Network Coding Protocols for Smart Grid Communications

IEEE TRANSACTIONS ON SMART GRID, VOL. 5, NO. 3, MAY 2014 1523

Network Coding Protocols for Smart GridCommunications

Rui Prior, Daniel E. Lucani, Yannick Phulpin, Maricica Nistor, and João Barros

Abstract—We propose a robust network coding protocol for en-hancing the reliability and speed of data gathering in smart grids.At the heart of our protocol lies the idea of tunable sparse networkcoding, which adopts the transmission of sparsely coded packetsat the beginning of the transmission process but then switches toa denser coding structure towards the end. Our systematic mech-anism maintains the sparse structure during the recombination ofpackets at the intermediate nodes. The performance of our pro-tocol is compared by means of simulations of IEEE reference gridsagainst standard master-slave protocols used in real systems. Ourresults show that network coding achieves 100% reliability, evenfor hostile network conditions, while gathering data 10 times fasterthan standard master-slave schemes.Index Terms—Data gathering, network coding, smart grids com-

munication protocols, sparse coding.

I. INTRODUCTION

E LECTRICAL power systems have recently experienced afast evolution driven by the development of smart appli-

cations, mainly at the distribution level. Information and com-munication technologies are actually deployed to enable newfunctionalities, avoiding or delaying costly grid reinforcementsmade necessary by the massive integration of distributed gener-ation based on renewable energies and new electricity uses, suchas electric mobility and heat pumps [2]. In practice, real-time ap-plications such as distribution state estimation [3], emergencycontrol [4], and dynamic pricing of electricity [5] are at an earlystage yet, as many utilities consider only advanced meteringfunctionalities for deployment projects [6]. Many smart me-tering infrastructures rely on bidirectional data exchanges be-tween a base station (BS) and the distribution network userswith a periodicity between 10 and 60 minutes [7]. The commu-nications correspond to a data stream of the order of 1 Mbit persubstation per data collection period, e.g., 100 bits by distribu-tion network users and 10 000 users.

Manuscript received April 15, 2013; revised September 14, 2013; acceptedOctober 17, 2013. Date of publication March 06, 2014; date of current ver-sion April 17, 2014. This work is partially financed by the Green Mobile Cloudproject granted by the Danish Council for Independent Research (Grant No.10-081621), Fundação para a Ciência e Tecnologia (Portuguese Foundation forScience and Technology) under grant SFRH-BD-61953-2009 and CodeStreamproject (PTDC/EEI-TEL/3006/2012). Paper no. TSG-00292-2013.R. Prior is with Instituto de Telecommunicações, FCUP, 4169-007 Porto, Por-

tugal (e-mail: [email protected]).D. E. Lucani is with the Department of Electronic Systems, Aalborg Univer-

sity, 9220 Aalborg Denmark (e-mail: [email protected]).Y. Phulpin was with Inesc, 4200-465 Porto, Portugal (e-mail:

[email protected]).M. Nistor and J. Barros are with Instituto de Telecomunicações, Faculdade

de Engenharia, Universidade do Porto, 4200-465, Porto, Portugal (e-mail:[email protected], [email protected]).Color versions of one or more of the figures in this paper are available online

at http://ieeexplore.ieee.org.Digital Object Identifier 10.1109/TSG.2013.2288868

To handle such information exchanges, several communi-cation technologies have been envisioned [8]. They can beclassified in three categories, namely power line communi-cation (PLC), cable communication (by telephone or opticfiber), and wireless communication (GSM, GPRS, ISM, RF).As cable communications generally involve the developmentof a dedicated infrastructure with high costs, PLC and wirelesscommunications are usually considered as the most promisingalternatives [7]. Nevertheless, these technologies raise practicalissues that challenge the large-scale deployment of smart metersin distribution systems. In particular, PLC approaches may failto connect every single household (or substation) of the grid dueto the strong attenuation of the communication signal [9], andsuffer from interference as the spectrum is unregulated [10], [11].As for wireless communication, the main challenges are relatedwith transmissionmedia characteristics, including signal fading,noise, and path loss [12]. In addition, both PLC and wirelesscommunications rely on accessible transmission and are, thus,subject to security issues, including potential malicious attacks[13] and provision of privacy guarantees [14]. Hence, neitherPLC or wireless communication infrastructures achieve thequality of service requirements for smart grids. In fact, studiesof advanced metering infrastructures have highlighted that bothtechnologies suffer a lack of reliability, as the information lossoften exceeds 1% [12], [15], even after employing reliablecommunication methods to improve performance.To address this bottleneck, research has mainly focused on

physical and MAC layers, leading to the development of ad-vanced PLC standards, for example [16]. Alternatively, it hasbeen suggested that a significant gain in reliability could be real-ized by leveraging the broadcasting properties of PLC and wire-less communications [17]. Until now, the conventional proto-cols based on master-slave or flooding techniques are prevalentsmart grid communications [5].Given the disruptive nature and benefits of network coding

[18] to enhance network performance in terms of throughput,delay, robustness, and energy consumption [19], [20], we setout to understand the potential of this technique in smart grids.Network coding can be deemed well-suited for dealing with thechallenges of smart grid communication protocols because of itsproven high efficiency and reliability in wireless broadcastingand mesh networks [21] and in distributed wireless sensor net-works [22]. In practice, various network coding protocols havebeen demonstrated, e.g., [23], [24]. In the context of smart gridcommunications, network coding has been addressed in [17] forwireless and PLC communication and [25] for wireless. Refer-ence [17] presents network coding strategies to be implementedin smart grid applications, while [25] supports a QoS-provi-sioning MAC protocol.However, the main challenge here is to deal with very small

data packets generated at different sources across the network,which introduce constraints on the type and structure of coding

1949-3053 © 2014 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

1524 IEEE TRANSACTIONS ON SMART GRID, VOL. 5, NO. 3, MAY 2014

that can be supported, as well as the number of packets that canbe coded without incurring a high signalling overhead due tocoding coefficients [21]. Our contributions are:• Network coding protocols: We use network coding insmart grid communications, e.g., for advanced meteringinfrastructures, and propose two protocols for leveragingthe benefits of this approach for information gatheringand broadcasting of data messages, while still addressingthe challenge of coding across small packets comingfrom thousands of different data sources. Both proposedprotocols are based on tunable sparse network coding.

• Comparison results: We show the benefits of networkcoding for smart grids in the contexts of wireless communi-cations by comparing its performance with a master-slavescheme and some data collection approaches in deploy-ments of up to about 4000 smart meters demonstratingthat our approach can provide an order of magnitudeimprovement in collection time with respect to currentlydeployed schemes while ensuring higher reliability.

II. BACKGROUND

Asmentionedbefore, bothwireless andPLCare twoexamplesof technologies used in smart grids [26]. The main challenges inwireless communications are with respect to the channel itself,which is unpredictable, and the signal that propagates throughit may suffer fluctuations, reflections and attenuation. Here,the signal can be characterized by the path loss and shadowingmodels, where the first one is related to the dissipation of thepower radiated by the transmitter and the second to obstaclesbetween the transmitter and receiver that may attenuate thesignal power [27]. In fact, channel characterization in wirelesscommunications constitutes a wide research area, where modelsare developed for any type of environment, indoor, outdoor oreven for transitions from one environment to another [28]–[30].Regarding PLC communication, transmitting data over powerlines is becoming increasingly used, even if the channel condi-tions are very harsh in this case. These channels are characterizedby various types of noises, such as background, narrowband,impulse, and also frequency-dependent attenuation and multi-path distortion [31]. Current research works focus on proposingchannel models for indoor [32], indoor-to-outdoor [33], [34],but also for outdoor environment for different frequency ranges[35]. Our protocols are evaluated on broadcast communicationsthroughwireless radio, but they are compatible with PLC, whichalso shares a broadcast channel. In fact, some of our notions, e.g.,the definition of downstream nodes in Section III, are perhapsmore natural in a PLC setting.The idea behind network coding is that intermediate nodes

in the network can mix the packets through algebraic opera-tions, breaking the traditional store-and-forward approach. Inparticular, random linear network coding (RLNC) [36] providesa fully distributed methodology for network coding, wherebyeach node in the network selects independently and randomlya set of coefficients and uses them to form linear combinationsof the data symbols (or packets) it receives (called also inno-vative coded packets). More recently, tunable sparse networkcoding was introduced in [1], where the coding is done at dif-ferent levels of sparsity, i.e., more sparse at the beginning oftransmission (coding coefficients mostly zero) and denser to-wards the end, while keeping the transmitted coded packets in-novative with high probability. This scheme reduces the delayand decoding complexity, and our work was inspired in part bythis idea.

III. NETWORK CODING STRATEGIES FOR SMART GRIDS

The network is modelled here as a graph represented by theset of nodes. We consider a BS, also called the informationsink, in charge of data gathering, coordination, and control ofthe nodes (secondary substations and households). The BS isaware of the number of online nodes, and the network topologyis multi-hop. The terminals are responsible for gathering mea-surements and forwarding messages to subsystems connectedto them. We assume that all the terminals have the same perfor-mance capabilities in terms of processing and storage. More-over, the links between the BS and nodes are noisy and the signalcan fade, such that the data transmitted from the terminals to theBS and backward can be lost.Our goal is to periodically (e.g., every 15 min) convey one

information packet from each node to the BS. The main idea isto ascertain the potential benefits of the use of network coding,as well as providing insights on how to manage large-scale datagathering applications with small packets, rather than fully de-veloping a specific network coding protocol. We consider a rea-sonable clock synchronization (a few seconds’ drift is accept-able) of the BS and the data senders.The concept of downstream node is used, meaning a node

that is closer to the sink than the local node. We intend the in-formation to flow as waves towards the sink and compute thedownstream nodes by using a shortest path algorithm. Other ap-proaches have been considered in [37].

A. Main Challenges and Design Features

Our scenario for data communication in smart grids brings thefollowing main challenges: 1) network topology is very com-plex, since the number of nodes in the smart grid network in-cluding the secondary substations and the households is verylarge, e.g., 3936 smart meters in some of our numerical results;2) as a consequence, the number of data packets to be trans-mitted through the network to the BS is also considerable, es-pecially since the paths tend to be very large in terms of hopcount; 3) the size of the data packets is small, meaning that themain constraints are related to the type and structure of codedpackets, and also to the number of the data packets to be codedwithout inducing a high overhead due to coding coefficients; 4)given the large amount of data to be sent to the BS, the key chal-lenge is focused on gathering the data packets before a deadline,since the transmission process has some predefined delay con-straints; 5) in terms of design criteria, routing in such a largenetwork requires too much time and network resources devotedto signalling.To overcome these challenges and ensure reliability, we use

the idea from [1]. While tunable sparse codes can be applied tomany other communication problems, smart grid communica-tions is an application for which they are particularly well suited,as would be others involving rather small packets and a largenumber of nodes, like sensor networks in general. Here, the levelof sparsity is tuned, such that at the beginning coding is sparserand then, towards the end, the density of the code is increased.Uncoded packets are transmitted first. Then, each node startsgenerating coded packets using a limited number of packetsfrom the set of already decoded by that node. In general, thecoded packet has the format shown in Fig. 1, where the “header”can be any extra information needed, and are the randomcoding coefficients for packet and , respectively, and and

are the identification of each source,and . In particular, for it is unnecessary to sendthe coefficients’ values, since a non-zero coefficient is equal

PRIOR et al.: NETWORK CODING PROTOCOLS FOR SMART GRID COMMUNICATIONS 1525

Fig. 1. Structure of the coded packet for the general case.

Fig. 2. Tunable sparse codes example from the BS perspective.

to 1. Besides decoding, the IDs in the coded packets are alsoused to inform upstream neighbors of packets already decodedat the transmitting node, providing implicit feedback [38]. Third,after a predefined number of received packets at the sink, the BSsends a negative acknowledgement, called explicit feedback, in-dicating the missing packets. This allows the nodes to eliminatefrom their buffers the packets that have already being decodedby the BS. The new coded packets will be effectively denser and,thus, innovative with higher probability than if sparse codedpackets would be used instead.We illustrate the effect of increasing this effective density

at the BS by an example in Fig. 2. The effective density oc-curs when the innovation probability is increased, thus, makinga packet to look denser. We assume that four linear indepen-dent coded packets are received at the BS, and packets

can be successfully decoded. By using an explicitfeedback, the BS announces the nodes in the network about themissing packets and . The terminals spread this feedbackamong the network and the nodes eliminate from their buffersthe packets already decoded at the BS, namely .Thus, in the following transmissions, the nodes perform com-binations of packets using only those not yet decoded. For in-stance, the combinations obtained can be and , asFig. 2 illustrates. To achieve the same performance without thisexplicit feedback, the last combinations would have to be per-formed across all packets using a higher actual density, i.e., con-tributions from more packets as in the bottom case. Reducingthe set of packets being combined yields a packet that has ahigher effective density because the other coefficients used inthe combination would be cancelled anyway. The direct advan-tage of using explicit feedback in our system is that the overheadneeded to send coefficients can be maintained at the same level.An added benefit lies in the reduction of decoding complexity.

B. Network Coding Protocols

1) Overview: As previously stated, the goal of the networkcoding protocol is to collect, at the data sink, one data packetfrom each data sender per data collection period, designatedround. On their way to the sink, the packets will be combinedwith other packets from the same round but different sources.Spreading the information content of each source packet across

many different combinations is beneficial because it makes theprotocol resilient to the loss of individual packets.In order to keep the decoding complexity low, the coding

process is performed over , meaning that linear combi-nations are carried out by selecting the packets involved in thecombination and XORing them. For this task to be performedin an entirely decentralized fashion and without requiring statesynchronization among the nodes, we adopted a generation-based coding scheme [39], inspired by RLNC. Here, a gener-ation corresponds to a round. RLNC requires the coding coef-ficients (global encoding vector) to be sent in each packet as aheader, adding to the protocol overhead. In the globalencoding vector is a bitmap indicating the source packets in-cluded in the combination. Since the overhead of carrying thisbitmap would become too high in topologies with large gener-ation sizes (corresponding to a large number of nodes, in ourcase), our protocol header conveys instead a list with an iden-tification of each source packet included in the combination, asFig. 1 shows. Keeping the overhead low implies that only a lim-ited number of source packets can be mixed together in anygiven combination. In the proposed protocol, a node can onlyinclude in the combinations it generates packets that it has al-ready decoded. This aspect facilitates keeping the sparsity of thecombinations as they progress downstream towards the sink.The combinations are transmitted in broadcast mode,

meaning that their reception is not acknowledged. However,individual feedback of the combinations is not important, whatreally matters to the intermediate nodes is knowing whichpackets have been decoded by downstream nodes and can,hence, be removed from the set of packets to include in futurecombinations. Since the nodes can only combine packets theyhave themselves already decoded, and since the nodes canoverhear other nodes’ transmissions, the protocol uses an im-plicit feedback mechanism in which the inclusion of a sourcepacket in a combination transmitted by a downstream node andoverheard by an upstream node is interpreted by the upstreamnode as an acknowledgement for that packet. Note, that theinclusion by the downstream node of a given decoded packet isprobabilistic, implying that this implicit feedback may not beimmediate. This mechanism allows information to progress asa wave towards the BS.It is worth noting that even though some of the protocol

specific aspects that are necessary to make it practical (use of, sparse coding, exclusive use of decoded packets in

combinations) lower the efficiency of network coding, we stillbenefit from its advantages—recoding the packets and miti-gating the coupon collector’s problem [40], [41]. It is known,however, that sparse coding does not eliminate coupon col-lector’s problem altogether [1]. We use implicit acknowledgesby pruning out the unnecessary packets from the queues, thusincreasing the effective density of the packets. Furthermore,in a variant of the protocol called the two-phase, we introducea simple scheme for one shot explicit feedback from the sink.Whenever the number of packets not yet decoded by the sinkbecomes lower than a given threshold, the sink transmits anexplicit feedback packet containing a list of the missing sourcepackets. This message is flooded to all the nodes, uncoded.After receiving the feedback from the BS, a node will nolonger include in future combinations packets not identifiedin the list. By not including source packets that were alreadydecoded at the BS, the new combinations have an effect in theperformance similar to that of an increased number of mixed-insource packets (a denser code).


Fig. 3. Small example of network coding protocol using tunable sparse codes for 4 nodes. (a) Spread uncoded packets, e.g., node 1. (b) Push coded packets, e.g.,node 4. (c) Explicit feedback from the BS.

TABLE IAN EXAMPLE OF NETWORK CODING WITH EXPLICIT FEEDBACK USING TUNABLE SPARSE CODES FOR 4 NODES

We provide an example in Fig. 3 and Table I to illustrate thereason behind the effective density. Four data senders want tosend their packets to the BS. The transmissionrange is shown by the dashed lines, but packets can be lost evenwithin range. The data packets sent by a downstream node andreceived by an upstream node are either 1) discarded if the datacontains packets from the upstream nodes or 2) not saved if thedata refers only to the downstream packets. The initial bufferstate is given in slot 1. The first stage is illustrated in Fig. 3(a)and in slots 2 to 5 in Table I. In slot 6 the systems starts thetransmission of coded packets and continues until slot 8. In thisexample, the number of source packets that can be combinedin a coded packet is limited to 2. For instance, in slot 7, node2 broadcasts , since it can generate only combinationsfrom the packets already decoded. When node 4 sends ,the BS decodes , and, using the implicit feedback mechanism,node 1 discards from its buffer [Fig. 3(b)]. The threshold formissing coded packets at the BS is chosen to be 2, meaning thatthe BS generates an explicit feedback in slot 9 acknowledgingthe absence of and [Fig. 3(c)]. This feedback is floodedto all nodes in the network, and node 2 discards from itsencoding queue, node 3 discards , and node 4 discardsand . The process continues until all the packets are receivedat the BS.2) Protocols Details:One-Phase Protocol (NC): The protocol starts by estab-

lishing which nodes are downstream from others. As previouslystated, a shortest path algorithm is used for building a pathtree rooted on the BS. This tree is not used for routing, butonly to identify whether a given node is downstream. All datasenders generate a packet at the beginning of each round andstart by transmitting that packet uncoded at an instant selecteduniformly at random within a short time window, e.g., 1 s. Anew transmission is scheduled with a delay chosen uniformly at

random from within a similar window. At the instant of trans-mission, the node creates a coded packet by XORing a certainnumber of packets, , selected among those that have alreadybeen decoded at that node . Since identifiers of the sourcepackets mixed in a combination must be sent in the header andwe want to limit the overhead, is upper-bounded by a number(i.e., combinations are -sparse). The actual number of sourcepackets contained in a combination is , whereis an integer selected uniformly at random in the range

. All packets are sent in broadcast mode, meaning thatthis is no link-by-link reliability, but nodes can overhear thetransmissions of other nearby nodes. Since the combinationsare generated from packets that have already been decoded atthe node, when a node overhears a transmission of a down-stream node, it marks the IDs of the mixed-in source packets asflushed, and will not include them in future combinations.Decoding the original packets from the combinations is

done by using a modified version of Gauss-Jordan elimination,(though we are aware of a more efficient algorithm proposed in[1]). Thus, when a coded packet arrives, we start by removingthe already decoded packets, and then we run Gauss-Jordanelimination, but skipping the all-zero columns. The process istriggered on the arrival of each combination, and may or maynot result in the immediate decoding of one or more originalpackets. The overall worst-case complexity is for anentire full rank matrix, and for each processing steptriggered by the reception of a single combination. In practice,since the information flows in “waves,” the average complexityis lower and partial decoding is possible. The processing over-head should not be a practical limitation of the network codingprotocols for typical values of the generation time, neithershould the energy consumption due to this computation, sincethe devices draw power from the grid. Memory consumption isalso a non-issue. The packets have a small size, and the genera-


tions and limited number of packets used in each combinationshelp to keep the overhead low. Only, one bit is required for eachentry in the array used for Gauss-Jordan elimination,and only up to packets can be buffered at any time in eachnode.

Two-Phase Protocol (NC-Fb): This is an extension of theNC protocol. On crossing a given threshold on the number ofmissing packets, the BS sends an explicit feedback packet listingthe IDs of yet undecoded packets. On hearing this packet forthe first time, each node resends an identical copy, effectivelyflooding the packet to all nodes in the topology. The absence ofan ID means that the BS has already decoded the correspondingpacket. Thus, all IDs missing from the received explicit feed-back are marked as flushed, and the nodes stop including thecorresponding packets in new combinations.

IV. PERFORMANCE EVALUATION

A. Comparison Schemes

We compare the network coding protocols with two otherschemes using wireless communication. The multi-hop pathsare determined based on a reliability metric, the expected prob-ability of successful one-hop transmission between each pair ofnodes without MAC layer retransmissions. A graph is built con-sidering only the pairwise “links” with a probability higher thana given threshold, and a shortest paths algorithm is run on thatgraph. Note that we simulate only one single area, but in prac-tice the protocol can be applied to multiple BS, each one corre-sponding to a different area.Master/Slave Reference Protocol: The Master/Slave pro-

tocol (MS) relies on a mechanism where the master requestsdata from each slave in a round-robin fashion. The round timeis divided by the number of slaves from which data packetsmust be collected, yielding the time slot. The master querieseach slave in turn, waiting for a slot time for the response. If theresponse arrives within the slot time, the master proceeds im-mediately to the next slave, otherwise, it waits until the end ofthe slot and proceeds to the next slave without retrying. Shouldthe response arrive late, it is still accepted. No reliability mech-anism is implemented in this layer. However, since both thequery and response packets are unicast, the link layer protocol(802.11 MAC for wireless) provides link-by-link reliability bymeans of acknowledges and retransmission of frames (up toseven times). The MS is similar to protocols already deployed[42].Unsolicited Packet Reference Protocols: We can avoid the

overhead of queries in MS by considering that each node spon-taneously generates an unsolicited packet every round time andsends it unicast to the BS. We test two versions. In the first ver-sion, called UOrd, the round time is divided into slots and eachsender transmits its packet at the beginning of the correspondingslot, in the same order of the MS. In the second version, we de-fine a transmission window at the beginning of each round, andeach sender selects its transmission instant uniformly at randomwithin this window. In order to avoid large losses due to con-gestion caused by the funnelling effect [43], the window shouldnot be too short. We used three different window lengths from300 s to 900 s, hence, we name the schemes as UR300, UR600,and UR900.

B. Performance Settings

In order to gain a clearer idea on the gain that can be attainedin a realistic practical setting, we implemented NC, NC-fb, and

Fig. 4. IEEE 123 node test feeder.

the reference protocols in the ns-2 network simulator (Version2.34) and conducted several experiments in a wireless setting(802.11). Recall that communication through wireless radiomight not be the preferred choice for smart grids, as most smartmetering infrastructures use alternative communication means(PLC, GPRS), but it is a widespread technology and has a verylow cost. The goal of the protocols is to collect one packet of100 bytes from each sender every 15 min (900 s). RLNC wasnot tested in the simulations simply because it would not befeasible with such large number of nodes.We start with the topology based on the IEEE 123 node test

feeder [44], [45] illustrated in Fig. 4, where the BS and the sec-ondary substations are positioned within a rectangle of 3300 mper 2500m. Around each substation, we add 32 households, dis-tributed uniformly at randomwithin an annulus with a minimumradius of 10 m and a maximum of 250 m.We use the Shadowing2 propagation model with a path loss

exponent of 3.5 and a shadowing deviation of 7.0 dB, which areaverage values for outdoor communication in urban areas [46].The validity parameter is set to 0.999, meaning that packets arenever delivered to nodes where the probability of successful re-ception is lower than 0.1% (this value speeds up the simulationsand does not noticeably affect the results).The experiments simulate 32 complete rounds, equal to 8

simulated hours, and data for the first round is discarded. Weperform 4 runs of each simulation, with different PRNG seeds.The simulations use fixed routing, for several reasons. First,we want our results to be independent of the routing protocol.Second, with such a large topology, ad-hoc routing protocolswould consume too much of the network’s capacity, hinderingour ability to fairly compare the intrinsic merits of each ap-proach. Third, the physical topology of these networks is stable(no node mobility), and the addition or removal of a node isvery sporadic. Finally, network coding does not even requirerouting to be in place, only a notion of whether a node is down-stream. The routing table is computed by running an all-pairsshortest path algorithm considering only the “links” with 70%or better reliability (the highest value for which the resultinggraph is connected), as determined by 32 runs of preliminarysimulations where 400 packets are broadcasted from each node(12800 packets from each node), well-spaced in time, and wherethe nodes receiving each packet are logged.The packets for the network coding schemes must be larger,

because of the IDs of the mixed-in packets. In our case, theoverhead is 20 bytes, corresponding to up to 10 mixed-in


Fig. 5. (a) CDF of the number of hops to the sink. (b) Average percentage of collected packets and (c) average collection time.

Fig. 6. (a) CDF of the per source node packet delivery probability. (b) Average collection time for different numbers of households per secondary substation. (c)Average collection time for 4 households per secondary substation.

packets identified by their 16-bit IDs. Thus, the number ofsource packets included in a combination is equal to 10 (i.e.,

). Since the packet size affects the probability of re-ception, different routing tables had to be computed for thenetwork coding and the reference protocols. This fact cannotyield any advantage to the network coding protocols, since theloss probability is higher for longer packets. In practice, it madelittle to no difference in the cumulative distribution function(CDF) of path lengths, as can be observed in Fig. 5(a), wherethe lines are indistinguishable.

C. Network Coding vs. Reference Protocols

In the first experiment we compare the performance ofthe network coding protocols and the reference protocols,as Fig. 5(b) and (c) show. The error bars represent the 99%confidence intervals for the mean. The network coding pro-tocols achieve 100% reliability, which is not the case for anyof the other protocols, and the collection time is much shorterthan that of the reference protocols. We configured the BS tosend explicit feedback when 60 or fewer packets remain tobe decoded (60 of 16-bit IDs fit in a packet with 120 bytes ofpayload). This improvement further reduced the collection timeto less than on tenth of that of MS, while obviously keeping theability to decode 100% of the packets.

With the reference protocols, the average packet losses arehigh, and, moreover, some of the nodes can have reasonablyhigh packet delivery probability while others have a very lowone. The CDF of the per-node packet delivery probability isshown in Fig. 6(a). Since the delivery probability of the net-work coding protocols is 100% for all nodes, we only show thereference protocols. With MS, a large number of sources havea very low delivery probability, severely limiting the usefulnessof the communication, and all reference protocols have a largenumber of sources with lower than 50% packet delivery prob-ability. The high reliability of the network coding represents ahuge improvement over the other protocols.

D. Varying Density of Households

We conducted a second experiment in order to evaluate the ef-fects of different densities of households around the secondarysubstations for NC-fb. We used the same topology as before,but with only 4, 8, or 16 households per secondary substation,also randomly distributed within a similar annulus around each.In spite of the higher loss probability for individual packets dueto the sparser distribution of the nodes, the BS was able to de-code 100% of the packets in all cases. The average collectiontimes are shown in Fig. 6(b). Although the number of packetsto decode is lower, the collection time is larger in the topologywith 4 households per secondary substation than in those with


Fig. 7. (a) IEEE 37 node test feeder. (b) Average percentage of collected packets. (c) Average collection time.

8 or 16, probably due to the higher packet loss probability thatstems from the larger average distance between the nodes. Inthe topologies with fewer than 32 households per substation,the graph constructed using a 70% reliability threshold becomesdisconnected, making it impossible to compute a routing matrixwhere all nodes can reach the sink. With 4 households per sub-station it becomes connected only using a threshold as low as10%. This is a limitation for non-coded protocols, where it isnecessary to have routes from all sources to the sink and thedelivery of each individual packet is important. For the NC-fbprotocol, only an indication of whether a node is downstreamis necessary, and that works even if this information is missingfor some nodes. This is a clear indication of the advantages ofnetwork coding in terms of robustness of the network.

E. Varying Reliability Threshold

In a third experiment, we evaluated the effects in the per-formance of the NC-fb of using different reliability thresholdsfor including each pairwise “link” in the graph. We use thetopology with 4 households per secondary substation and variedthe threshold from 10% to 90%. With more than 10%, the graphbecomes disconnected, meaning that some nodes have infinitedistance to the source. Such nodes are never considered to bedownstream from any other node, which can adversely impactthe implicit feedback mechanism (but not the delivery of codedpackets). The results of this experiment show that the effectdoes not hinder the protocol’s ability to deliver all data to thesink, and all packets are decoded in all instances. The collectiontime using routes computed with different reliability thresholdsis shown in Fig. 6(c). Even though some nodes become infinitelydistant from the sink with higher reliability threshold values,this affects only a small fraction of the nodes. This effect seemsto be compensated by the fact that larger reliability thresholdslead to longer paths in the routing tables, which, in turn, imply asharper vision of the relative location of the nodes with respectto the sink (upstream, same distance or downstream). Hence,the network coding protocols are robust to imprecisions in theestimation of the topology of the wireless network.

F. Smaller, Denser Topology

We conducted additional experiments with a topology corre-sponding to a smaller but denser urban area. This topology isbased on the IEEE 37 node test feeder [44], [45], illustrated in

Fig. 7(a), fitting in a rectangle of 1000 m per 1400 m. We alsoadded 32 households, distributed uniformly at randomwithin anannulus around each secondary substation, but with a minimumradius of 10 m and a maximum radius of 100 m. The remainingconditions are similar to the first experiment.The results are shown in Figs. 7(b) and (c). Compared to

Fig. 5(b) and (c), the performance of the reference protocolsis much better, due to shorter paths to the sink, and probablyalso to the closer proximity between the nodes. However, it stilldoes not even come close to that of the network coding proto-cols, particularly in terms of reliability.In an additional experiment, we varied the reliability

threshold for inclusion of a pairwise “link” in the graph usedfor computing the routes. Due to the higher density of thenodes, the graph is connected even when using a reliabilitythreshold of 90%. Thus, no nodes are prevented from everreaching/being reached from the BS due to the absence ofroutes in the reference protocols, that could meaningfullybe included here. Contrary to the network coding protocols,data collection in the non-coding reference protocols is highlyaffected by the inclusion of lower reliability links in the routes(and never reached 100%), as Fig. 8 shows. This aspect must bebalanced against the fact that using high reliability thresholdscan lead to disconnected graphs with unreachable sources. Thecollection delay is also affected by the threshold, particularly inMS, where the sink waits for a certain period for the arrival ofthe packet from each source until proceeding to the next source.Network coding protocols also benefit from using higherthresholds because the implicit feedback is more effective witha finer granularity notion of downstream nodes.

V. CONCLUSION

We proposed two network coding schemes based on tunablesparse codes, with and without explicit feedback from theBS, for wireless communication in smart grid, more exactlyadvanced meter readings in a MV network. Our results, testedon the IEEE 123 and IEEE 37 node test feeders with ns-2simulator, are strongly indicative of the advantages of usingnetwork coding protocols when comparing with other schemes,namely a master-slave approach used nowadays and someother data collection protocols. Its robustness against adversenetwork conditions (losses, low density networks) makes net-work coding protocols, especially the scheme with feedback,


Fig. 8. IEEE 37 node test feeder: (a) average percentage of collected packetsand (b) average collection time.

particularly well suited for data collection in smart grids. Withnetwork coding, the data packets are collected much soonerthan the defined deadline. Compared with the MS referenceprotocol, our proposed protocols exhibited a threefold (withoutexplicit feedback) to tenfold (with explicit feedback) improve-ment in the collection time.The protocol with explicit feedback can be further improved,

by discarding from the node’s queue the seen packets at the BS,instead of only the decoded packets.In the course of this work, it has become apparent that there is

a lot to be gained in re-designing the communication networksand protocols for smart grids, particularly by improving the reli-ability of commodity communication interfaces (such as WiFi)that have a low cost but do not yet meet the standards of utilitycompanies. In conclusion, our findings call for a promising op-portunity of using a network coding approach in smart grid ap-plications. Extensions to other network topologies or services insmart grids are part of our future work.

REFERENCES[1] S. Feizi, D. E. Lucani, and M. Médard, “Tunable sparse network

coding,” in Proc. Int. Zurich Seminar Commun., Mar. 2012, pp.107–110.

[2] “European electricity grid initiative roadmap and implementationplan,” Tech. Rep. May 2010, ENTSOE and EDSO.

[3] M. Baran and A. Kelley, “State estimation for realtime monitoring ofdistribution systems,” IEEE Trans. Power Syst., vol. 9, pp. 1601–1609,Aug. 1994.

[4] G. Heydt, “The next generation of power distribution systems,” IEEETrans. Smart Grid, vol. 1, pp. 225–235, Dec. 2010.

[5] G. Bumiller, L. Lampe, and H. Hrasnica, “Power line communica-tion networks for large-scale control and automation systems,” IEEECommun. Mag., vol. 48, pp. 106–113, Apr. 2010.

[6] “Status review on regulatory aspects of smart metering (electricity andgas) as of May 2009,” Tech. Rep., Oct. 2009, ERGEG.

[7] G. Deconinck, “An evaluation of two-way communication means foradvanced metering in Flanders (Belgium),” in Proc. IEEE IIMTC, May2008, pp. 1–6.

[8] D. Marihart, “Communications technology guidelines for EMS/SCADA systems,” IEEE Trans. Power Del., vol. 16, pp. 181–188,Apr. 2001.

[9] Q. Gao, J. Yu, P. Chong, P. So, and E. Gunawan, “Solutions for thesilent node problem in an automatic meter reading system using power-line communications,” IEEE Trans. Power Del., vol. 23, pp. 150–156,Jan. 2008.

[10] B. Sivaneasan, E. Gunawan, and P. So, “Modeling and performanceanalysis of automatic meter-reading systems using PLC under im-pulsive noise interference,” IEEE Trans. Power Del., vol. 25, pp.1465–1475, Jul. 2010.

[11] S. Galli, A. Scaglione, and Z.Wang, “For the grid and through the grid:The role of power line communications in the smart grid,” Proc. IEEE,vol. 99, pp. 998–1027, Jan. 2011.

[12] V. Gungor, B. Lu, and G. Hancke, “Opportunities and challenges ofwireless sensor networks in smart grid,” IEEE Trans. Ind. Electron.,vol. 57, pp. 3557–3564, Oct. 2010.

[13] P. McDaniel and S. McLaughlin, “Security and privacy challenges inthe smart grid,” IEEE Security Privacy, vol. 7, pp. 75–77, May/Jun.2009.

[14] C. Efthymiou and G. Kalogridis, “Smart grid privacy via anonymiza-tion of smart metering data,” in Proc. IEEE Smart. Grid. Commun.,Oct. 2010, pp. 238–243.

[15] M. Souryal, C. Gentile, D. Griffith, D. Cypher, and N. Golmie, “Amethodology to evaluate wireless technologies for the smart grid,” inProc. IEEE Smart Grid Commun., Oct. 2010, pp. 356–361.

[16] M. Hoch, “Comparison of plc g3 and prime,” in Proc. IEEE Int. Symp.Power Line Commun. Its Appl., Apr. 2011, pp. 165–169.

[17] Y. Phulpin, D. Lucani, and J. Barros, “Network coding in smart grids,”in Proc. IEEE Smart Grid Commun., Oct. 2011, pp. 37–42.

[18] R. Ahlswede, N. Cai, S.-Y. Li, and R. Yeung, “Network informationflow,” IEEE Trans. Inf. Theory, vol. 46, pp. 1204–1216, Jul. 2000.

[19] J. Widmer, C. Fragouli, and J. L. Boudec, “Low-complexity en-ergy-efficient broadcasting in wireless ad-hoc networks using networkcoding,” in Proc. NetCod, Apr. 2005.

[20] L. Keller, E. Drinea, and C. Fragouli, “Online broadcasting with net-work coding,” in Proc. NetCod, Jan. 2008, pp. 1–6.

[21] C. Fragouli, J. L. Boudec, and J. Widmer, “Network coding: An instantprimer,” ACM Comput. Commun. Rev., Jan. 2006.

[22] A. Kamra, V. Misra, J. Feldman, and D. Rubenstein, “Growth codes:Maximizing sensor network data persistence,” Sigcomm, Sep. 2006.

[23] S. Katti, H. Rahul, W. Huss, D. Katabi, M. Medard, and J. Crowcroft,“XORs in the air: Practical wireless network coding,” in Proc. ACMSigcomm, 2006.

[24] S. Chachulski, M. Jennings, S. Katti, and D. Katabi, “Trading structurefor randomness in wireless opportunistic routing,” in Proc. ACM SIG-COMM, 2007.

[25] H. Su and X. Zhang, “Network coding based qos-provisioning mac forwireless smart metering networks,” in Proc. 7th Int. Conf. Heteroge-neous Netw. Quality, Reliability, Security, Robustness, Nov. 2012.

[26] S. Güzelgöz, H. Arslan, A. Islam, and A. Domijan, “A review of wire-less and plc propagation channel characteristics for smart grid environ-ments,” J. Electr. Comput. Eng., vol. 2011, Jun. 2011.

[27] A. Goldsmith, Wireless Communications 2005.[28] M. Pätzold, “Mobile radio channel models for present and future wire-

less communication systems,” in Proc. Int. Conf. ATC, 2008.[29] H. Okamoto, K. Kitao, and S. Ichitsubo, “Outdoor-to-indoor propaga-

tion loss prediction in 800-mhz to 8-ghz band for an urban area,” IEEETrans. Veh. Technol., vol. 58, pp. 1059–1067, Mar. 2009.

[30] T. Sarkar, Z. Ji, K. Kim, A. Medouri, and M. Salazar-Palma, “A surveyof various propagation models for mobile communication,” AntennasPropag. Mag., vol. 45, pp. 51–82, Jun. 2003.


[31] J. Meng and A. A. Marbl, “Effective communication strategies fornoise limited power line channels,” IEEE Trans. Power Del., Jan. 2006.

[32] F. Canete, J. Cortés, L. Díez, and J. Entrambasaguas, “A channel modelproposal for indoor power line communications,” IEEE Commun.Mag., vol. 49, pp. 166–174, Dec. 2011.

[33] F. Kakar, M. Khalid, and F. Suri, “Enhanced outdoor-to-indoorcoverage estimation in microcells,” in Proc. LAPC, Mar. 2008, pp.421–424.

[34] C. Múller, H. Georg, M. Putzke, and C. Wietfeld, “Performance anal-ysis of radio propagation models for smart grid applications,” in Proc.IEEE Int. Conf. Smart Grid Commun., Oct. 2011, pp. 96–101.

[35] M. Korki, N. Hosseinzadeh, H. Vu, T. Moazzeni, and C. Foh, “Achannel model for power line communication in the smart grid,” inProc. IEEE Power Syst. Conf. Expo., Mar. 2011, pp. 1–7.

[36] R.Koetter andM.Medard, “An algebraic approach to network coding,”IEEE/ACM Trans. Netw., vol. 11, pp. 782–795, Oct. 2003.

[37] D. Lucani, M. Medard, and M. Stojanovic, “Underwater acoustic net-works: Channels models and network coding based lower bound totransmission power for multicast,” IEEE J. Sel. Areas Commun., vol.26, pp. 1708–1719, Dec. 2008.

[38] D. Lucani, M. Médard, and M. Stojanovic, “Network coding schemesfor underwater networks—The benefits of implicit acknowledgement,”inProc.WorkshopUnderwater Netw. (WuWNet), Sep. 2007, pp. 25–32.

[39] P. Chou, Y. Wu, and K. Jain, “Practical network coding,” in Proc. 41stAllerton Conf. Commun., Control, Comput., Oct. 2003.

[40] R. Motwani and P. Raghavan, Randomized Algorithms. Cambridge,U.K.: Cambridge Univ. Press, 1995.

[41] C. Fragouli, J. Widmer, and J.-Y. L. Boudec, “On the benefits of net-work coding for wireless applications,” in Proc. 4th Int. Symp. Mod-eling Optim. Mobile, Ad Hoc, Wireless Netw., 2006, pp. 1–6.

[42] A. Zaballos, A. Vallejo, M. Majoral, and J. Selga, “Survey and perfor-mance comparison of AMR over PLC standards,” IEEE Trans. PowerDel., vol. 24, pp. 604–613, Apr. 2009.

[43] C. Wan, S. Eisenman, A. Campbell, and J. Crowcroft, “Siphon: Over-load traffic management using multi-radio virtual sinks in sensor net-works,” in Proc. SenSys, 2005, pp. 116–129.

[44] W. Kersting, “Radial distribution test feeders,” IEEE Trans. PowerSyst., vol. 6, pp. 975–985, Aug. 1991.

[45] W. Kersting, “Radial distribution test feeders,” in Proc. IEEE PowerEng. Soc. Winter Meet., 2001, vol. 2, pp. 908–912.

[46] K. Fall and K. Varadhan, “The ns Manual (formerly ns Notes andDocumentation),” in The VINT Project, A Collaboration BetweenResearchers at UC Berkeley, LBL, USC/ISI, and Xerox PARC, May2010.

Rui Prior (S’06–M’07) received the Lic. and M.Sc.degrees in electrical and computer engineering fromthe Faculty of Engineering of the University of Porto,Portugal, in 1997 and 2001, respectively, and a Ph.D.in computer science from the Faculty of Sciences ofthe University of Porto in 2007. He has previouslyworked as a researcher at INESC-Porto and LIACC,and is currently an Assistant Professor at the Depart-ment of Computer Science of the Faculty of Sciencesof the University of Porto and a researcher at Institutode Telecomunicações. His research interests lie in the

broad field of computer networks (protocols, QoS, mobility, network coding).

Daniel E. Lucani (S’04–M’10) received his B.S.(summa cum laude) and M.S. (with honors) degreesin electronics engineering from Universidad SimónBolívar, Venezuela in 2005 and 2006, respectively,and the Ph.D. degree in electrical engineeringfrom the Massachusetts Institute of Technology,Cambridge, MA, USA (MIT) in 2010. He is anAssociate Professor in the Department of ElectronicSystems, University of Aalborg, Denmark. He wasan Assistant Professor at the Faculty of Engineeringof the University of Porto and a member of the

Instituto de Telecomunicações (IT) from April 2010 to July 2012 beforejoining Aalborg University. His research interests lie in the general areas ofcommunications and networks, network coding, information theory and theirapplications to highly volatile wireless sensor networks, satellite and under-water networks, focusing on issues of robustness, reliability, delay, energy,

and resource allocation. Prof. Lucani was a visiting professor at MIT. He isthe general co-chair of the 2014 International Symposium on Network Coding(NetCod2014) and was the general co-chair of the Network Coding Applica-tions and Protocols Workshop (NC-Pro 2011). Dr. Lucani has also served asTPC member for international conferences and as reviewer for high impactjournals, such as IEEE JOURNAL OF SELECTED AREAS IN COMMUNICATIONS,IEEE TRANSACTIONS ON INFORMATION THEORY, IEEE TRANSACTIONS ONCOMMUNICATIONS, IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS,and IEEE/ACM TRANSACTIONS ON NETWORKING.

Yannick Phulpin (S’06–M’10) graduated fromboth Supelec, France (2003), and TU Darmstadt,Germany (2004) in electrical engineering and re-ceived the Ph.D. degree from the Georgia Instituteof Technology, Atlanta, GA, USA (2009). He hasbeen successively an Assistant Professor at Supelec(2004–2010), a Senior Researcher at the INESCTEC, Portugal (2011), and a Researcher at EDFR&D, France (2011–). His research interest isenergy policy and decision-making in power systemoperation and planning, including aspects related

with modelling, control, optimization, and economics of energy systems.

Maricica Nistor received her Engineer’s Degree atTechnical University of “Gh. Asachi,” Electronicsand Telecommunications Faculty, Romania, in 2007,and her Master’s degree at the same university in2008. She was awarded a Doctoral Scholarshipfrom the Portuguese Foundation for Science andTechnology and she is currently pursuing the Ph.D.degree at the Universidade do Porto (UP), Porto,Portugal. She is a also a researcher of the Institutode Telecomunicações (IT Porto), where she has beensince 2008. In 2012, she was a Visiting Researcher at

the Cisco Systems, California. Her research interests include network coding,communication networks, wireless sensor networks, and smart grids.

João Barros (S’98–M’04–SM’11) received hisundergraduate education in Electrical and ComputerEngineering from the Universidade do Porto (UP),Portugal and Universitaet Karlsruhe, Germany,a performing arts degree in flute from the MusicConservatory of Porto, and the Ph.D. degree inelectrical engineering and information technologyfrom the Technische Universitaet Muenchen (TUM),Germany. He is an Associate Professor of Electricaland Computer Engineering at the University ofPorto and Founding Director of the Institute for

Telecommunications (IT) in Porto, Portugal, which counts almost 100 activemembers. He was a Fulbright scholar at Cornell University and has been aVisiting Professor with the Massachusetts Institute of Technology, Cambridge,MA, USA (MIT) since 2008. He also teaches at the Porto Business School andco-founded two recent startups, Streambolico and Veniam, commercializingwireless video and vehicular communication technologies, respectively. Be-tween 2009 and 2012, Dr. Barros served as National Director of the CarnegieMellon Portugal Program, a five-year international partnership funded by thePortuguese Foundation of Science and Technology, with a total budget of 56MEuros. In recent years, Dr. Barros has been Principal Investigator (PI) and Co-PIof numerous national, European, and industry funded projects, co-authoringone book and 145 research papers in the fields of networking, informationtheory and security, with a special focus on smart city technologies, networkcoding, physical-layer security, sensor networks, and intelligent transportationsystems. Dr. Barros has received several awards, including the 2010 IEEECommunications Society Young Researcher Award for the Europe, MiddleEast, and Africa region, the 2011 IEEE ComSoC and Information TheorySociety Joint Paper Award, the 2012 BES National Innovation Award, and astate-wide best teaching award by the Bavarian State Ministry of Sciences,Research and the Arts. Dr. Barros is frequently invited as an expert speaker byinternational organizations such as the European Commission, OECD, ITU,EuroDIG, and IEEE. He also works as an independent consultant for variousorganizations and projects. Dr. Barros is fluent in Portuguese, German, English,French, and Spanish.

Network Coding Protocols for Smart Grid Communications

Documents