Home Energy Management System Using NILM and Low-Cost HAN

Qasim Khalid, Naveed Arshad, Nasir Khan, Taha Hassan, Fahad Javed, and Jahangir Ikram

1. Introduction

Most of the energy consumed in the world today is generated by using fossil fuels, such as coal, oil, and natural gas. These sources of energy have a detrimental impact on the environment because of carbon dioxide (CO2) and other greenhouse gases. For example, according to the Department of Energy, USA, to generate one kilowatt-hour(kWh) by coal, a staggering amount of more than two pounds of CO2is released to the atmosphere. This is the amount of carbon dioxide (CO2) produced from high efficiency plants (furthermore low efficiency plants are even worse). To reduce our dependence on fossil fuels,more renewable sources of energy like wind, solar, and hydro are being introduced. However, our mainstay of energy production for many years is hinged on fossil fuels.Therefore, energy efficiency is considered as the most important fact for the future. Of all the electricity generated in the world, buildings consume more than 40% of the energy[1]. Studies have shown that up to 20% savings in their energy consumption could be achieved by just communicating the real time energy information to the consumers[2]. Further to this, home energy management systems (HEMs) are being developed that provide more fine-grained control of the devices in buildings. An essential component of HEMs is a home area network(HAN) that forms a communication link between the household consumer and the installed devices[3]. As shown in Fig. 1, a typical HAN architecture involves installing sensors at devices that communicate to a central node or form a mesh network, which includes two components: i) a network for transceiving a signal for communication and ii)user-end devices such as washing machines, televisions,and refrigerators etc. The central part in HAN is the network that connects all the smart devices in a particular area (home). In such a network various technologies can be used, for example ZigBee, Z-wave, and Wi-Fi. Although the costs of HEMs and HANs have reduced in recent years,their cost is still prohibitive enough to hold back a mass deployment especially in developing countries.

Fig. 1. Typical HAN model.

Therefore, in this paper we present the idea of using non-intrusive load monitoring (NILM) in combination with a very low cost radio frequency-based single-waycommunication HAN to develop a home energy management system. NILM is the disaggregation of aggregate demand profile (ADP) of a household into individual signatures of appliances switched on at a particular instant or within a specified time period. This disaggregation is carried out without any intrusive physical sensors on individual appliances. ADP or instantaneous power consumption is typically observed via a single energy meter installed at the service entry point of the household.

2. Related Work

In this section we present the related work with respect to HANs and non-intrusive load monitoring.

2.1 HAN

HAN is used for communication with appliances,which can be classified into two categories: a) wired HAN and b) wireless HAN.

a) Wired HAN—It is that in which the communication takes place through electric wire. This wire can be telephone line cables, optical cables, coaxial cables, twisted cables, ribbon cables, or Ethernet cables. Following are the different technologies for wired HAN.

· Ethernet: It is based on IEEE 802.3 Standard. It is a very popular technology and has largely replaced all other competing wired local area network (LAN) technologies.Firstly, it was introduced with coaxial cables but later it was replaced by twisted pairs and optical fibers. Now, its data rates ranges from 10 Mb to 100 Gb per second. In HAN, this technology can be used easily with computers,laptops, liquid crystal display (LCD) screens, printers,scanners, and gaming consoles but cannot be readily available with other appliances, like washing machines, air conditioners, refrigerators etc. Therefore, high cost and power is needed to deploy Ethernet in all other appliances to create HANs. It is being used widely.

· HomePlug: It has IEEE P1901 Standard. It is a power line communication technology that allows networking over the existing home electrical wiring. It focuses on low power and throughput for applications, for example smart meters as well as in HANs, too. Its max speed is 14 Mb to 200 Mb per second and its range is about 300 m. It is being used at a moderate level.

· X10: It is based on X10 Standard. It is an industry standard and a protocol, which is developed by X10 Wireless Company. It is used for communication between electric appliances. It also allows remote control of home devices and appliances. That is why it uses brief radio frequency (RF) signals in the forms of bursts which represent digital information. But it has some shortcomings,such as hard to install new appliances in a network, slow speed, larger overheads, and absence of encryption. Its max speed is about 50 Kb to 60 Kb per second and its range is about 300 m. It is being used at a moderate level.

· Insteon: It is based on X10 Standard, too. It is a technology by which we can connect lights, fans, and all appliances without any extra wirings. It has its own protocol which permits all devices to connect and communicate in peer to peer, so that, each device can transmit and receive the messages directly without any master controller. Its max speed is about 1.2 Kb per second and its range is about 3000 m. It is being used at a moderate level.

· ITU G.hn: It is based on G.hn Standard. It is a pure home network technology developed by the International Telecommunication Union. It can support data networking over all types of cables, specially coaxial cables, power cables, and telephone lines. Its data rate is up to 1 Gb per second. It gives secure connectivity with Internet Protocol version 4 (IPv4) as well as Internet Protocol version 6(IPv6). The benefit of G.hn over all other wired technologies is that it is compatible with all types of wires and needs no extra installation and equipment. It is used rarely because it was not included in the final NIST Roadmap for Smartgrid Interoperability Standard.

b) Wireless HAN—In a wireless network, the most desirable attributes for communication technologies are that they should be low cost and consume low power (must be battery operated). These technologies can be used in HANs for security, monitoring, and automatic controlling of devices. Following are the different technologies for wireless HANs.

· ZigBee: ZigBee is a wireless network with a mesh topology based on IEEE 802.15.4 Protocol. In the mesh technology, every node is interconnected to each other such that there are multiple paths of communication between any two nodes. Each node is self-routing and routing information is updated dynamically. ZigBee provides high reliability and a wider range. Its max speed is about 250 Kb per second at 2.4 GHz and 40 Kb per second at 915 MHz frequency band. The typical range of ZigBee with mesh networking is about 10 m to 75 m.

· Z-wave: Z-wave is an authorized communication protocol. It is specially designed for controlling appliances in homes and medium sized buildings. It is not an open source but still widely used. It uses 900 MHz frequency band which is unlicensed in many countries. It is better than 2.4 GHz frequency band because working in the 900 MHz frequency band provides a wider range and consumes lower power. Its max speed is 40 Kb per second and the typical range is about 30 m in open environments.

· Wi-Fi: Wi-Fi is a well-known wireless technology which is generally used with computers, mobile phones,and now in HANs. It is based on IEEE 802.11 Standard. It is an internet protocol (IP) based protocol which delivers data signals from the transmitter to the receiver in a packet switched internetwork. Due to interoperability of Wi-Fi, it can be easily use for communication of HANs & the smart grids architecture. But the range of the Wi-Fi technology is its main limitation factor. Its max speed is up to 300 Mb per second and its range is about 100 m.

· OneNet: It is an open source standard for wireless networking. It is also an IP based technology. We can use it with readily available RF modules and microcontrollers made by any manufacturer as our choice. It is a low cost and simple technology. Its speed is about 38 Kb per second to 230 Kb and its range is about 60 m.

· 6LoWPAN: It is the abbreviation of IPv6 low power wireless personal area network. It is a standard from Internet Engineering Task Force (IETF), which is based on the idea that the IP address should be applied to every device even to smallest ones. It uses the IEEE 802.15.4 Standard and transfers data in the form of packets with a very low bandwidth. It allows communication with appliances with the help of Internet without using ZigBee to IP addresses conversion. Its range is about 10 m to 75 m.

2.2 Non-Intrusive Load Monitoring (NILM)

Contemporary research on implementation of a NILM system typically addresses the following design choices.

· Granularity over time of ADP: Granularity over ADP refers to the rate at which the installed meter is able to observe and report the overall instantaneous power consumption. This design choice is usually associated with the distinction between event-based and non-event based operation. In event-based operation, the decision frequency for NILM is the frequency at which appliances in the household change their states[4]–[6]. This choice is appropriate for real-time diagnostic feedback in residential and commercial energy feedback systems (REFS and CEFS)and is considered for this paper. Non-event based NILM operates on aggregated consumption data and is useful for long-term or medium term diagnoses on energy consumption of particular households or for larger settings such as micro-grids and distribution sectors, particularly as an enabling technology for demand-side management.

· Def i nitions of load signatures: The def i nitions of load signatures refer to the metrics used to identify operating characteristics of individual devices connected within the setting in question. Various existing implementations of NILM systems primarily differ in choice of steady-state or transient, fundamental frequency or harmonic frequency signatures. For instance, [7] and [8] employed steady-state and transient changes of active real power as the basis of classif i cation, respectively. In contrast, [6] and [5] used harmonic magnitudes associated with step changes in the overall load for establishing and comparing appliances prof i les. Liang and colleagues[4]have demonstrated the use of the raw single-cycle current, instantaneous power, and admittance waveforms for the purpose. A comprehensive review of signature types evaluated for NILM appeared in[9] and [10]. Reference [11] studied typical bases for load taxonomies and introduced wave-shape features (WS) as a competitive new basis using hierarchical clustering.

· Initial acquisition of load signatures: Hart’s canonical work refers to NILM two distinct modes of operation, with different degrees of intrusiveness[7]. The first uses a one-time calibration period where appliance signatures are manually collected for supervised learning algorithms. The mode is referred to as manual-setup NILM (MS-NILM) or semi-automated energy disaggregation and is considered for this paper. References [4], [7], and [8] are typical implementations of MS-NILM. The second operational mode for NILM, termed automatic-setup NILM, uses a priori information about expected load characteristics and unsupervised machine learning algorithms to automatically disaggregate ADP. References [12] and [13] are recent investigations of unsupervised energy disaggregation.MS-NILM is computationally convenient, however, the intrusive collection and labeling of signatures makes it tedious to set up and adapt to new appliances.

2.3 Selection of Learning and Optimization Algorithms

This refers to the parameter search and performance optimization of learning algorithms, for instance, the artif i cial neural network or support vector method employed to learn appliance signatures. This choice of algorithms for disaggregation is also fundamentally relative to both the required response time for a NILM system and the choice of load signatures. A wide variety of supervised and unsupervised learning algorithms[4],[8],[12]–[15]and optimization strategies, such as integer programming and metaheuristics[16]–[19], have been observed effective in their capacity for load prof i ling and disaggregation. This capacity is typically distinct relative to load conditions at the time of operation, for instance, the number of simultaneously operating appliances, noise levels, electrical interference from the neighboring system[4].

3. Proposed Model

The proposed HEM model uses energy usage and device status information from the smart energy meter installed in the building. A very low cost RF-based HAN employs one-way communication to send the signals to devices to change their status. The proposed model is shown in Fig. 2.

As shown in Fig. 2, a smart meter is installed at the service entry point of the building. Any usage of electricity is being captured at this point. Every device in the house has its unique signature as well as the load that can be observed when the device is turned on.

Fig. 3 is a set of eight different devices have different load signatures when plotted as a voltage vs. current graph at 16 kHz frequency. In our model we have used these data at 16 kHz to train a machine learning algorithm commonly known as adaptive boost. With adaptive boost we have achieved 99% accuracy. Once the energy is disaggregated using NILM, the status of devices and their power consumption can be calculated easily. This information is used by HEM to send control signals by the HAN. To reduce the cost of developing, we have used a one-way RF-based HAN. Following are the design and implementation details of HAN.

Fig. 2. Proposed HEM.

Fig. 3. Current-voltage (VI) signatures from eight devices.

Fig. 4. Master circuit test model.

Fig. 5. Master circuit block diagram.

Fig. 6. Slave circuit block diagram.

Fig. 7. Slave circuit block diagram.

Fig. 8. Model block diagram.

3.1 One-Way RF-Based HAN Implementation

To develop a low-cost HAN we employ a low power radio transceiver which is the least expensive among the available radio transceivers. The main advantage of this model is that we can use low power RF transceivers to cover a larger area and floors. Our model is divided into two main parts.

· Master circuit: The master circuit acts as a transmitter.It is connected to a computer through the serial cable and we can send control signals from the master module to the slave module by using one-way communication. The control circuit is composed of a common use transceiver connected to a programmable interface controller (PIC)which is further connected to a computer as shown in Fig. 4 and Fig. 5.

· Slave circuit: The slave part of the circuit receives the control signal from the master module and acts accordingly as shown in Fig. 6 and Fig. 7.

The master circuit transmits a command signal (turn on and turn off) to the slave circuit using a mesh network for communication. We use one-way communication between the controller and the slave circuit which help us reduce the cost of communication and power consumption. We have implemented our proposed model by designing the test model as shown in Fig. 8.

Our test model includes PIC 16F877A, PIC 16F676,and MX232 IC for serial communication with the computer,RF module, relays (instead for real appliances),opto-couplers (for isolation of circuits), crystal oscillator,transistors, and voltage regulators integrated circuit (IC).The RF module uses the 434 MHz RF receiver module,which is readily available and commonly used in car door locking system. It is secure as it has a unique hatch code number thus making it optimal and economical for HAN security and implementation.

Table 1: HAN test results for distance covered

Table 2: HAN results for penetration along rooms, floors and walls

Table 3: Median training/test/overall precision of prediction (%)with reference energy disaggregation dataset[11]. Rows represent employed benchmark load signatures (LS); columns represent learning algorithms used for classif i cation/prediction.

4. Results

Since our proposed model consists of both HAN and NILM, we have evaluated them in the following two sections separately.

4.1 HAN Results

We tested the one-way HAN using different scenarios.Our criteria for testing the HAN are 1) the distance it can cover, 2) the number of walls it can penetrate, and 3) the number of floors it can penetrate.

We tested the model in a concrete floored multi-story building. The following Tables 2 and 3 summarize the performance matrix for the said model.

Our HAN works without any problem for 80 feet across all rooms, floors and walls. Beyond 80 feet the performance is dependent on the hindrances. This is an excellent result considering that our total HAN cost does not surpass 30 dollars for one master circuit and four slave circuits. Any solution from ZigBee and Z-wave is easily over 100 dollars even for minimal two node implementation tested in a concrete floored multi-story building.

4.2 NILM Results

To assess the energy disaggregation using NILM we have used reference energy disaggregation dataset (REDD)that is a publicly available dataset containing detailed energy usage information of several households over extended periods and in two granularities[20]. The low granularity data is the average real power consumption of multiple households (both the mains and individual circuits) at a frequency of approximately 1 Hz for the mains and 0.33 Hz for individual circuits. High granularity data is the AC voltage and current waveform data from household mains acquired using commercial load monitors at a frequency of 16.5 kHz. We employed four machine learning algorithms namely artificial neural network (ANN), artificial neural network + evolutionary algorithm (ANN+EA), support vector machines (SVM), and adaptive boost (AdaBoost)across three variations of load signatures (LS) namely power quality (PQ), wave shape (WS) and harmonics (HAR). Table 3 lists training, test, and overall precision of prediction for ANN, ANN+EA, SVM, and AdaBoost with decision stumps(on an Intel Core i7 machine, CPU clock 3.1 GHz, 8 GB of RAM) over PQ, WS, and HAR. Adaptive boost exhibits that superior precision of prediction is relative to ANN,ANN+EA, and SVM for original switching events extracted from REDD.

5. Discussions

The cost of the proposed system is very low. In most houses it is assumed that a smart meter is already installed.These smart meters are capable of providing instantaneous energy consumption. Therefore, the real cost for developing this system is the cost of developing the one-way HAN.Since we have used RF communication, this cost is very low. In laboratory settings we have developed this HAN for four devices using around $30. The only downside of the system is that NILM requires a learning time where machine learning algorithms are trained in the household devices. With a new device the system may needs to be trained on its voltage current (VI) signatures.

6. Conclusions

In this paper we have proposed a new model for implementing HEMs. This HEM is unique in a sense that it uses the total energy consumption data of the house to disaggregate the load profiles of various electric devices.By using a one-way communication HAN, the HEM can communicate with the electric devices.

We have shown in this paper that NILM and a low cost one-way RF could reduce the cost in HEMs further. This will enable large-scale deployments of HEMs which will further reduce energy usage. Ultimately any energy that is saved will result in lesser cost for the consumer and less pollution for the environment.

[1] L. Prez-Lombard, J. Ortiz, and C. Pout, “A review on buildings energy consumption information,” Energy and Buildings, vol. 40, no. 3, pp. 394-398, 2008.

[2] A. Faruqui, S. Sergici, and A. Sharif, “The impact of informational feedback on energy consumption, a survey of the experimental evidence,” Energy, vol. 35, no. 4, pp.1598-1608, 2010.

[3] A. Kailas, V. Cecchi, and A. Mukherjee, “A survey of communications and networking technologies for energy management in buildings and home automation,” Journal Computer Networks and Communications, doi:10.1155/2012/932181.

[4] J. Liang, S. K. K. Ng, G. Kendall, and J. W. M. Cheng,“Load signature study part 2: Disaggregation framework,simulation and applications,” IEEE Trans. on Power Delivery, vol. 25, no. 2, pp. 561–569, 2010.

[5] C. Laughman, K. Lee, R. Cox, S. Shaw, S. Leeb, L. Norford,and P. Armstrong, “Power signature analysis,” IEEE Power and Energy Magazine, vol. 1, no. 2, pp. 56–63, 2003.

[6] K. D. Lee, S. B. Leeb, L. K. Norford, P. R. Armstrong, J.Holloway, and S. R. Shaw, “Estimation of variable-speeddrive power consumption from harmonic content,” IEEE Trans. on Energy Conversion, vol. 20, no. 3, pp. 566–574,2005.

[7] G. Hart, “Non-intrusive load monitoring,” Proc. of the IEEE,vol. 80, no. 12, pp. 1870–1891, 1992.

[8] A. Cole and A. Albicki, “Nonintrusive identif i cation of electrical loads in a three-phase environment based on harmonic content,” in Proc. of IEEE Instrumentation and Measurement Technology Conf., Baltimore, 2000, pp. 24–29.

[9] M. Zeifman and K. Roth, “Nonintrusive appliance load monitoring: Review and outlook,” IEEE Trans. on Consumer Electronics, vol. 57, no. 1, pp. 76–84, 2011.

[10] Z. Ahmed, A. Gluhak, M. Imran, and S. Rajasegarar,“Non-intrusive load monitoring approaches for disaggregated energy sensing: A survey,” Sensors, vol. 12, no. 12, pp.16838-16866, 2012.

[11] H. Lam, G. Fung, and W. Lee, “A novel method to construct taxonomy of electrical appliances based on load signatures,”IEEE Trans. on Consumer Electronics, vol. 53, no. 2, pp.653–660, 2007.

[12] O. Parson, S. Ghosh, M. Weal, and A. Rogers,“Non-intrusive load monitoring using prior models of general appliance types,” in Proc. of the 26th Conf. on Artif i cial Intelligence, Toronto, 2012, pp. 356–362.

[13] H. Kim, “Unsupervised disaggregation of low frequency power measurements,” M.S. thesis, University of Illinois at Urbana-Champaign, 2012.

[14] K. Lee, “Electric load information system based on non-intrusive power monitoring,” Ph.D. dissertation,Massachusetts Institute of Technology, 2003.

[15] S. Leeb, “A conjoint pattern recognition approach to non-intrusive load monitoring,” Ph.D. dissertation,Massachusetts Institute of Technology, 1993.

[16] H. Lai, G. Fung, H. Lam, and W. Lee (2007). Disaggregate loads by particle swarm optimization method for non-intrusive load monitoring. Proc. of the Int. Conf. on Electrical Engineering. [Online]. Available:http://www.icee-con.org/papers/2007/Oral_Poster%20Papers/14/ICEE-672.pdf

[17] K. Suzuki, S. Inigaki, T. Suzuki, H. Nakamura, and K. Ito,“Nonintrusive appliance load monitoring based on integer programming,” in Proc. of SICE Annual Conf., Tokyo, 2008,pp. 2742–2747.

[18] Y.-H. Lin, M.-S. Tsai, and C.-S. Chen, “Applications of fuzzy classif i cation with fuzzy c-means clustering and optimization strategies for load identif i cation in NILM systems,” in Proc. of IEEE Int. Conf. on Fuzzy Systems,Taipei, 2011, pp. 859–866.

[19] T. Hassan. (2012). Bi-level characterization of manual setup non-intrusive demand disaggregation using enhanced differential evolution. Proc. of the 1st Inter-National Workshop on Non-Intrusive Load Monitoring. [Online].Available: http://www.ices.cmu.edu/psii/nilm/posters/hassan_Lahore_ NILM2012_poster.pdf

[20] J. Kolter and M. J. Johnson. (2011). REDD, a public dataset for energy disaggregation research. Proc. of the SustKDD Workshop on Data Mining Applications in Sustainability.[Online]. Available: http://www.cs.cmu.edu/~zkolter/pubs/kolter- kddsust11.pdf