, and G. P. Hanche, "Using Neural networks for non-intrusive monitoring of industrial electrical loads," in Proc. 1994 IEEE Instrumentation and Measurement Technology Conf., pp. 1115-1118. [3] H. T. Yang, H. H. Chang, C. L. Lin, "Design a Neural Network for Features Selection in Non-intrusive Monitoring of Industrial Electrical Loads," Computer Supported Cooperative Work in Design, 11th
1127
