About this Journal Submit a Manuscript Table of Contents
International Journal of Distributed Sensor Networks
Volume 2013 (2013), Article ID 372982, 8 pages
http://dx.doi.org/10.1155/2013/372982
Research Article

Power Load Distribution for Wireless Sensor and Actuator Networks in Smart Grid Buildings

Department of Computer Science and Statistics, Jeju National University, Jeju-Do 690-756, Republic of Korea

Received 22 October 2012; Accepted 24 December 2012

Academic Editor: Jiman Hong

Copyright © 2013 Junghoon Lee and Gyung-Leen Park. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

Abstract

This paper presents a design and analyzes the performance of an actuator operation scheduler for wireless sensor and actuator networks, aiming at efficiently managing power consumption and distributing peak load in smart grid buildings. To create a schedule within an acceptable response time, a genetic algorithm is designed, and the scheduler places the operations of activated tasks to appropriate time slots in the allocation table. For genetic operations, each schedule is encoded to an integer-valued vector, where each element represents either start time or binary allocation map of the associated task according to the task type. The fitness function evaluates the schedule quality by estimating the load of the peaking slot. Out-task model defines P-Penalty and N-Penalty to account for the extrapower load brought by the delayed start of task operation. The performance measurement results obtained from a prototype implementation reveal that our genetic scheduler reduces the peak load by up to 35.2% for the given parameter set compared with the Earliest scheduling scheme, intelligently compromising two conflicting requirements of even load distribution and small initiation delay.

1. Introduction

It is too much well known that wireless sensor networks, or WSNs in short, are extending their application areas even to harsh environments such as battlefields and chemical process facilities. Moreover, they are commonly integrating a variety of actuators specific to the system goal, forming wireless sensor and actuator networks (WSANs). Their typical application is real-time monitor-and-control of critical systems. In the mean time, the smart grid is apparently also a promising target of WSANs. As a future power system, the smart grid pursues continuous monitoring, pervasive communication, self-healing reliability, and timely reaction to meet the growing demand for sustainable and clean electric energy [1]. In the smart grid, WSANs provide essential infrastructure for automatic metering and remote system monitoring. There are many standard WSAN technologies available to the smart grid including Bluetooth, Wi-Fi, ultrawideband, and Zigbee [2]. Upon those protocols, many different energy-related applications can work without human intervention, further benefiting from M2M technologies [3].

In WSANs, sensor applications capture the current status of the monitoring target via sensor readings, while the actuators are triggered by the control logic. Every sensor and actuator operation accompanies power consumption. Power management has long been a major issue in WSANs and achieved great improvement especially in wireless communication [4]. For example, Zigbee devices are known to be able to last for several years without battery replacement [5]. Hence, the requirement on power efficiency is gradually moving from sensor part to actuator part. Particularly, in smart grid buildings, heaters or air conditioners are turned on or off based on continuous monitoring of environment variables such as temperature [6]. The simultaneous operation of multiple actuators may lead to sharp increase in power consumption, deteriorating the frequency quality and thus jeopardizing the safe operation of the power supply system [7]. The scheduling-based coordination of actuator operations can shift the power load to light-loaded time interval, alleviating the above-mentioned problems.

In the smart grid perspective, the energy-efficient actuator schedule belongs to the demand response. Not only actuator operations but also the execution of sensing applications themselves require power consumption. In smart buildings which are equipped with many actuator devices such as lighting facilities, elevator management, HVAC (Heating, Ventilation, and Air Conditioning) systems, and electric pumps, peak load shifting is more important [6]. Furthermore, charging facilities for electric vehicles will be installed in buildings. For actuator operation scheduling, time window is divided into fixed-length time slots, and each operation is aligned with the slot. Namely, actuators can be started, suspended, or resumed at each slot boundary. In addition, the load profile or interchangeably consumption profile contains the sequence of power demand along the time axis for each actuator operation. Currently, information on power consumption statistics is available in public for diverse appliances [8].

The scheduling problem for actuator operations is quite similar to real-time task scheduling in that tasks have their own deadlines [9]. Each actuator operation can be taken as a processing task while tasks arrive when the control logic decides the control action. At each arrival, task scheduling is performed. However, in actuator scheduling, tasks can run in parallel within the provisioned power cable capacity. Here, by task scheduling, some actuator operations can be delayed while others start at the moment they are activated by the control logic. The delayed start may lead to the change in power consumption dynamics. For example, the delayed initiation of air conditioners or heaters can possibly extend the operation length or increase the amount of power consumption. The scheduler must take into account this factor in schedule generation. However, such task scheduling is in most cases a complex time-consuming problem quite sensitive to the number of tasks, so conventional optimization schemes are impractical due to their extremely long execution time.

For better responsiveness, suboptimal search techniques are indispensable in spite of optimality loss. Genetic algorithms are one of the most widely used suboptimal search techniques in many different areas, not restricted to just engineering problems [10]. It can generate a task schedule within an acceptable time bound, and its execution time is even controllable by adjusting the number of genetic iterations. Moreover, it can combine a variety of efficient heuristics such as initial population selection. In this regard, this paper designs an actuator operation scheduling scheme for WSANs based on genetic algorithms. To this end, first, it is necessary to encode a schedule to a chromosome, which is represented by an integer-valued vector. Second, a fitness function must be defined to evaluate the quality of each schedule, accounting for different task behaviors according to the delayed start time. Finally, genetic operators, such as selection, crossover, and mutation, are tailored for the schedule generation based on the given task model.

This paper is organized as follows: after issuing the problem in Section 1, Section 2 surveys the background and related work of this paper. Section 3 explains actuator operation scheduler, focusing on encoding scheme design, the fitness function definition, and genetic operation customization. After performance measurement, results are demonstrated and discussed in Section 4; Section 5 finally summarizes and concludes the paper with a brief introduction of future work.

2. Background and Related Work

Reference [1] overviews the promising applications of WSNs for electric power systems, where timely information is essential for reliable power transmission and distribution from generation units to end users. Those applications include wireless automatic meter reading, remote system monitoring, equipment fault diagnostics, and the like. WSNs in the smart grid are expected to efficiently cope with harsh environmental conditions, reliability and latency requirement, packet errors with variable link capacity, and resource constraint. It focuses on the measurement of the WSN link quality in the electric-power-system environment in terms of background noise, channel characteristics, and 2.4 GHz band attention. Hence, extensive field tests have been conducted on IEEE 802.15.4-compliant sensor nodes. The experiment environments include a 500 kV substation, a main power control room, and an underground network transformer vault. After all, the wireless channel has been modeled using a log-normal shadowing path-loss model.

Reference [11] considers joint problems of control and communication in WSANs for building control systems, where WSAN serves as components of control loops. It focuses on controlling the environment variables, such as temperature, humidity, and illumination by means of heating, air conditioning, ventilating, and lighting. This system is built on top of 3-tier design consisting of network, control, and user interface. The authors propose both centralized and distributed control schemes, respectively. For both of them, Kalman filters compensate for packet losses and delays in wireless channels to estimate the current WSAN status, while the control action is decided by means of the control objective function. The performance comparison between centralized and distributed schemes is conducted in terms of control performance, energy efficiency, computational complexity, and packet loss rate. However, their control decision procedure does not consider explicitly the power consumption dynamics in actuators.

Reference [5] addresses that WSNs will play a key role in the deployment of the smart grid towards residential premises, hosting various demand and energy management applications. The authors evaluate the performance of in-home energy management schemes focusing on the energy cost reduction. This scheme allows its applications to be flexibly built on top of wireless sensor networks in home area. Here, the Zigbee technology is exploited for communication between the energy management unit and other power entities such as appliances, smart meters, and storage devices. The management system also incorporates SEP (Smart Energy Profile) 2.0 developed by Zigbee Alliance for the standardized message format and exchange procedure in automatic metering applications. The performance of the sensor network part depends on the packet size generated by the monitoring application. The simulation-based experiment finds out that the smaller packet size is preferred for smart grid applications from the viewpoint of packet delivery ratio, delay, and jitter. Anyway, standard WSNs can be seamlessly integrated into smart grid systems.

Reference [12] presents a strategy for deep demand response of power loads considering the availability of renewable energies. Its monitoring and controlling architecture also employs WSNs, where each sensor is attached to a low-power wireless mote running an IPv6-compatible networking layer. A set of motes forms an ad hoc network while a laptop computer provides a gateway to the global network. The authors introduce the notion of slack, which means the potential of an energy load to be advanced or deferred without affecting earlier or later operations. According to the operation model integrating the slack, the distribution of appliance operations can lead to a better match of energy consumption to generation. The thermostatically controlled load is the main target of the time-dependent power consumption management, and this scheme mainly focuses on the coordination of renewable energy. Their power consumption model is reasonable, and our scheme will adapt it for explicit appliance scheduling.

In the mean time, our research team has been conducting researches on demand response schemes for the purpose of applying the research results promptly in Jeju area, which has established one of the world’s largest smart grid testbeds [13]. Main focus is put on efficient scheduling of power device operations and charging of electric vehicles in smart homes, buildings, and charging stations. Based on the task model consisting of preemptive and nonpreemptive tasks, our strategies build an operation schedule for the given time window either by genetic algorithms or by exhaustive searches combined with a heuristic [14]. They are mainly concerning how to reduce peak load by distributing each task operation as evenly as possible. In addition, for the efficient integration of renewable energies in a smart grid unit, a dual battery management scheme decides when to charge or discharge each battery according to the availability of wind power generation and power demand approximation [15]. Until now, the power consumption is assumed to be constant irrespective of its start time.

3. Actuator Operations Scheduler Design

3.1. System and Task Models

Figure 1 depicts our system model which is inherited from the typical wireless WSAN architecture. As our design is targeted at smart grid buildings, sensors and actuators are selected for building environment control. After the sensor data analysis, a series of control actions are determined. The actuator control process notifies the power scheduler of those tasks. Each action has its own time constraint that must be completed within a specific time instant. Its power consumption is specified by load profile which is a sequence of power amount consumed on each time interval. At each task arrival, the power scheduler generates a new schedule or modifies an existing schedule considering the load profile and already admitted tasks. According to the schedule, the switch controller unit turns on or off power switches which connect respective electric devices to the power supply, be it the main power line or local renewable energies.

372982.fig.001
Figure 1: System model.

The operation schedule is denoted by a time table, where is the number of tasks, and is the number of time slots. For task specification, task can be modeled by the tuple of . First, indicates whether is preemptive or nonpreemptive. In addition, is the activation time of , is the deadline, and denotes the operation length, which corresponds to the length of the consumption profile entry. A nonpreemptive task can start from its activation to the latest start time, which can be calculated by subtracting from . Actually, is equivalent to the slack of . When a start time is selected, the profile entry is just copied to the allocation table one by one, as the task must not be preempted once it has started. In contrast, the preemptive task case is quite complex. To meet its time constraint, out of ( ) slots are picked for device operation. The operation can be intermittent.

The load power profile is practical for characterizing the power consumption behavior of each task. As an example, power consumption for ventilation process depends on the selected program. Along the time axis, the activated set of devices will be different stage by stage, and accordingly the amount of power consumption is different for each stage. The scheduler can assume that the power requirement of each operation step is known in priori. Here, the power consumption pattern for every electric device is aligned to the fixed-size time slot. Actually, each actuator has its own time scale in its power consumption. However, we can take the average value during each time slot considering voltage regulating equipments commonly available in most buildings. The length of a time slot can be tuned according to the system requirement on the schedule granularity and the computing time. This length usually coincides with the period of price signal change in power trade systems.

In addition to the standard consumption behavior specified by the load profile, we define two parameters to adapt the consumption dynamics according to the delayed start of actuator operations. First, P-Penalty represents the increase in the power consumption for each slot of delay. If P-Penalty is 5% and 2 slots are delayed, the power consumption increases by 10% in every slot. Second, N-Penalty denotes the amount of delay which leads to one slot extension in operation length. If N-Penalty is 3 and the actuator operation is delayed by 6 slots, the operation length increases by 2 slots. The power consumption in the extended slots is equal to the per-slot average. Additionally, if it is delayed by 4 slots and the average per-slot power consumption is 3, the operation length increases by 2 slots. Here, the power consumption in the first-extended slot is 3, and that in the second slot is 1, which corresponds to a third of the average per-slot consumption. This adaptation model accounts for the common behavioral tendency that the power consumption largely increases in proportion to the amount of delay [12].

3.2. Encoding and Fitness Function

In the development of genetic scheduler, it is necessary to represent an allocation table by a chromosome or an integer-valued vector. Our design takes integers for a schedule, where each element is associated with a task. A vector element has different meanings according to the task type. For a nonpreemptive task, the integer element denotes the slot from which its task operation starts. A single number is enough to represent the task operation in the allocation table, as a nonpreemptive task goes to the end without being suspended, once it has started. It can start from its activation time to the latest start time, which can be calculated by subtracting from . In contrast, for preemptive tasks, the integer element means the binary map by which task operations are assigned to the time slots. Hence, the map tells how to select , the length of the task operation, out of slots, the length of the map. The number of choice options is bounded by for nonpreemptive tasks while by for preemptive tasks [14].

Now, a fitness function evaluates a schedule according to the given system goal. For peak load estimation, the allocation table is converted from a chromosome. Then, we can calculate the power demand for each slot, and the maximum of them will be the peak load. The smaller the peak load, the better the schedule. In addition to this basic encoding scheme, the effect of delayed start must be further taken into account in the fitness function. It can be better described by an example shown in Figure 2. As shown in Figure 2(a), there are 4 tasks, so is 4. and are nonpreemptive, while and are preemptive. In addition, we assume that the time window consists of 10 slots, making equivalent to 10. is activated at the time slot 1 and must be completed by time slot 6. Its operation length is 3, so its profile entry has three numbers, each of which represents the power consumption on its time slots when the task is run.

fig2
Figure 2: Encoding and allocation table.

Next, Figure 2(b) shows how a chromosome is converted to an allocation table. For an encoded vector of (2, 3, 58, 13), the element associated with is 2. It means that begins from time slot 2. Hence, the profile entry (3, 2, 3) is sequentially copied to row of the allocation table from time slot 2. The allocation for can be explained in the same way. For , a preemptive task, its start time is 2 and deadline is 8. As its operation length is 4, 4 out of 7 slots need to be selected. In the vector, is associated with 58, and its binary equivalent is 0111010. The profile entry is mapped to the allocation table for each appearance of 1, from time slot 2. Hence, (0, 1, 2, 3, 0, 4, 0) is the allocation result for , where the first 0 is the power demand in slot 2. Then, Figure 2(b) also shows the per-slot power demand just below the allocation table. The fitness function finds the maximum of them, so 9.0 is the fitness value for this allocation.

Figure 2(c) illustrates how to modify the allocation table by P-Penalty and N-Penalty. First, has N-Penalty of 4, and its start is delayed by 1 slot, while average per-slot power consumption is 2.67. Hence, , namely, 1 extraslot is added, and the its power consumption is 2.67 4 = 0.67, as shown in time slot 5. Second, starts 3 slots later than its activation time, and its N-Penalty is 2. , namely, 2 slots are added for this delay. They are slots 5 and 6. As average power consumption is 3.0, slot 5 has power consumption of 3 while slot 6 has 1.5. Next, has P-Penalty of 0.2, and it is delayed by 1 slot. Hence, power consumption in every slot is increased by 20%. For , each slot increases by 10%. If a preemptive task has N-Penalty and slots are additionally needed, slots are selected out of slots from its start time (not its activation time) to deadline. The peak load takes place at the 5th slot, and peak value has changed to 10.6. It must be mentioned that slot length extension can lead to the violation of deadline constraints for some tasks. However, those schedules will be eliminated in genetic iterations.

3.3. Genetic Iterations

With the design of an encoding scheme and the definition of the fitness function, genetic operations are executed, continuously improving the quality of population. The overall procedure is described in Figure 3. Each evolutionary step generates a population of candidate solutions and evaluates them according to a given fitness function. Even though the genetic algorithm can possibly fail to find an optimal solution, its efficiency makes itself very practical, as the schedule must be created within the system’s tolerance. For initial population, a predefined number of chromosomes are generated randomly. In a chromosome, each element is selected from th e valid range of the associated task. That is all schedules in the population can meet the time constraint of each task. This restriction narrows the search scope and cuts off the chance of creating a better schedule by mating invalid schedules. However, without this restriction, the scheduler spends too much time in processing invalid schedules, which violate the time constraint of some tasks.

372982.fig.003
Figure 3: Scheduler operation.

The iteration mainly consists of selection and reproduction. Selection is a method that picks parents according to the fitness function. The Roulette wheel selection gives more chances to chromosomes having better fitness values for mating. Actually, we have also tried other strategies such as the tournament selection; however, the Roulette wheel scheme performs better than others in most cases. Reproduction or crossover is the process of taking two parents and producing offspring with the hope that the offspring will be a better solution. This operation randomly selects a pair of two crossover points and swaps the substrings from each parent. Reproduction may generate the same chromosome with currently existing ones in the population. It is meaningless to have multiple instances of a single schedule, so they will be replaced by new random ones. Additionally, mutation exchanges two elements in a single chromosome. In our scheme, the meaning of the value is different for preemptive and nonpreemptive tasks. Hence, the mutation across the different task domains must be prohibited.

4. Performance Measurement

This section implements the proposed allocation method using Visual C++ 6.0, making it run on the platform equipped with Intel Core2 Duo CPU, 3.0 GB memory, and Windows Vista operating system. The time slot length is implicitly selected to be 5  , considering the period of price signal change. For the scheduling window of 2 hours, is set to 24. For a task, the start time is selected randomly between 0 and , while the operation length and the slack exponentially distribute with the averages of 5.0 and 3.0 slots, respectively. A task will be discarded and replaced if the finish time, namely, the sum of start time and operation length, exceeds . In addition, the power level for each time slot ranges from 1.0 to 10.0. The power scale is not explicitly specified, as it depends on the grid type and included devices. As for genetic operations, each population includes 96 chromosomes, while the number of iterations is set to 1,000. This configuration can create a schedule within 1 second. Actually, even with more iterations, the fitness value hardly gets improved.

For performance comparison, the Earliest scheduling scheme initiates tasks as soon as they get ready and makes it run without preemption. It adopts no control strategy but provides a measure for a comparative assessment for the efficiency of other charging strategies. The experiment mainly focuses on peak load and total load for the schedule found by the proposed scheme. Extrapower load is not added to the Earliest scheme, as the activation time is the same as the start time for every task. Main performance metrics include the number of tasks, P-Penalty, N-Penalty, and slack. In the subsequent experiments, the number of preemptive tasks is set to 3, as the nonpreemptive tasks are more common. For each parameter setting, 30 sets are generated, and their results are averaged. In the performance comparison graphs, the curve marked by Genetic plots the performance of the proposed genetic scheduler.

The first experiment measures both peak load and total load according to the number of tasks, while the experiment results are plotted in Figure 4. Here, the number of tasks ranges from 4 to 20. For the experiment, every preemptive task is assumed to have P-Penalty of 0.05, while every nonpreemptive task is assumed to have N-Penalty of 3. As the average power consumption is almost the same for each task, for a sufficiently large number of tasks, both peak load and total load linearly increase along with the number of tasks. As shown in Figure 4(a), the proposed scheme reduces the peak load by 35.2% when the number of tasks is 8. The performance gap is generally uniform for the range of a given number of tasks. Total load indicates how much overhead is added due to peak load distribution. As shown in Figure 4(b), just a small difference is found between two cases of Earliest and proposed schemes. The difference tends to get larger when there are more tasks, but it remains less than 5%. This experiment discovers that our scheme achieves significant peak load reduction just with small increase in the total load.

fig4
Figure 4: Effect of the number of tasks.

The next experiment measures peak load and total load according to the P-Penalty ranging from 0 to 0.1, and the results are shown in Figure 5. In this experiment, the number of tasks is set to 10, and N-Penalty is to 3. Peak load and total load do not change in the Earliest scheme, as the number of tasks is fixed to 10. However, Figure 5(a) shows that peak load is rarely affected by the change of P-Penalty also in the proposed scheme. Peak load lies between 22.1 and 22.3. This result indicates that the proposed scheme can stably find an efficient schedule, even if peak load is expected to linearly increase along with P-Penalty. According to the graph, our genetic scheduler reduces peak load by around 33.1%, compared with the Earliest scheme. In addition, the total load increases when P-Penalty gets larger almost linearly as shown in Figure 5(b). The difference in the total load is about 10.1% between the two schemes when P-Penalty reaches 0.1.

fig5
Figure 5: Effect of P-Penalty.

In addition, Figure 6 traces the effect of N-Penalty to peak and total load, respectively. In the experiment, N-Penalty is changed from 1 to 10, while the number of tasks is set to 15 and P-Penalty to 0.05. A smaller value of N-Penalty makes the total power consumption increase more. If N-Penalty is 1, each slot of delay leads to one slot of extension. As in the previous experiment, both peak load and total load remain constant in the Earliest scheme, as the number of tasks is fixed. Our genetic scheduler places those tasks having severe penalty as earlier as possible to minimize the influence of delayed starts. As a result, the genetic scheduler reduces peak load by 26.2%, even when N-Penalty is 1. In addition, just 9.2% of load is added to the total power consumption, compared with the Earliest scheme.

fig6
Figure 6: Effect of N-Penalty.

Finally, Figure 7 measures the effect of task slack to peak load and total load, respectively. Here, the slack ranges from 1 to 8 slots, while the number of tasks is fixed to 15, P-Penalty to 0.05, and N-Penalty to 3. The larger the slack is, the more options scheduler has in placing task operation in the allocation table. Actually, the large number of options means the increase in the search space size. The genetic scheduler has an additional tradeoff between the execution speed and accuracy. However, as the penalty also increases if the task operation is delayed for better distribution, it is difficult to design a straightforward policy in task scheduling. For the Earliest scheme, peak load is also affected by the slack size, even if not so much, as a smaller slack value is highly likely to make several tasks activated at the same time slot.

fig7
Figure 7: Effect of task slack.

Figure 7(a) shows that peak load of the proposed scheme is smaller than that of the Earliest scheme by up to 35.7%, when the slack is 4 slots. The improvement is smallest when task operations are tight; namely, tasks have small slacks. Even though this experiment cannot find a regular improvement pattern due to two conflicting factors of even distribution and increased penalty, our scheme outperforms by at least 26.1% for the whole slack range. Additionally, according to Figure 7(b), total load gets larger according to the increase of the slack. The gap increases from 3.1% just to 6.8%. After all, this experiment discovers that our genetic scheduler takes advantage of more scheduling options stemmed from large slack, just with a little increase in the total load.

5. Conclusions

The smart grid is a future power system which essentially employs WSANs for its applications such as wireless meter reading and remote system monitoring. The power management for actuators gets more important, as many standard protocols such as Zigbee have already achieved significant improvement in sensor network part. This paper has designed an actuator operation scheduler capable of reducing peak load in power consumption of actuator tasks, taking advantage of genetic algorithms. Based on the load profile specification and the task model consisting of nonpreemptive and preemptive tasks, each schedule is represented by a chromosome. Here, a schedule is also equivalent to an allocation time table. The fitness function calculates per-slot power consumption to find the peaking slot. In schedule evaluation, P-Penalty and N-Penalty account for the extended operation length or increased power consumption brought by the delayed start of some actuator operations, respectively.

Extensive experiments have been conducted to measure the performance of the proposed scheme mainly in terms of peak load and total load according to the number of tasks, P-Penalty, N-Penalty, and task slack. The experiment reveals that our genetic scheduler reduces the peak load by up to 35.2% for the given parameter set compared with the Earliest scheduling scheme. Moreover, genetic operations can find an efficient schedule, smartly compromising two conflicting objectives of even power load distribution and small actuator initiation delay. Our genetic scheduler can possibly integrate an intelligent heuristic in initial population selection and genetic loop customization. In addition, a new system goal can be defined such as cost reduction and renewable energy integration.

As future work, we are planning to integrate renewable energies and charging facilities for electric vehicles into our WSAN-based power management system for smart buildings. Future buildings are highly likely to install distributed power generation equipments for solar and wind energies, while electric vehicles put significant load on the power system if they are plugged in to the power grid simultaneously. Computational intelligence can overcome the complexity in coordinating many different power entities having their own roles in power consumption scenarios.

Acknowledgment

This paper was financially supported by the Ministry of Knowledge Economy (MKE), Korea Institute for Advancement of Technology (KIAT) through the Inter-ER Cooperation Projects.

References

  1. V. C. Gungor, B. Lu, and G. P. Hancke, “Opportunities and challenges of wireless sensor networks in smart grid,” IEEE Transactions on Industrial Electronics, vol. 57, no. 10, pp. 3557–3564, 2010. View at Publisher · View at Google Scholar · View at Scopus
  2. D. Gislason, ZIGBEE Wireless Networking, Newnes, Melbourne, Australia, 2008.
  3. Z. M. Fadlullah, M. M. Fouda, N. Kato, A. Takeuchi, N. Iwasaki, and Y. Nozaki, “Toward intelligent machine-to-machine communications in smart grid,” IEEE Communications Magazine, vol. 49, no. 4, pp. 60–65, 2011. View at Publisher · View at Google Scholar · View at Scopus
  4. S. Bu, R. Yu, Y. Cai, and X. Liu, “When the smart grid meets energy-efficient communications: green wireless cellular networks powered by the smart grid,” IEEE Transactions on Wireless Communications, vol. 11, no. 8, pp. 3014–3024, 2012.
  5. M. Erol-Kantarci and H. T. Mouftah, “Wireless sensor networks for cost-efficient residential energy management in the smart grid,” IEEE Transactions on Smart Grid, vol. 2, no. 2, pp. 314–325, 2011. View at Publisher · View at Google Scholar · View at Scopus
  6. H. Doukas, K. D. Patlitzianas, K. Iatropoulos, and J. Psarras, “Intelligent building energy management system using rule sets,” Building and Environment, vol. 42, no. 10, pp. 3562–3569, 2007. View at Publisher · View at Google Scholar · View at Scopus
  7. R. H. Katz, D. E. Culler, S. Sanders et al., “An information-centric energy infrastructure: the Berkeley view,” Sustainable Computing: Informatics and Systems, vol. 1, no. 1, pp. 7–22, 2011. View at Publisher · View at Google Scholar · View at Scopus
  8. S. Shao, T. Zhang, M. Pipattanasomporn, and S. Rahman, “Impact of TOU rates on distribution load shapes in a smart grid with PHEV penetration,” in Proceedings of the IEEE PES Transmission and Distribution Conference and Exposition, April 2010. View at Publisher · View at Google Scholar · View at Scopus
  9. T. Facchinetti, E. Bibi, and M. Bertogna, “Reducing the peak power through real-time scheduling techniques in cyber-physical energy systems,” in Proceedings of the 1st International Workshop on Energy Aware Design and Analysis of Cyber Physical Systems, 2010.
  10. S. Sivanandam and S. Deepa, Introduction to Genetic Algorithms, Springer, Berlin, Germany, 2008.
  11. X. Cao, J. Chen, Y. Xiao, and Y. Sun, “Building-environment control with wireless sensor and actuator networks: centralized versus distributed,” IEEE Transactions on Industrial Electronics, vol. 57, no. 11, pp. 3596–3605, 2010. View at Publisher · View at Google Scholar · View at Scopus
  12. J. Taneja, D. Culler, and P. Dutta, “Towards cooperative grids: sensor/actuator networks for renewables integration,” in Proceedings of the 1st IEEE International Conference on Smart Grid Communications, pp. 531–536, October 2010. View at Publisher · View at Google Scholar · View at Scopus
  13. J. Lee, H. Kim, G. Park, and H. Jeon, “Genetic algorithm-based charging task scheduler for electric vehicles in smart transportation,” in Intelligent Information and Database Systems, vol. 7196 of Lecture Notes in Artificial Intelligence, pp. 208–217, Springer, 2012.
  14. J. Lee, H. Kim, G. Park, and M. Kang, “Energy consumption scheduler for demand response systems in the smart grid,” Journal of Information Science and Engineering, vol. 28, pp. 955–969, 2012.
  15. J. Lee and G. Park, “Dual battery management for renewable energy integration in EV charging stations,” NeuroComputing. In press.