Gina Martinez，Shufang Li，and Chi Zhou

（1.Department of Electrical and Computer Engineering，Illinois Institute of Technology，IL 60616，USA；2.Telecommunication Engineering Institute，Beijing University of Posts and Telecommunication，Beijing 100876，China）

An Optimal Lifetime Utility Routing for 5G and Energy-Harvesting Wireless Networks

Gina Martinez1，Shufang Li2，and Chi Zhou1

（1.Department of Electrical and Computer Engineering，Illinois Institute of Technology，IL 60616，USA；2.Telecommunication Engineering Institute，Beijing University of Posts and Telecommunication，Beijing 100876，China）

Harvesting energy from environmental sources such as solar and wind can mitigate or solve the limited-energy problem in wireless sensor networks.In this paper，we propose an energy-harvest-aware route-selection method that incorporates harvest availability properties and energy storage capacity limits into the routing decisions.The harvest-aware routing problem is formulated as a linear program with a utility-based objective function that balances the two conflicting routing objectives of maximum total and maximum minimum residual network energy.The simulation results show that doing so achieves a longer network lifetime，defined as the time-to-first-node-death in the network.Additionally，most existing energy-harvesting routing algorithms route each traffic flow independently from each other.The LP formulation allows for a joint optimization of multiple traffic flows.Better residual energy statistics are also achieved by such joint consideration compared to independent optimization of each commodity.

routing；5G；energy-harvesting；wireless sensor networks

1 Introduction

Wireless sensor networks（WSN）continue to be an affordable，convenient solution to a broad range of wireless communication challenges today.However，the limited energy capacity of wireless network components continues to be the main obstacle to WSN application.With research on 5G wireless networks already underway，there is a pressing need to ensure that components have sufficient energy capacity to support next -generation wireless technologies.

5G wireless networks will be characterized by high capacity and high data rate.The trend in wireless services is rapid increase in multimedia traffic and ever-increasing demand for bandwidth.Therefore，energy efficiency and power-saving system services are important in 5G research.Recently，energy harvesting has been more widely studied to satisfy the high energy demands of emerging wireless technologies.In an energy-harvesting wireless sensor network（EH-WSN），nodes are able to harvest energy，e.g.，solar，thermal and mechanical，from the surrounding environment.

Two types of technologies make EH-WSN feasible:energy-harvesting and energy-management［1］.An energy-harvesting technology converts environmental energy into electrical energy；e.g.，photovoltaic（PV）panels convert solar radiation；piezoelectric transducers convert mechanical stress；and wind turbines convert kinetic wind.Energy management involves intelligently constraining energy consumption while still meeting certain performance and energy objectives，such as maximizing network lifetime or packet delivery rate.Our work focuses on managing energy through harvest-aware routing.

Network lifetime can be defined in several ways.Here，we define it as time-to-first-node-death（TTFND）.Maximizing the lifetime of a battery-powered network involves a tradeoff between two objectives:1）minimizing the total amount of energy consumed in routing the data packets and 2）maximizing the minimum residual network battery level.In an EH-WSN，there is additional energy“consumption”in the form of energy wastage due to overcharge of finite-capacity energy buffers.Minimizing total energy consumption，including wastage，in routing is equivalent to maximizing the total residual energy of the network［2］.

Our preliminary simulation results confirm the previously mentioned tradeoff.Fig.1a shows the minimum battery level and Fig.1b shows the average battery level in a network of 25 nodes.In the maximum total remaining（MAX-TOTAL）objective，nodes along shortest-path routes are depleted faster than in maximum minimum remaining（MAX-MIN）because these routes consume the least total energy.When these nodes are depleted，the network may become partitioned，which shortensthe network lifetime.On the other hand，the average battery level of the network for MAX-MIN is less than that for MAXTOTAL because MAX-MIN may purposely select longer routes in order to avoid low-energy nodes，decreasing the overall available network energy.

▲Figure 1.（a）Minimum and（b）average battery levels for MAX-TOTAL and MAX-MIN routing.

In our previous work［2］，we discussed the objective of maximizing the total remaining network energy.We proposed a simple yet effective route-selection method for maximizing a lifetime utility function，which seeks to balance the routing objectives of MAX-TOTAL and MAX-MIN energy level in the network.We show that this achieves higher lifetime than either objective alone.This method is applicable online because it takes into account the uncertainty of future traffic rather than requiring offline，deterministic traffic-arrival information.

We investigated the effects of network factors，such as topology，energy consumption，and prediction error，on the energy savings.In doing so，we gained insight into how some network design decisions and characteristics affect target performance and energy requirements.

2 Related Work

There have been several novel works published on routing for EH-WSNs.In［3］，an energy-management and routing method that maximizes energy efficiency by modeling the energy buffer as a G/G/1（/N）queue is presented.Routes are chosen according to the probability of the buffer depletion of the nodes.The authors of［4］and［5］formulate link-cost metrics based on a node's energy availability，which generally encompasses initial battery and energy harvest.Routes are then found using Bellman-Ford，Directed Diffusion（DD），or other applicable shortest-path or least-cost algorithms with respect to these metrics.Most of the works in［6］discuss this method but not the exploitation of joint optimal routing of multiple commodities in the network.

The tradeoff between minimum energy route consumption and maximum residual energy routing mainly arises from unknown future traffic arrivals.That is，the online traffic arrival is unknown.For this reason，several heuristic online maximum lifetime algorithms for battery-powered WSNs have been investigated.In［7］，conditional MAX-MIN battery capacity routing（CMMBCR）is introduced.With CMMBCR，routes that would result in a minimum battery below a threshold are discarded. Among the remaining routes，the minimum consumption route is chosen.Alternatively，in［8］，the authors present a MAXMIN zPmin algorithm.In this algorithm，of the routes with at most zPmin energy consumption，the route resulting in maximum minimum energy is chosen.As previously shown，this tradeoff also exists for EH-WSNs［2］，［9］.Here，we introduce a simple maximum lifetime utility function that balances the two objectives.

We formulate multi-commodity routing as an LP problem that additionally takes into account finite battery capacities. Formulating the maximum lifetime routing problem as linear programming（LP）optimization has been well-explored for battery-based wireless sensor networks where the energy source is fixed and non-rechargeable.The authors of［10］provided a basis for flow conservation and energy constraints used in many LP-based routing formulations，including the one in this work. In［11］and［12］，the LP formulation takes node operation modes，such as transmit，receive and idle，into account.In［13］，routing is made more robust by taking into account uncertainties in the defined constants，such as the energy consumption costs and residual battery levels.In［14］，the LP routing problem formulation specifically addresses multi-commodity traffic and analyzes the interaction of multiple routes.However，these works were not designed for energy harvesting sensor networks，where the harvest availability and finite capacity limits of the energy buffers need to be considered.

3 System Model

We consider a WSN network with a flat，multihop topology. The network can be described by a directed graph G（V，E），where V is the set of vertices representing the nodes，and E is the set of edges representing the links.There is an edgee（i，j）∈E between nodes viand vjif vjis within the transmission range of vi.Each node vihas an associated energy cost offor transmitting a packet to node vjand an associated energy cost offor receiving a packet from vk.

Our problem formulation is amenable to multiple sink nodes；however，in this case，we assume that the sink nodes freely and quickly exchange information，and one is designated to perform the routing optimization.Without loss of generality，we designate a single node vdas a sink in the network.This node is assumed to be tethered with unlimited energy supply. Routing is performed as a centralized algorithm implemented at the sink.All other nodes vi∈V，vi≠vdcan be either a source or router node.

Operation time is divided into time slots Thof equal length. A commodity（or connection stream）cmis associated with a source and sequence number（vs，sn），and C=｛c1，...，cM｝is the set of M commodities active within a time slot.Additionally，energy harvest and packet origination rates of sources are assumed to be constant within a time slot.

4 Online Maximum Lifetime Utility Routing

where Biis the maximum battery capacity of vi.The energy units eiandare，respectively，the current battery level and expected energy harvest of vifor the future time horizon Th.

This formulation，however，creates a“time coupling”［15］between time slots because eiof the next time slot depends heavily onin the current time slot.This results in a non-linear，sequential program that is very costly to solve.In this work，we instead formulate an online utility objective function that de-couples the dependency using the MAX-TOTAL and MAX-MIN tradeoff previously mentioned.

Each packet being routed in the network is identified to belong to a certain commodity cm∈C，where C is the set of commodities active in the current time slot.is the set of downstream neighbors of node vitowards the destination of commodity cm.is the rate at which node virelays packets belonging to cmto its downstream neighbor node vj.is the energy cost of transmitting a packet from vito vj，andis the corresponding cost at vifor receiving a packet from vj.For clarity，the notations used in this section's discussion are summarized in Table 1.

We formulate the online maximum lifetime utility（OMLU）routing objective function U as a weighted function between

whereαandβare weighting parameters，and NVis the number of nodes in V excluding the sink.The vectorqˉis the set of optimal packet rates of each link e（i，j）with respect to the traffic of each commodity cm.

The set of constraints in（4）guarantees the conservation of flows［10］，which specifies that the total packet outflow of each node must equal the total inflow plus the self-generated traffic of the nodeitself if it is the source of cm.The constraint does not apply if the node is the destination.

The constraints in（5）satisfy the energy requirement that the total consumption of a node（due to reception and transmission）must not exceed the node's available energy，i.e.，the minimum between the maximum battery capacity，and the effective energy，i.e.，the sum of the residual battery level and expected harvest.

As previously mentioned，MAX-TOTAL is equivalent to the minimization of the sum of total route energy consumption and total wasted energy due to finite maximum battery capacities. Therefore，compared to simply minimizing the total route con-sumption，MAX-TOTAL has an additional incentive to avoid low-energy nodes if there is energy wastage that can be minimized.However，especially in the case where there is little energy wastage to leverage，MAX-TOTAL still results in quick depletion of nodes along shortest-path routes［9］because these are the routes that result in minimum total energy consumption.

Energy-aware routing can also prioritize the avoidance of low -energy nodes，such as in MAX-MIN.However，in general，MAX-MIN consumes more energy and results in lower overall network residual energy，as will be seen in the simulation results.Therefore，the utility function in（2）attempts to balance these two conflicting objectives.Additionally，the LP optimization can be applied to both single commodity optimization and multi-commodity joint optimization.In section 5.5，we provide simulation results that show the advantage of the latter.

5 Simulation Results

5.1 Simulation Settings

In the following simulations，25 nodes are placed within a fixed 500×500 m area（Fig.2）.The transmission range of each node is 150 m.Each node has an energy cost of=10 μWh/bit and=0?viand vj.All the nodes have rechargeable batteries that are initially full a capacity Bmaxof 500 Wh.

In this work，we consider an EH-WSN that harvests solar energy to supplement the initial batteries of the nodes.We use the solar radiation data available from the National Solar Radiation Data Base（NSRDB）［16］for our closest geographical location，which is Chicago O'Hare International Airport.Fig.2c shows the hourly solar radiation data for July 1，2010.The PV panels are assumed to have a combined efficiency and sizing factor of 0.2，which means the nodes can only harvest 20%of the energy shown in Fig.2c.

The simulation begins at 10 a.m.and has a duration Tsof 12 hours to coincide with the maximum availability of sunlight. This duration Tsis divided into timeslots with a duration Th. The simulation is terminated when a battery is depleted，when the LP no longer has a feasible solution，or when the end of Tsis reached.Although a feasible solution is found for the timeslot，a battery may become depleted within the timeslot because of some variation in the residual battery level readings obtained by the RREQs or energy consumption of the control packets that are not taken into account in the LP formulation.

Traffic is characterized by a connection stream or commodity between a randomly chosen source and the fixed destination node.Within a commodity，vioriginates packets in a constant bit rate（CBR）manner at a specified packet rate Qi.The source continues to produce packets within the timeslot，after which a new source is randomly chosen.All packets are 512 bytes. Last，the weightsαandβare both set to 1 so that in the following simulations，OMLU equally balances both objectives.

5.2 On-Demand Routing

We run the simulations using NS2 by extending the on-demand routing protocol Dynamic Source Routing（DSR）in order to implement OMLU.DSR's source routing mechanism allows us to collect per-node energy availability information（ei+ehi）without significantly modifying the Route Request（RREQ）header structure，which already collects per-hop node addresses.Moreover，the set of downstream neighborsof vifor cm， can be easily inferred from the source routes received at the sink.The route replies（RREP）include a packet rate assigned to that route.Unlike traditional DSR，the resulting route replies are derived from the optimal solutionqˉand may not correspond to specific source routes received from an RREQ.Currently，it is assumed that all RREPs are received by the source.

5.3 Performance Metrics

To evaluate performance，we compare TTFND of MAX-TOTAL，MAX-MIN，and OMLU，which are denoted ttfndMT，ttfnd-MM，and ttfndOMLU，respectively.Additionally，the average and minimum of network battery levels at ttfnd0are compared，where ttfnd0=min｛ttfndMT，ttfndMM，ttfndOMLU｝.Performance metrics may also be measured at ttfnd1，which is the second-lowestvalue in the set｛ttfndMT，ttfndMM，ttfndOMLU｝.

▲Figure 2.Simulation settings:（a）grid topology，（b）random topology instance，and（c）solar irradiation for July 1，2010.

5.4 Simulation Results for Single-Commodity Traffic

In this section，we investigate the advantage of OMLU routing objective over MAX-TOTAL or MAX-MIN by considering single commodity traffic.That is，at each time slot，only a single source is chosen to originate packets for the destination. The timeslot duration This 200 s，and the destination node v18.

5.4.1 Topology and Traffic

Network topology and traffic variation can affect the performance of routing protocols.Truly optimal performance can be achieved through offline optimization only if future traffic is known；however，this is generally not realistic.Therefore，we run simulations on different topologies and with different traffic realizations to determine whether OMLU's performance improvement can be seen across traffic and topology variations. In the following single-commodity simulations，the source at each timeslot originates packets at a rate of 3 packets per second，that is，Qi=3 for source node vi

We first run a simulation on the grid topology shown in Fig. 2a.In this network，only non-diagonal neighbors are able to communicate directly with each other.Fig.3a shows the resulting performance metrics discussed in section 5.3.Fig.3b compares the MAX-TOTAL and OMLU statistics measured at ttfnd0.The right graph compares MAX-MIN and OMLU statistics measured at ttfnd1.The average and minimum battery levels are normalized to Bmax，and ttfnd is normalized to Ts.The presented results are averaged over five simulations of random traffic on the grid network.

OMLU results in 46.48%longer TTFND than MAX-TOTAL. Moreover，at ttfnd0，which is when the first node is depleted in MAX-TOTAL and occurs around 18，500 seconds into the simulation，the average battery level of the network with OMLU is 1.3%higher than that of MAX-TOTAL.In general，we can expect the average battery of OMLU to be comparable or perhaps slightly lower than that of MAX-TOTAL because maximizing the total（or average）battery is the objective of MAX-TOTAL whereas it is not the only objective of OMLU.Also，at ttfnd0，the minimum battery of OMLU is still more than 50%of Bmaxwhereas one or more batteries in MAX-TOTAL have been depleted.

By definition，TTFND itself is very close to the definition of MAX-MIN.Therefore，we expect MAX-MIN to perform well with respect to this metric.Even so，F(xiàn)ig.3a shows that OMLU results in a longer lifetime than MAX-MIN，although the increase is smaller at 2.06%.Because MAX-MIN aggressively focuses on the minimum network battery without regard for the overall energy availability in the network，more total energy is consumed due to longer routes.However，in the long term，this leads to a shorter lifetime.At ttfnd1，which is the time of first node death for MAX-MIN for these particular set of simulations，we also see higher average and minimum battery levels for OMLU compared to MAX-MIN.

▲Figure 3.Averaged TTFND and battery level statistics on（a）grid and（b）random topologies.

These results highlight the effect of unknown future traffic on current decisions.Within a time slot，MAX-MIN may choose longer routes in order to avoid low-energy nodes.However，this reduces the energy of nodes that may potentially be critical routes of future traffic.Therefore，it is advantageous to be conservative in prioritizing either objective in anticipation of the randomness of future traffic.

Here，we determine performance across various network topologies by running simulations on networks with nodes randomly placed within the area according to a uniform distribution.Fig.3b shows the same performance metrics averaged over five simulations on randomly generated topologies，such as shown in Fig.2b，as well as randomly generated traffic.In this case，OMLU results in 63%and 6.24%longer lifetime than MAX-TOTAL or MAX-MIN routing，respectively.These results show a similar trend to that in the grid topology.OMLU results in comparable or higher average battery than MAX-TOTAL at ttfnd0，and also achieves much higher minimum batterylevels at both ttfnd0and ttfnd1compared to the other two schemes.

5.4.2 Energy Consumption Rate

Here，Qiis varied to determine the effect of network energy consumption rate on OMLU performance.Table 2 shows the resulting TTFND values for the three routing schemes，averaged over five simulations of random traffic and normalized to Ts.Of all packet rates，MAX-TOTAL results in the shortest lifetime compared with MAX-MIN and OMLU.

▼Table 2.Lifetime values normalized toTSfor various packet rates

For very low packet rates，the lifetime of OMLU is comparable or slightly shorter than that of MAX-MIN because nodes are able to replenish energy very easily due to the low energy consumption rate.Moreover，the energy consumed by the longer routes in MAX-MIN may otherwise be wasted if not used due to battery storage limit.Therefore，the larger total consumption does not adversely affect the overall residual energy of the network.

However，as the packet rate increases，OMLU achieves a much longer lifetime than either MAX-TOTAL or MAX-MIN. For example，at 3 packets/s，OMLU's lifetime is 46.48%longer than that of MAX-TOTAL and 2.06%longer than MAXMIN.At 4 packets/s，the performance gain increases to 140.8%and 20.3%，respectively.At 5 packets/s，the lifetime gain remains high at 128.8%and 26.6%，respectively.Finally，at 6 packets/s，OMLU still achieves longer lifetime than either scheme，but the respective gains are lower at 115.8%and 4.1%.The OMLU objective function，like that of MAX-TOTAL，includes the maximization of total or average residual energy，which is equivalent to the minimization of the sum of total energy consumption and network energy wastage.At high consumption rates，there is hardly any wastage to minimize； therefore，the advantage is reduced.Moreover，as consumption increases，theminterm of the objective function decreases faster than theavgterm，making the OMLU results more comparable to that of MAX-MIN.

5.4.3 Harvest Prediction

The formulations in（1）-（5）rely on the predicted energy harvest.In this subsection，the effect of harvest prediction error is explored.The previous simulations used a perfect prediction model in which，obtained at the start of the time slot，reflects the actual total energy harvested by the node within the time slot.To investigate prediction error，we compare OMLU performance with results derived from using a constant prediction model of 309.25 Wh/m2（Fig.2c）that has an RMSE of 331 Wh/m2.As in previous sections，these results are averaged over five simulations of random traffic.

Fig.4 shows the resulting TTFND values achieved by the corresponding routing scheme using the perfect and constant prediction models.As can be expected，the error-free prediction model resulted in longer lifetimes than the error-prone constant model，indicating that the accuracy of harvest prediction has non-negligible effect on performance.

Additionally，a comparison of the resulting TTFND values and battery statistics among MAX-TOTAL，MAX-MIN and OMLU using a constant harvest prediction model show similar trends to that seen in the error-free case.That is，OMLU still achieves longer TTFND compared to MAX-TOTAL and MAXMIN when there is prediction error.Moreover，its average and minimum battery levels are also higher compared to MAX-TOTAL at ttfnd0and MAX-MIN at ttfnd1.Last，although not shown in these results，harvest prediction errors introduce another adverse effect in that the routing LP problem may yield an infeasible solution due to available energy being underestimated，leading to premature termination.

5.5 Simulation Results for Multi-Commodity Traffic

Another advantage of formulating routing as an LP problem，such as in OMLU，is that we can exploit the joint optimization of multi-commodity traffic.Optimal multi-commodity routing fundamentally differs from separately routing each traffic flow without consideration of the interactions of other traffic flows in the network.In order to jointly optimize multiple commodities，the respective routing requests must be available.While this cannot be assumed in many practical settings，there are al-so many cases in which this scenario is applicable，especially in event monitoring WSNs in which nodes within a given area detect a certain event and simultaneously produce data to be sent to the server.In this section，we investigate the performance improvement of OMLU when multiple traffic commodities are jointly optimized as opposed to when commodities are optimized independently.

▲Figure 4.TTFND values for the error-free and constant prediction models.

To simulate multi-commodity traffic，M nodes are randomly chosen at the start of each timeslot to be respective sources for commodities C=｛c1，...，cM｝，at a specified packet generation rate of，which is constant within the time slot.In this section，M is set to 3，each with uniform packet rate of 2 packets/ s.Also，in order to emphasize the advantage of joint optimization，the timeslot duration is increased to Th=1000s.

Again，the results are averaged over five simulations of random traffic.Due to the heavy total packet rate in the network，the simulations terminate after about two or three hours.The battery level observed during the runtime of one such simulation is shown in Fig.5.These results show longer lifetime and higher average and minimum battery levels for joint OMLU as well as lower standard deviation of network battery levels compared to independent OMLU.

Independent OMLU solves the LP routing problem for each commodity，and Joint OMLU waits for the RREQs of all three commodities for each timeslot before solving the LP problem. The independent case terminates after a little more than one hour whereas the joint case has about 2.5 hours of runtime. This indicates the advantage of joint optimization of commodities.In Fig.5b，both cases terminate with the minimum battery still at about 10%of Bmaxbecause of the large Th.Because the LP formulation seeks a solution that theoretically ensures operation throughout the entire timeslot，the simulations terminate due to the infeasibility of the solution.

On average，the joint OMLU yields 35%longer TTFND.At the first node death of the independent case，the average battery level of the joint case is about 0.38%higher.The minimal increase is also mainly due to the relatively large Thcompared to the short runtime of two to three hours.Because routes are“held”in use within the entire timeslot，only several nodes are ever used，leaving most other nodes to remain full.This results in a high average level for both cases.However，the performance improvement is much more evident in the minimum battery where it remains at about 17%of Bmaxfor the joint case when the first node of independent OMLU is depleted.

6 Conclusion

Energy-harvesting technology is a viable solution for aggressive energy-saving in 5G wireless networks.Energy-harvesting enables WSNs to support power-hungry applications that may otherwise be impractical because of current battery capacity limits.In this paper，we have considered a WSN in which initial battery is augmented by solar energy harvested using PV panels.

▲Figure 5.Average，minimum，and standard battery level for multi-commodity traffic for independent and joint OMLU optimization of routes.

We proposed a linear programming-based solution to routing with a simple utility-based objective function that seeks a balance between the two objectives of maximizing the total and the minimum residual energy.These two objectives are generally conflicting because MAX-TOTAL can deplete nodes along shortest path routes relatively quickly.On the other hand，MAX-MIN can consume more total energy because longer routes may be selected in order to avoid low-energy nodes，resulting in larger total consumed energy.

The simulation results showed that MAX-TOTAL yielded the shortest lifetime compared to MAX-MIN and our online maximum lifetime utility routing scheme，OMLU.Comparable lifetimes were obtained for MAX-MIN and OMLU in lower energy consumption rates；however，very large lifetime gains over MAX-MIN were achieved by OMLU in high packet rates.Moreover，in the presence of harvest-prediction error，OMLU still resulted in performance gain over the other two routing schemes. OMLU achieves the best of both worlds in that its average battery level is comparable to that of the MAX-TOTAL and the minimum battery level is comparable or higher that that of MAX-MIN.

In this paper，we also investigated the joint optimization of multi-commodity traffic in which multiple traffic flows between different source and destination nodes are jointly optimized. We showed that a joint solution can lead to better performance compared to an independent solution for each traffic flow.

In future work，we plan to formulate network lifetime formally across multiple time slots.However，because energy levels across time slots are not independent，the problem is no longer linear but probabilistic because future traffic is unknown. Moreover，we plan to more extensively investigate practical settings that were not considered in this work，such as mobility and stochastic solar harvest availability.

［1］Z.G.Wan，Y.K.Tan，and C.Yuen，“Review on energy harvesting and energy management for sustainable wireless sensor networks，”in IEEE 13th International Conference on Communication Technology，Jinan，China，Sept.2011，pp. 362-367.doi:10.1109/ICCT.2011.6157897.

［2］G.Martinez，S.Li，and C.Zhou，“Maximum residual energy routing in wastageaware energy harvesting wireless sensor networks”，IEEE 79th Vehicular Technology Conference，Seoul，South Korea，May 2014，pp.1-5.doi:10.1109/VTCSpring.2014.7022982.

［3］L.X.Cai，Y.Liu，T.H.Luan，et al.，“Sustainability analysis and resource management for wireless mesh networks with renewable energy supplies，”IEEE Journal on Selected Areas in Communications，vol.32，no.2，pp.345-355，F(xiàn)eb. 2014.doi:10.1109/JSAC.2014.141214.

［4］L.Lin，N.Shroff，and R.Srikant，“Asymptotically optimal energy-aware routing for multihop wireless networks with renewable energy sources，”IEEE/ACM Transactions on Networking，vol.15，no.5，pp.1021-1034，Oct.2007.doi: 10.1109/TNET.2007.896173.

［5］M.K.Jakobsen，J.Madsen，and M.Hansen，“DEHAR:a distributed energy harvesting aware routing algorithm for ad-hoc multi-hop wireless sensor networks，”in IEEE International Symposium on a World of Wireless Mobile and Multimedia Networks，Montreal，Canada，June 2010，pp.1-9.doi:10.1109/WOWMOM.2010.5534899.

［6］D.Hasenfratz，A.Meier，C.Moser，et al.，“Analysis，comparison，and optimization of routing protocols for energy harvesting wireless sensor networks，”in IEEE International Conference on Sensor Networks，Ubiquitous，and Trustworthy Computing，Newport Beach，USA，June 2010，pp.19-26.doi:10.1109/ SUTC.2010.35.

［7］C.K.Toh，“Maximum battery life routing to support ubiquitous mobile computing in wireless ad hoc networks，”IEEE Communications Magazine，vol.39，no. 6，pp.138-147，Jun.2001.doi:10.1109/35.925682.

［8］Q.Li，J.Aslam，and D.Rus，“Online power-aware routing in wireless ad-hoc networks，”in 7th Annual International Conference on Mobile Computing and Networking，Rome，Italy，Jul.2001，pp.97-107.doi:10.1145/381677.381687.

［9］G.Martinez，S.Li，and C.Zhou，“Wastage-aware routing in energy-harvesting wireless sensor networks，”IEEE Sensors Journal，vol.14，no.9，pp.2967-2974，Apr.2014.doi:10.1109/JSEN.2014.2319741.

［10］J.Chang and L.Tassiulas，“Maximum lifetime routing in wireless sensor networks，”IEEE/ACM Transactions on Networking，vol.12，no.4，pp.609-619，Aug.2004.doi:10.1109/TNET.2004.833122.

［11］K.R.Krishnan，D.Shallcross，and L.Kant，“Joint optimization of transmission schedule and routing for network capacity，”in IEEE Military Communications Conference，San Diego，USA，Nov.2008，pp.1-7.doi:10.1109/MILCOM.2008.4753383.

［12］B.K.Cetin，N.R.Prasad，and R.Prasad，“A novel linear programming formulation of maximum lifetime routing problem in wireless sensor networks，”in 7th International Wireless Communications and Mobile Computing Conference，Istanbul，Turkey，Jul.2011，pp.1865-1870.doi:10.1109/IWCMC.2011.5982819.

［13］I.C.Paschalidis and R.Wu，“On robust maximum lifetime routing in wireless sensor networks，”in 47th IEEE Conference on Decision and Control，Cancun，Mexico，Dec.2008，pp.1684-1689.doi:10.1109/CDC.2008.4738804.

［14］V.Kolar and N.B.Abu-Ghazaleh，“A multi-commodity flow approach for globally-aware routing multi-hop wireless networks，”Fourth Annual IEEE International Conference on Pervasive Computing and Communications，Pisa，Italy，Mar.2006，pp.308-317.doi:10.1109/PERCOM.2006.3.

［15］S.Chen，P.Sinha，N.Shroff，and C.Joo，“Finite-horizon energy allocation and routing scheme in rechargeable sensor networks，”in IEEE INFOCOM，Shanghai，China，pp.Apr.2011，2273-2281.doi:10.1109/INFCOM.2011.5935044.

［16］U.S.Dept.of Energy.National solar radiation database［Online］.Available: http://rredc.nrel.gov/solar/old_data/nsrdb/

Manuscript received:2014-07-15

Biographiesphies

Gina Martinez（gmartine@iit.edu）received her BS degree in Computer Engineering from the University of Illinois at Urbana-Champaign in 2005.She received her MS in Computer Engineering from Illinois Institute of Technology（IIT）in 2009.She is currently a PhD student in the Department of Electrical and Computer Engineering，IIT.Her research interests include wireless sensor networks（WSN），green communications，WSN routing，energy-harvesting and resource management in WSNs.She is an IEEE student member.

Shufang Li（lisf@bupt.edu.cn）received her PhD degree in electrical engineering from Tsinghua University，China，in 1997.She is professor and PhD adviser in School of Information and Communication Engineering，Beijing University of Posts and Communications（BUPT）.She is also the director of the Joint Lab of BUPT and State Radio Monitoring Center（SRMC），China.Her primary research interests include EMI/EMC，simulation and optimization for radiation interference on highspeed digital circuit，simulation on electro-magnetic environment and countermeasure for interference control，simulation on SAR specific absorption rate）for electromagnetic radiation of cellular phones，and testing technology on EMC，and the theory and design of radio frequency circuits in wireless communication.Professor Li has been a senior member of the IEEE since 2009.

Chi Zhou（zhou@iit.edu）received double degrees in both automation and business administration from Tsinghua University，China，in 1997.She received her MS and PhD degrees in electrical and computer engineering from Northwestern University，USA，in 2000 and 2002.Since 2006，she has been working in the Department of Electrical and Computer Engineering，Illinois Institute of Technology，and is currently an associate professor there.Her primary research interests include wireless sensor networks，scheduling for OFMA/MIMO systems，network coding for wireless mesh networks，and integration of optical and wireless networks.She is a senior member of the IEEE.