You are on page 1of 10

Control Engineering Practice 92 (2019) 104075

Contents lists available at ScienceDirect

Control Engineering Practice


journal homepage: www.elsevier.com/locate/conengprac

Design and experimental testing of an optimization-based flow control


algorithm for Energy Harvesting Wireless Sensor Networks
Antonio Frezzetti 1 , Sabato Manfredi a ,∗
a Department of Electrical Engineering and Information Technology, University of Naples Federico II, Italy

ARTICLE INFO ABSTRACT


Keywords: In this paper a distributed flow control law is proposed to maximize throughput and to minimize energy
Wireless Sensor Networks consumption in Energy Harvesting Wireless Sensor Networks (EH-WSNs). We preliminary recast the control
Energy harvesting problem in terms of primal–dual optimization one taking into account the bandwidth and energy autonomy
Distributed optimization
node constraint. Then, we devise a distributed flow rate control implemented at each node that allows the
Distributed control
overall network to converge to the optimal solution of the original problem. The closed loop EH-WSN stability
Flow control algorithms
and convergence to the optimal equilibrium are proven. The effectiveness of the proposed control law in
terms of throughput and network lifetime performance is experimentally validated by a small representative
EH-WSN. The experimental results are in a good agreement with the theoretical ones.

1. Introduction densely deployed over a field, the spatial correlation of measurements


can be used to severely under-sample the sensed signal and reconstruct
In recent years, the progressive and inexorable downscaling of it with a non-linear technique known as compressive sensing (Como,
embedded sensing systems has been exploited for the development of Lovisari, & Savla, 2015). However, when the sampled information is
fully autonomous wireless sensor networks (WSNs). Consisting of a not spatially correlated, or the network is not sufficiently dense, a more
certain number of nodes capable of sensing environmental information general approach must be considered.
and communicate them to a gathering node (named Fusion center, In Zhang, Chen, Zhou, Chen, and Shen (2016), authors propose a
FC), WSNs are going to be overwhelming important in many modern control scheme aimed at balancing the residual energy among sen-
applications ranging from agriculture, home automation, healthcare, sors to guarantee network energy sustainability. The proposed control
networked cyber–physical systems (Manfredi, 2017; Shaikh & Zeadally, scheme consists to dynamically select transmitting nodes and their
2016), transportation and power systems.
transmission power according to the mutual residual energy of those
In most of those applications, WSNs are deployed in remote fields
involved into the transmission. The optimal node is selected in a
and once placed, it is desirable to reduce as much as possible the human
distributed manner with a selection system based on priority flags.
intervention for maintenance. Energy autonomy is the main critical
To achieve the goal, the receiving node analyses the residual energy
concern for WSNs (Curry & Smith, 2016; Mathuna, ODonnell, Martinez-
of each source using direct transmission. An higher priority flag is
Catala, Rohan, & OFlynn, 2008; Pourazarm & Cassandras, 2017; Rault,
assigned to the nearer source node with sufficient residual energy
Bouabdallah, & Challal, 2014). In Energy Harvesting WSN (EH-WSN),
renewable sources are used to extend the network lifetime (Anastasi, for the transmission. This optimization is, however, achieved to the
Conti, Di Francesco, & Passarella, 2009 and references within). Differ- detriment of network throughput that, along with the network lifetime,
ent approaches have been proposed to increase their lifetime. As the is to be maximized in WSN.
major energy consumption of a node is due to transmission power, Because node communication rate is in trade-off with the energy
maximization of network lifetime is achieved either by managing the consumption (Zhai & Fang, 2006), the design of an efficient EH-WSN
network data to reduce node’s throughput, either by controlling the requires the combined optimization of network throughput and energy
nodes’ transmission power. autonomy. Network throughput is limited by the maximum capacity
Reducing nodes’ throughput is possible with techniques consisting of node to receive data before being congested with a consequent
in compressing data before to send them (Du, Gkatzikis, Fischione, & loss of packets (Babayo, Anisi, & Ali, 2017). The harvested energy
Xiao, 2017; Frezzetti & Manfredi, 2015). In fact, when sensor nodes are constraint is achieved if the amount of energy spent for node does

∗ Corresponding author.
E-mail address: sabato.manfredi@unina.it (S. Manfredi).
1
Independent Scholar.

https://doi.org/10.1016/j.conengprac.2019.06.014
Received 11 December 2018; Received in revised form 24 April 2019; Accepted 18 June 2019
Available online 12 September 2019
0967-0661/© 2019 Elsevier Ltd. All rights reserved.
A. Frezzetti and S. Manfredi Control Engineering Practice 92 (2019) 104075

not exceed the available renewable energy (Besbes, Smart, Burana- distributed and converges to the optimal solution of the original primal
panichkit, Kloukinas, & Andreopoulos, 2013; Frezzetti, Manfredi, & problem.
Pagano, 2015; Pimentel & Musílek, 2010), this condition is addressed Differently, Authors in Nakayama et al. (2014) formulate an opti-
as the energy neutrality condition. In this paper, we formulate a dis- mization problem that maximizes the energy margin in battery. There-
tributed flow control law to maximize the network throughput while fore the control law accounts for the effect of the energy harvesting
energy neutrality condition and maximum bandwidth constraints are management (i.e. the residual energy in battery), instead of the en-
fulfilled. ergy harvested and available time by time at node. Additionally, we
formulate a distributed and low computational demanding algorithm
1.1. Related works that controls the WSN to an equilibrium point that is the optimal
solution of the original problem. Differently, in Dehwah et al. (2017)
In the last years the problem of flow control in communication and Nakayama et al. (2014) greedy algorithms are introduced in or-
networks has been extensively analysed and solved by different tech- der to overcome the high computational complexity of the original
niques (see i.e. Low, Paganini, & Doyle, 2002 and references therein). control with the loss of the optimality of the final equilibrium point.
A general approach is to optimize a network utility function 𝒰(𝐠) by Moreover, the strategies in Dehwah et al. (2017) and Nakayama et al.
setting the packet rate 𝑔𝑖 of a traffic source 𝑖 under rate capacity (2014) are routing-based as they dynamically reconfigure the network
constraints. This approach, originally formulated for ATM networks routing matrix in order to increase the network lifetime. Differently,
(see i.e. Low et al. (2002) and references therein) is as follows. Let 𝐠 be
we propose a flow-control based algorithm based on primal–dual opti-
the vector of nodes rate 𝑔𝑖 , 𝐑(𝐠) the vector of nodes’ incoming rate and
mization approach that controls at each node the communication rate
𝑐 = [𝑐𝑖 ] be the vector of the maximum output links capacity, then 𝐠 is
to trade off between data throughput and node power consumption.
determined by solving the following problem (Kelly, Maulloo, & Tan,
Others primal–dual based approaches are proposed in the literature as
1998; Low, 2000):
in Chiang (2005) and Zhang and Zheng (2006), where the main aim
max 𝒰(𝐠) (1) is to maximize the network throughput without explicitly accounting
𝐠
of the energy harvested at nodes. Differently, our node utility function
s.t. 𝐑(𝐠) ≤ 𝐜 includes both throughput and energy spent to support it in the commu-
where 𝒰(𝐠) is a concave utility function and 𝐑(𝐠) ≤ 𝐜 defines the nication. Moreover, we explicitly account for the energy harvested at
link capacity constraints. This formulation is attractive for WSN be- node and related constraint in the optimization problem formulation.
cause it can be recast to a dual formulation that yields a distributed As it will be shown by the experimental results, this yields to a closed
flow rate algorithm. This approach has been extensively proposed for loop WSN equilibrium point that trades off between throughput and
large scale wired and wireless communication networks (Akyildiz, Su, lifetime performance.
Sankarasubramaniam, & Cayirci, 2002; Du et al., 2011). Finally, we validate the proposed distributed algorithm on a rep-
In Nakayama et al. (2014) it is proposed an optimization problem resentative EH-WSN. A dynamic scenario under time-varying energy
subject to link capacity and power constraint aimed to balance nodes harvesting and link capacity conditions is considered. The results show
power consumption. The optimization variable is the routing matrix that under the considered dynamic scenarios, the controlled steady-
resulting in a discrete NP-hard problem. To significantly reduce the state rates are in a good agreement with the theoretical ones. We
topology complexity, authors proposed a sub-optimal solution by di- also show that our approach increases the lifetime of the network, as
viding the network into a set of (assumed independent) sub-networks compared with an algorithm not accounting for energy harvested at
and applying the optimization approach to each of them. In Dehwah, nodes.
Shamma, and Claudel (2017) the authors present a max–min problem The paper is organized as follows. In Section 2 the EH-WSN model
using both a Mixed Integer Linear Programming formulation and for- is provided, in Section 3 the primal–dual optimal flow control problem
ward dynamic programming approach. The objective is to maximize the formulation is presented. In Section 4 a distributed dynamic flow
lowest energy in the wireless sensor network at a fixed time horizon, control algorithm is given while the convergence and global asymptotic
under the sensing, communication and power constraints. Authors also stability analyses are presented in Section 5. In Section 6 the experi-
propose and validate a distributed greedy policy over the network mental validation of the proposed approach on a network composed
aimed to reduce the high computational complexity of the original of 6 nodes is carried out. Finally, conclusions and future work are
policy.
provided in Section 7.
In Chiang (2005) a distributed non linear control algorithm based
on primal–dual decomposition approach is formulated to regulate node
generation rate and transmission power in order to maximize the 1.3. Notation
network throughput under capacity constraint. Robustness evaluation
of the proposed algorithm in presence of packet loss and estimation of For the sake of clarity, we will adopt the following notation through-
local link capacity is carried out. out the manuscript.
In Zhang and Zheng (2006) the authors devise a distributed primal– 𝑎 = [𝑎𝑖 ] is a column vector of elements 𝑎𝑖 , 𝑏 = 𝑎 𝑇 is the transpose of
dual algorithm to optimize the throughput subject to rate constraints in 𝑎 and 𝑁 = {1, … , 𝑛} is a set of elements 1, … , 𝑛.
presence of communication noise. The convergence algorithm is proven 𝑑𝑓
Let a function 𝑓𝑖 (𝑡), 𝑓̇𝑖 (𝑡) = 𝑑𝑡𝑖 (𝑡) denotes its derivative function
both in deterministic and stochastic case. while 𝑓 (𝑡) = [𝑓𝑖 (𝑡)] (resp. 𝑓̇ (𝑡)) is a column vector of elements 𝑓𝑖 (𝑡) (resp.
𝑓̇𝑖 (𝑡)) at time 𝑡.
1.2. Contribution and paper outline Notation [𝑓𝑖 (𝑡)]+ denotes a function that equals 𝑓𝑖 (𝑡) if 𝑓𝑖 (𝑡) is
positive, otherwise it is equal to zero.
In this paper we propose a distributed flow control law to deal
with congestion and energy autonomy problem in EH-WSN. In partic-
ular, we firstly formulate a primal optimization problem to maximize 2. Energy harvesting wireless network model
communication rate under energy and output link capacity constraint.
Using the primal–dual decomposition approach, we devise a distributed In the following we will introduce the EH-WSN model considered
rate control law that maximizes network throughput accounting for throughout the paper presenting both data traffic and energy harvesting
the energy harvested at nodes during time. The resulting algorithm is dynamic formulations.

2
A. Frezzetti and S. Manfredi Control Engineering Practice 92 (2019) 104075

2.1. Data traffic model Assuming such a linear behaviour of 𝑝𝑖 (𝑡) in WSN is quite standard in
the literature and is suitable for most of real applications of WSNs (see
We consider a EH-WSN composed of 𝑛 energy harvested wireless i.e. Karakus, Gurbuz, & Tavli, 2013 and references therein). Throughout
sensor nodes and a fusion center (later shortly FC). Each node 𝑖 ∈ 𝑁 = the experiment, we will characterize the power consumption model
{1, … , 𝑛} sends data to the FC through a communication path composed of the adopted WSN nodes in order to validate and confirm this
of nodes assumption.
𝑁𝑝 (𝑖) = {𝑗 ∈ 𝑁|j belongs to node 𝑖’s path} with 𝑁𝑝 (𝑖) ⊂ 𝑁. The routing
matrix 𝐋 ⊂ R𝑛×𝑛 has elements 𝑙(𝑗, 𝑖) defined as: 3. Optimal flow control in EH-WSNs
{
1 𝑖𝑓 𝑗 ∈ 𝑁𝑝 (𝑖) 𝑜𝑟 𝑖𝑓 𝑗 ≡ 𝑖 In this section, we formulate the optimization problem for the
𝑙(𝑗, 𝑖) = (2)
0 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒 EH-WSN aimed to maximize the overall network utility while the con-
straints on energy autonomy and link congestion are fulfilled. Firstly,
𝐋 is assumed to be a full row rank matrix.
a primal–dual formulation of the original problem is given. Then, the
Let 𝑔𝑖 (𝑡) be the packet generation rate of node 𝑖 at time 𝑡, the
∑ corresponding dual-problem is used to decompose the original primal
aggregate input rate 𝑟𝑖 (𝑡) at node 𝑖 is defined as 𝑟𝑖 (𝑡) = 𝑗∈𝑁 𝑙(𝑖, 𝑗)𝑔𝑗 (𝑡).
problem into 𝑛 sub-problems to be individually solved by each node in
Notice that 𝑟𝑖 (𝑡) includes packets coming from the other nodes and
a decentralized fashion.
packets sampled and generated at node 𝑖. In the following for sake
of brevity, we will refer to 𝑟𝑖 (𝑡) as rate at node 𝑖. Each node 𝑖 with a
3.1. Primal–dual problem formulation
generation rate 𝑔𝑖 (𝑡) evaluates an aggregate price or cost 𝑞𝑖 (𝑡), defined

as 𝑞𝑖 (𝑡) = 𝑗∈𝑁 𝑙(𝑗, 𝑖)𝜌𝑗 (𝑡), where 𝜌𝑗 (𝑡) is a penalty function at node
In the proposed problem formulation, we will explicitly take into
𝑗. Let 𝐠(𝑡) = [𝑔1 (𝑡), … , 𝑔𝑛 (𝑡)]𝑇 be the generation rate vector, 𝐑(𝑡) =
account two realistic features of a node: the maximum node commu-
[𝑟1 (𝑡), … , 𝑟𝑛 (𝑡)]𝑇 the aggregate rate vector, 𝜌 (𝑡) = [𝜌1 (𝑡), … , 𝜌𝑛 (𝑡)]𝑇 the nication rate (related to the link capacity) and the available energy
penalty vector and 𝐪(𝑡) = [𝑞1 (𝑡), … , 𝑞𝑛 (𝑡)]𝑇 the aggregate penalty vector. (provided by the harvesting system). To this end, we consider each
With the definitions made, it results: node to be characterized by the following constraints: (1) the energy
𝐑(𝑡) = 𝐋 ⋅ 𝐠(𝑡) neutrality condition: 𝑝𝑖 (𝑡) = 𝜙𝑟𝑖 (𝑡) + 𝜀 ≤ ℎ̃ 𝑖 (𝑡), and (2) link congestion:
(3) 𝑟𝑖 (𝑡) ≤ 𝑐𝑖 . Indeed the term (ℎ̃ 𝑖 (𝑡) − 𝜀)∕𝜙 represents the maximum
𝐪(𝑡) = 𝐋𝑇 ⋅ 𝜌 (𝑡)
packet rate at time 𝑡 supported by the node according to the available
Notice that, according to (3), 𝐑(𝑡) is function of 𝐠(𝑡), so that 𝐑(𝑡) = harvested energy. For clearness of presentation, thereafter we will write
𝐑(𝐠(𝑡)). Functions 𝐠(𝑡) and 𝜌 (𝑡) are non-negative valued. In our frame- ℎ𝑖 in place of ℎ̃ 𝑖 − 𝜀. With the recent technology development, 𝜀 is
work, 𝐪(𝑡) and 𝐑(𝑡) are dynamically coupled by the following differen- usually negligible than the dynamic power consumption. Additionally,
tial equations: a well designed harvesting system is able to cope with the static power
̇ = 𝐅(𝐠(𝑡), 𝐪(𝑡))
𝐠(𝑡) consumption of a node. It is therefore reasonable to assume ℎ𝑖 = ℎ̃ 𝑖 −𝜀 >
(4) 0 such that the energy autonomy condition is fulfilled when 𝑟𝑖 (𝑡) = 0.
𝜌̇ (𝑡) = Γ(𝜌𝜌(𝑡), 𝐑(𝑡))
Therefore the above constraints can be recast in terms of the aggregate
where 𝐅 = [𝐹𝑖 (𝑔𝑖 (𝑡), 𝑞𝑖 (𝑡))] and Γ = [𝛾𝑖 (𝜌𝑖 (𝑡), 𝑟𝑖 (𝑡))] are the vector of rate 𝑟𝑖 (𝑡):
node generation rates 𝐹𝑖 and the vector of node penalty dynamics 𝛾𝑖 , ℎ𝑖
𝛿𝐹 𝑟𝑖 (𝑡) ≤
respectively. We assume 𝐹𝑖 continuously differentiable with 𝛿𝑞 𝑖 ≠ 0 𝜙
𝑖 (7)
in 𝐃 = {(𝑔𝑖 (𝑡), 𝑞𝑖 (𝑡))|𝑔𝑖 (𝑡) > 0, 𝑞𝑖 (𝑡) > 0}. Under this hypothesis, the im- 𝑟𝑖 (𝑡) ≤ 𝑐𝑖
plicit function theorem ensures that there exists a unique continuously
differentiable function 𝑓𝑖 ∶ 𝑔(𝑡) → 𝑞𝑖 (𝑡) such that:
Remark. In the considered framework each node is a source of
( ) information and is subject to capacity and energy constraints. The FC
𝐹𝑖 𝑔𝑖 (𝑡), 𝑓𝑖 (𝑔𝑖 (𝑡)) = 0 (5) ∑
is usually constrained by 𝑟𝐹 𝐶 (𝑡) ≤ 𝑐𝐹 𝐶 being 𝑟𝐹 𝐶 (𝑡) = 𝑖∈𝑁 𝑔𝑖 (𝑡), and
Notice that (4) represents a general state dynamic formulation of 𝑐𝐹 𝐶 the maximum network throughput that depends on the specific
generation rate and penalty function (depending on the congestion application requirement, network protocol and FC device technology.
level) at nodes. Similar assumptions and framework are defined in the
original papers (Kelly et al., 1998; Low, 2000). Therefore, by (7), at each time the maximum node rate is con-
strained by the 𝑚𝑖𝑛(𝑐𝑖 , ℎ𝑖 ∕𝜙). In order to include the above constraints in
the problem formulation, we introduce the set 𝑁𝑒 = {𝑖 ∈ 𝑁|𝑚𝑖𝑛(𝑐𝑖 , ℎ𝑖 ∕𝜙)
2.2. Energy model
= ℎ𝑖 ∕𝜙} and 𝑁𝑞 = {𝑖 ∈ 𝑁|𝑚𝑖𝑛(𝑐𝑖 , ℎ𝑖 ∕𝜙) = 𝑐𝑖 } such that 𝑁𝑞 ∩ 𝑁𝑒 ≡ ∅ and
𝑁𝑞 ∪ 𝑁𝑒 ≡ 𝑁.
A suitable function 𝐹𝑖 satisfying the assumption will be formulated
Let 𝐏𝑒 (𝑡) = {𝑝𝑖 (𝑡)} and 𝐇𝑒 = {ℎ̃ 𝑖 } for 𝑖 ∈ 𝑁𝑒 and 𝐑𝑞 = {𝑟𝑖 (𝑡)},
later on in this paper. The following integral model for the node sensor
𝐂𝑞 = {𝑐𝑖 } for all 𝑖 ∈ 𝑁𝑞 . The optimal node generation rate 𝐠 solution of
battery energy level 𝑒𝑖 (𝑡) is considered:
the following primal problem is:
𝑒̇ 𝑖 (𝑡) = ℎ̃ 𝑖 (𝑡) − 𝑝𝑖 (𝑡) (6) ∑
𝑛
max 𝒰(𝐠(𝑡)) = 𝑈𝑠 (𝑔𝑠 (𝑡)) (8)
where ℎ̃ 𝑖 (𝑡) is the power stored by the harvesting system and 𝑝𝑖 (𝑡) is 𝐠≥𝟎
𝑠=1
power consumed by the node at time 𝑡. Both 𝑝𝑖 (𝑡) and ℎ̃ 𝑖 (𝑡) can be easily s.t. 𝐏𝑒 (𝑡) ≤ 𝐇𝑒
measured by the harvesting system or by the state of change of the
𝐑𝑞 (𝑡) ≤ 𝐂𝑞
storage battery.
The proposed energy model is based on the following assumptions. with 𝑈𝑠 (𝑔𝑠 ) being the node utility functions of nodes source of in-
Firstly, we assume throughout the paper that ℎ̃ 𝑖 (𝑡) is a piecewise formation, assumed strictly concave. In the following of this paper,
constant function of time. This assumption is reasonable because the the optimal set solution of the primal problem will be denoted with
available level of energy harvesting is usually averaged over time to 𝐠∗ = [𝑔1∗ , … , 𝑔𝑛∗ ]𝑇 , and 𝒰 ∗ = 𝒰(𝐠∗ ) indicates the optimal value of the
filter out both noise and fast changes occurring at the renewable source. utility function.
Secondly, the power consumption 𝑝𝑖 (𝑡) is modelled as 𝑝𝑖 (𝑡) = 𝜙𝑟𝑖 (𝑡) + 𝜀, Notice that in the problem formulation (8), the harvested energy
where 𝜙 is the power consumption per packet [W/Packet] due to the ℎ̃ 𝑖 (𝑡) is considered constant during the time interval of problem solution
wireless transmission and 𝜀 is the static power consumption of node. as it is assumed to be a piecewise constant function of time.

3
A. Frezzetti and S. Manfredi Control Engineering Practice 92 (2019) 104075

The Lagrangian function of the primal problem (8) is: 4.1. Generation rate control 𝐅
∑ ∑ ( ) ∑ ( )
ℒ (𝐠, 𝜇 , 𝜃 ) = 𝑈𝑖 (𝑔𝑖 ) + 𝜃𝑗 ℎ̃ 𝑗 − 𝑝𝑗 + 𝜇𝑗 𝑐𝑗 − 𝑟𝑗 𝛿𝐹
𝑖∈𝑁 𝑗∈𝑁𝑒 𝑗∈𝑁𝑞
As stated above, 𝐹𝑖 must be continuously differentiable with 𝛿𝑞 𝑖 ≠ 0
𝑖
∑ ∑ ( ) ∑ ( ) (9) in 𝐃 = {(𝑔𝑖 (𝑡), 𝑞𝑖 (𝑡))|𝑔𝑖 (𝑡) > 0, 𝑞𝑖 (𝑡) > 0}, where 𝑔𝑖 (𝑡) is a non negative
= 𝑈𝑖 (𝑔𝑖 ) + 𝜃𝑗 ℎ𝑗 − 𝜙𝑟𝑗 + 𝜇𝑗 𝑐𝑗 − 𝑟𝑗 +
variable such that 𝑔𝑖 (𝑡) = [𝑔𝑖 (𝑡)] . We choose 𝐹𝑖 according to the
𝑖∈𝑁 𝑗∈𝑁𝑒 𝑗∈𝑁𝑞
following equation:
where 𝜇 = [𝜇𝑗 ], 𝜃 = [𝜃𝑗 ] are vectors of Lagrangian multipliers. 𝜙 ( 𝜙) 2
The Lagrangian (9) is recast as: 𝑔̇ 𝑖 (𝑡) = 𝐹𝑖 (𝑔𝑖 (𝑡), 𝑞𝑖 (𝑡)) = 1 − 𝑞𝑖 (𝑡) − − 𝛽1 𝑞𝑖 (𝑡) + 𝑔 (𝑡) (18)
𝜙0 𝜙0 𝑖
∑ ∑ ( )
ℒ (𝐠, 𝜌) = 𝑈𝑖 (𝑔𝑖 ) + 𝜌𝑗 𝑑𝑗 − 𝑟𝑗 (10) with 𝜙0 , 𝛽1 being positive coefficients with 𝜙0 is given in [W/packet]
𝑖∈𝑁 𝑗∈𝑁 and 𝛽1 is dimensionless. Eq. (18) is the dynamic of node generation
with 𝜌 = [𝜌𝑗 ] and 𝜌𝑗 defined as: rate that inversely depends on function of the congestion level and
{ transmission power consumption. The equilibrium points (𝑔 𝑖 , 𝑞 𝑖 ) of the
𝜙𝜃𝑗 𝑖𝑓 𝑗 ∈ 𝑁𝑒 chosen 𝐹𝑖 are calculated as follows:
𝜌𝑗 = (11)
𝜇𝑗 𝑖𝑓 𝑗 ∈ 𝑁𝑞 𝜙 ( 𝜙) 2 1 𝜙
1 − 𝑞𝑖 − − 𝛽1 𝑞 𝑖 + 𝑔 = 0 → 𝑞𝑖 = − (19)
and: 𝜙0 𝜙0 𝑖 1 + 𝛽1 ⋅ 𝑔 𝑖 2 𝜙0
{
ℎ𝑗 ∕𝜙 𝑖𝑓 𝑗 ∈ 𝑁𝑒 Definition of 𝑓𝑖 in (5) yields to:
𝑑𝑗 = (12)
𝑐𝑗 𝑖𝑓 𝑗 ∈ 𝑁𝑞 1 𝜙
𝑓𝑖 (𝑔𝑖 ) = − (20)
Let 𝒬(𝜌𝜌) = max𝐠≥𝟎 ℒ (𝐠, 𝜌 ), it results: 1 + 𝛽1 ⋅ 𝑔𝑖2 𝜙0
[∑ ∑ ( )] Assumptions on 𝜙0 and 𝛽1 yield 𝑓𝑖 (𝑔𝑖 ) to be a strictly decreasing
𝒬(𝜌𝜌) = max 𝑈𝑖 (𝑔𝑖 ) + 𝜌𝑗 𝑑𝑗 − 𝑟𝑗 function of 𝑔𝑖 (see Appendix A for details). This condition is crucial to
𝐠≥𝟎
𝑖∈𝑁 𝑗∈𝑁
∑ [ ∑ ] ∑ (13) prove convergence and stability property of the resulting closed loop
= max 𝑈𝑖 (𝑔𝑖 ) − 𝑔𝑖 ⋅ 𝑙(𝑖, 𝑗)𝜌𝑗 + 𝜌𝑗 𝑑𝑗 WSN.
𝑔𝑖 ≥0
𝑖∈𝑁 𝑗∈𝑁 𝑗∈𝑁

such that we may formulate the following dual problem: 4.2. Penalty function Γ

min 𝒬(𝜌𝜌) = min max ℒ (𝐠, 𝜌 ) (14)


𝜌 ≥𝟎 𝜌 ≥𝟎 𝐠≥𝟎 As stated in the previous sections, Γ is a generic functional depend-
ing on the Lagrangian multipliers 𝜌𝑖 (𝑡). Notice that the Lagrangian (9)
Dual formulation in Eq. (14) can be solved with an algorithmic
is function of variables 𝜃𝑖 (𝑡) and 𝜇𝑖 (𝑡) that are related to 𝜌𝑖 (𝑡) according
procedure that splits the primal problem over 𝑁 local sub-problems.
to Eq. (11) and Eq. (12). Therefore, in what follows, with slightly abuse
In particular, for a given set of 𝜌 , 𝐪 is calculated via 𝜌 by 𝑞𝑖 =
∑ of notation we will make use of 𝜌𝑖 (𝑡), 𝜃𝑗 (𝑡) and 𝜇𝑗 (𝑡) in interchangeably
𝑗∈𝑁 𝑙(𝑗, 𝑖)𝜌𝑗 . The controlled rate 𝑔𝑖 derives by the following local
way.
optimization problem (Eq. (13)):
[ ] Let the constraints in (7). For each 𝑖th node we have the following
max 𝑈𝑖 (𝑔𝑖 ) − 𝑔𝑖 ⋅ 𝑞𝑖 (15) constraints:
𝑔𝑖 ≥0
(𝑎) 𝑝𝑖 (𝑡) = 𝜙𝑟𝑖 (𝑡) ≤ ℎ𝑖
Finally, let 𝐠 be the new aggregate rate set of the locally optimal 𝑔𝑖 ,
∑ (𝑏) 𝑟𝑖 (𝑡) ≤ 𝑐𝑖
then 𝑟𝑖 = 𝑗∈𝑁 𝑙(𝑖, 𝑗)𝑔𝑗 is locally computed and the value of 𝜌 updated
throughout 𝛤 as in Eq. (4)). Now we consider nodes 𝑖 ∈ 𝑁𝑒 , ℎ𝑖 ∕𝜙 < 𝑐𝑖 .
The optimal solution of the primal problem is achieved locally at In this case,constraint (a) implies (b) and the following coupled
each node by iteratively repeating the above procedure. This makes dynamic (𝜂𝑖 (𝑡), 𝜌𝑖 (𝑡)) at the node are proposed:
the dual form in Eq. (14), very attractive in WSN since it can be solved {
(𝜙𝑟𝑖 (𝑡) − ℎ𝑖 ) 𝜂𝑖 (𝑡) > 0
in a distributed fashion. 𝜂̇ 𝑖 (𝑡) =
[𝜙𝑟𝑖 (𝑡) − ℎ𝑖 ] + 𝜂𝑖 (𝑡) = 0
Notice that with the hypothesis on concavity of 𝑈𝑖 (𝑔𝑖 ), the optimal
{ (21)
solution 𝑔𝑖∗ of local problem (15) satisfies: 𝛽2 (𝛼2 𝜂𝑖 (𝑡) + 𝜙𝑟𝑖 (𝑡) − ℎ𝑖 ) 𝜌𝑖 (𝑡) > 0
𝜌̇ 𝑖 (𝑡) = 𝛾𝑖 (𝜌𝑖 (𝑡), 𝑟𝑖 (𝑡)) =
𝑑
𝑈 (𝑔 ∗ ) − 𝑞𝑖 = 0 (16) [𝛽2 (𝛼2 𝜂𝑖 (𝑡) + 𝜙𝑟𝑖 (𝑡) − ℎ𝑖 )]+ 𝜌𝑖 (𝑡) = 0
𝑑𝑔𝑖 𝑖 𝑖
Notice that under the aforementioned condition 𝑖 ∈ 𝑁𝑒 , we have that
Moreover for Eq. (16), 𝑞𝑖 is given and 𝑔𝑖∗ depends on 𝑞𝑖 . We can 𝜌𝑖 (𝑡) = 𝜙𝜃𝑖 (𝑡) according to (12). 𝜂𝑖 (𝑡) in Eq. (21) represents the dynamic
therefore define the functional: of the battery energy and is used to increase the node energy penalty
𝑑 function when the consumed power exceeds the harvested one (𝑖 ∈ 𝑁𝑒 ).
𝑧𝑖 (𝑞𝑖 ) ∶ 𝑞𝑖 → 𝑔𝑖∗ , so that 𝑈 (𝑔 ∗ ) = 𝑞𝑖 , (17)
𝑑𝑔𝑖 𝑖 𝑖 In (21) 𝛼2 , 𝛽2 > 0 are positive constants, being 𝛼2 in [1/s], 𝛽2 is in
with 𝑧𝑖 (𝑞𝑖 ) = 𝑔𝑖∗ being the optimal solution of local problem Eq. (15). [packet/s W]. Let finally define 𝜁2 = 𝛼2 𝛽2 . Notice that 𝜂𝑖 (𝑡) and 𝜌𝑖 (𝑡) are
Notice that having 𝑈𝑖 (𝑔𝑖 ) strictly concave, it implies 𝑧𝑖 (𝑞𝑖 ) to be a strictly not negative functions and, in particular, 𝜌𝑖 (𝑡) = [𝜌𝑖 (𝑡)]+ .
decreasing function of 𝑞𝑖 . Equilibrium analysis of (21): let (𝜂 𝑖 , 𝜌𝑖 ) be an equilibrium point for
Let 𝐳(𝐪) = 𝑑𝑖𝑎𝑔{𝑧𝑖 (𝑞𝑖 )} be a diagonal matrix with 𝑧𝑖 (𝑞𝑖 ) defined in (21). We note that 𝜂 𝑖 cannot be positive. Indeed, assuming 𝜂 𝑖 > 0, from
( ) ℎ
Eq. (17). By Eq. (3) and Eq. (17) it results: 𝐑 = 𝐋 ⋅ 𝐳 𝐪 . 𝜂𝑖 (𝑡) dynamics, it results 𝑟𝑖 = 𝜙𝑖 ; consequently, 𝜌̇ 𝑖 = 0. As a result,
𝜂 𝑖 = 0, in contrast with the assumption on 𝜂 𝑖 . On the other hand, if

4. Distributed flow control law it is assumed 𝜂 𝑖 = 0, we have 𝑟𝑖 ≤ 𝜙𝑖 and from dynamic on 𝜌𝑖 (𝑡) two
ℎ ℎ
equilibrium conditions result: 𝜌𝑖 = 0 if 𝑟𝑖 < 𝜙𝑖 or 𝜌𝑖 > 0 if 𝑟𝑖 = 𝜙𝑖 .
In this section we present the distributed control laws implemented Therefore, we can conclude that the unique equilibrium point is

at each node 𝑖 to regulate the generation rate 𝑔𝑖 and evaluate the 𝜂 𝑖 = 0 and 𝜌𝑖 ≥ 0, while 𝜌𝑖 > 0 occurs for nodes with 𝑟𝑖 = 𝜙𝑖 . Notice
penalty function 𝜌𝑖 to solve the dual problem (14). To this end, we will that having 𝛼2 > 0 and 𝛽2 > 0 ensure the uniqueness of the equilibrium
define distributed version of dynamic equations 𝐅 and Γ introduced in point.
(4). Continuing with our analysis, we now consider the case 𝑖 ∈ 𝑁𝑞 .

4
A. Frezzetti and S. Manfredi Control Engineering Practice 92 (2019) 104075

Fig. 1. (a) Network topology considered for the experimental validation scenario; (b) A Wireless Sensor Node used for the experimental validation (photo by courtesy of Meditel
Spa., Salerno, Italy).

The following coupled dynamics (𝜔𝑖 (𝑡), 𝜌𝑖 (𝑡)) are proposed: 6. Experimental validation
{
(𝑟𝑖 (𝑡) − 𝑐𝑖 ) 𝜔𝑖 (𝑡) > 0
𝜔̇ 𝑖 (𝑡) = We consider the network topology in Fig. 1(a). Node 1 is the FC,
[𝑟𝑖 (𝑡) − 𝑐𝑖 ]+ 𝜔𝑖 (𝑡) = 0 nodes 2–3–6 directly communicate with FC, node 5 is connected to FC
{ (22)
𝛽3 (𝛼3 𝜔𝑖 (𝑡) + 𝑟𝑖 (𝑡) − 𝑐𝑖 ) 𝜌𝑖 (𝑡) > 0 by node 3, while node 4 can send packets to FC through node 3 or node
𝜌̇ 𝑖 (𝑡) = 6. With a such topology, node 3 and 6 have both an higher energy
[𝛽3 (𝛼3 𝜔𝑖 (𝑡) + 𝑟𝑖 (𝑡) − 𝑐𝑖 )]+ 𝜌𝑖 (𝑡) = 0
consumption spent for communication than node 2, and as matter of
where 𝛼3 , 𝛽3 > 0 are positive constants, respectively in [1/s] and in fact, they are supplied by an EH system. Moreover, node 4 can split its
[packet/s W]. 𝜁3 = 𝛼3 𝛽3 is defined. traffic between node 3 and 6 according to their congestion levels and
Notice that, according to (11), 𝜌𝑖 (𝑡) = 𝜇𝑖 (𝑡) for 𝑖 ∈ 𝑁𝑞 . ability/inability to receive traffic. Hereafter we show the experimental
𝜔𝑖 (𝑡) in (22) is the dynamic of the queue length used to increase the validation of the proposed distributed control flow algorithm over the
congestion penalty function when the node incoming traffic exceeds the above network under dynamic variations of generation rate and energy
link capacity (𝑖 ∈ 𝑁𝑞 ). harvested at nodes.
Equilibrium analysis of (22): let (𝜔𝑖 , 𝜌𝑖 ) be an equilibrium point for
(22). With similar arguments used for (21), it results: 6.1. Experiment setup
– 𝜌𝑖 = 0 if 𝑟𝑖 < 𝑐𝑖
– 𝜌𝑖 > 0 if 𝑟𝑖 (𝑡) = 𝑐𝑖 The used sensor nodes for this experiment are shown in Fig. 1(b) and
being the uniqueness of solution with 𝜔𝑖 = 0, while 𝜌𝑖 > 0 occurs only are mainly equipped with: (1) a transceiver module for communication
for nodes with 𝑟𝑖 (𝑡) = 𝑐𝑖 . In the proof is supposed 𝛼3 > 0 and 𝛽3 > 0. (i.e. DigiMesh’s XBEE module); (2) an energy harvesting system to
supply energy provided by solar source (i.e. demoboard DC2048A
5. Convergence and stability results from Linear Technology); (3) a low power micro-controller Microchip’s
PIC18F to implement the proposed distribute algorithms.
In this section, convergence to the optimal solution of the primal The communication module is the Series 1 of XBEE DigiMesh RF
problem (8) is enunciated. Additionally, the global asymptotic stability based on 2.4 GHz of the 802.15.4 standard protocol with a nominal
of the given control distributed law (18), (21) and (22) is also stated. transmission rate of 250 kb/s. Being the physical packet size set to 1024
bits, it results: 𝑔̂ = 244 packets/s. Despite XBEE modules require 50 mA
Theorem 1. Let 𝑓𝑖 (𝑔𝑖 ) defined as in (20), the equilibrium of the closed at 3.3 V in transmission state, such RF modules are characterized by an
loop EH-WSN dynamic (18), (21) and (22) equals the optimal solution of embedded power sleep protocol with low current consumption lower
the primal problem (8) with utility function 𝑈𝑖 (𝑔𝑖 ): than 50 μA. These features make XBEE technology attractive to be used
1 √ 𝜙 in low power applications such as in the case of EH-WSNs.
𝑈𝑖 (𝑔𝑖 ) = 𝑓 (𝑔 )𝑑𝑔 = √ ⋅ 𝑎𝑟𝑐𝑡𝑔( 𝛽1 ⋅ 𝑔𝑖 ) − ⋅𝑔 (23)
∫ 𝑖 𝑖 𝑖 𝛽 𝜙0 𝑖 Recalling the energy model of Section 2.2, parameters 𝜙 and 𝜀
1
need to be characterized. Therefore, the average power consumption
Theorem 2. The (unique) equilibrium point of the closed loop EH-WSN 𝑝𝑖 of sensors in Fig. 1(b) for different transmission rates has been
dynamic (18), (21) and (22) is stable in the Lyapunov sense. Moreover, measured (solid stars in Fig. 2) and a 𝑀𝐿𝑆 linear approximation is
if the routing matrix 𝐋 is of a full row rank, then the dynamic system is derived (continuous line). As a result, we have: 𝑝𝑖 (𝑡) = 𝜙𝑟𝑖 (𝑡) + 𝜀 with
globally asymptotically stable. 𝜙 = 3.93 mW/packet ∈ (0, 1) and 𝜀 = 14 mW being the average static
power consumption of the node.
The proofs of Theorems 1–2 are deferred to Appendices A–B for the The considered experimental setup (Fig. 3) is composed of 6 nodes
sake of readability. (Fig. 1(b)) connected according to the topology in Fig. 1(a) with node 1
The assessment
( 𝜙 of )the result requires the fulfilment of the condi- being the FC. During the experiments, network throughput and source
tions: 𝛽1 ≤ 12 𝜙0 − 1 , 𝜁2 > 0 and 𝜁3 > 0. It is worth to note that rate values have been time by time acquired through the FC.
𝑔̂𝑖
the considered utility (23) is a general function depending on the node
throughput and energy consumption. It may represent the utility of 6.2. Flow control implementation
node that tries to maximize its transmission rate (first term of (23))
accounting for the derived transmission power consumption (second The proposed distributed control laws (18)–(21)–(22) are imple-
term of (23)). However, other choice of the utility function can be mented with the given values of 𝜙 = 3.93 mW/packet and 𝑔̂ = 244
made according to the specific problem to weigh in different way the packets/s, and with: 𝛽1 = 0.0038,
√ 𝜙(0 = 0.9, 𝜁2 = 𝜁3 = 2. Notice that
throughput term than the energy consumption one, provided that the )
1 𝜙0
given parameters yield to 𝑔̂𝑖 = −1 .
strictly concavity assumption is satisfied. 𝛽 𝜙 1

In what follows, we experimentally validate the proposed dis- The distributed version of the proposed control law will be imple-
tributed control laws by a EH-WSN composed of off-the-shelf compo- mented according to the discussion in Section 3.1. The implementation
nents. The aim is to validate the effectiveness of the proposed method involves the evaluation of penalty function 𝜌𝑖 , according its level of
in representative EH-WSN scenario. congestion (i.e. link capacity constraint) and energy state condition

5
A. Frezzetti and S. Manfredi Control Engineering Practice 92 (2019) 104075

Table 1
ℎ𝑖 and 𝑐𝑖 considered for the experiments during five different time periods.
(ℎ1 , 𝑐1 ) (ℎ2 , 𝑐2 ) (ℎ3 , 𝑐3 ) (ℎ4 , 𝑐4 ) (ℎ5 , 𝑐5 ) (ℎ6 , 𝑐6 )
𝑡 ∈ (0, 30) min (XX,100) (XX,30) (25,50) (XX,15) (XX,15) (25,50)
𝑡 ∈ [30, 60) min (XX,100) (XX,10) (25,50) (XX,15) (XX,15) (25,50)
𝑡 ∈ [60, 90) min (XX,100) (XX,10) (00,50) (XX,15) (XX,15) (25,50)
𝑡 ∈ [90, 120) min (XX,100) (XX,10) (25,50) (XX,15) (XX,15) (00,50)
𝑡 ∈ [120, 150] min (XX,100) (XX,10) (25,50) (XX,15) (XX,15) (25,50)

Table 2
Optimal solution of the primal problem applied to the considered experimental setup
in Section 6: aggregate rate for FC node (𝑟1 ) and source rates 𝑔𝑖 for nodes from
2 to 6.
𝑟1 𝑔2 𝑔3 𝑔4 𝑔5 𝑔6
𝑡 ∈ (0, 30) min 100 17.5 17.5 30 17.5 17.5
𝑡 ∈ (30, 60) min 100 10 17.5 30 17.5 25.0
𝑡 ∈ [60, 90) min 100 10 0.0 15 0.0 35.0
Fig. 2. Characterization of node consumption 𝑝𝑖 (in Watt) as function of 𝑟𝑖 : 𝑡 ∈ [90, 120] min 100 10 17.5 15 17.5 0.0
experimental data (solid stars), fitting curve (continuous line). 𝑡 ∈ [120, 150] min 100 10 17.5 30 17.5 25.0

variables for each node is depicted. Fig. 4(a) shows the comparison
between the generation rates 𝐠(𝑡) provided by the proposed distributed
algorithm (continuous lines) and the corresponding optimal solutions of
the original primal problem (dashed lines). Notice that the steady state
values assessed by the proposed algorithm asymptotically converge to
the optimal solutions given in Table 2. In Fig. 4(b) and (c) the aggregate
rate 𝐑(𝑡) and network throughput 𝑟1 are shown, respectively. In what
follows, we will describe the experimental results depicted in Fig. 4.
During the first time interval (𝑡 ∈ [0, 30) min) the controlled WSN is
able to achieve the required maximum throughput (i.e. 𝑟1 = 𝑐1 = 100).
At the end of the first time slice (𝑡 = 30 min) a reduction of
generation rate at node 2 occurs, yielding an higher generation rate of
node 6. Under this condition, the network still achieves the maximum
throughput (i.e. 𝑟1 = 100).
At 𝑡 = 60 min, the harvested energy at node 3 becomes null,
causing node 3 to be unable to route packets towards the FC. In this
Fig. 3. Picture of the experiment carried out at the Embedded System Laboratory,
University Federico II of Naples. situation, node 5 is inhibited to send packets, while node 4 can still
send information to FC through the node 6. Accordingly, generation
rate at node 5 is zero for 𝑡 ∈ [60, 90), while source rate at node 4 is still
positive, although partially reduced than the previous time interval.
(energy neutrality condition). The regulation of node generation rate
Finally, in order to maximize network throughout, node 6 increases
𝑔𝑖 is based on feedback variables 𝑞𝑖 (function of 𝜌𝑖 ). In particular, if
its transmission rate. Notice that the achieved network throughput is
𝑞𝑖 > 0, 𝑔𝑖 is reduced, otherwise is increased. Equilibrium between 𝑞𝑖
𝑟1 = 70 in this time interval 𝑡 ∈ [60, 90), lower that the maximum one
and 𝑔𝑖 is reached when the optimal value of 𝑔𝑖 is allocated. A software-
(𝑐1 = 100).
architecture is developed to implement the control distributively at
For 𝑡 ≥ 90 min, light intensity increases at node 3 while becomes
each node.
zero at node 6. In this case, the rates of node 3 and 5 increase, while
6.3. Experimental results the rate of node 6 decreases to zero. Node 4 increases the transmitted
packets to node 3. Notice that the throughput value equals the value
The experimental validation of the proposed distributed control obtained during the previous period of time (𝑡 ∈ [60, 90)). In other
algorithm is carried out at the Embedded System Laboratory of Uni- terms, the control algorithm regulates the generation rates according
versity of Naples Federico II. The implemented setup of the network to the dynamical variations of network conditions.
in Fig. 1(a) is shown in Fig. 3. Nodes 3 and 6 are equipped with four At the final time slice (𝑡 ≥ 120 min) light intensity at node 6 turns
solar harvesting cells Sanyo AM-5907 (250 mW, 5 V) and the energy is back to the initial value. In this case the distributed algorithm regulates
provided by a tunable source of light. Remaining nodes are supplied by all the rates achieving again the maximum throughput (𝑟1 = 𝑐1 = 100).
battery. To each node a duple of values (ℎ𝑖 , 𝑐𝑖 ) describes time by time The experiment results show that the proposed algorithm based on
the boundary conditions of the average harvested energy and the link primal–dual approach is able to maximize the network throughput,
capacity, respectively. During the experiment, this duple of constraints while congestion is avoided at each node. Moreover the generation
changes 5 times according to the values summarized in Table 1. In bold rates at nodes are regulated according to the available harvested energy
is highlighted the more restrictive constraint between ℎ𝑖 and 𝑐𝑖 for each fulfilling the energy neutrality condition and, thus, increasing the
duple (ℎ𝑖 , 𝑐𝑖 ). overall network lifetime. To further demonstrate the positive impact
For the FC, the energy harvesting constraint is not active (i.e. ℎ1 = of this algorithm feature on EH-WSN network performance, a second
𝑋𝑋), and 𝑐1 = 100 corresponding to the network maximum achievable experiment under the same conditions as in Table 1 has been conducted
throughput. ℎ𝑖 = 𝑋𝑋 also for nodes 𝑖 = 2, 4, 5, because they are not after being inhibited the capability of the algorithm to detect changes
equipped with the energy harvesting system. in the harvested energy. This also restates the primal–dual approaches
The optimal solution of primal problem for the considered WSN is where only the capacity constraints are considered in the optimization
reported in Table 2, while in Fig. 4 the dynamic evolution of the state problem formulation. Results are reported in Fig. 5. From Fig. 5(a)

6
A. Frezzetti and S. Manfredi Control Engineering Practice 92 (2019) 104075

Fig. 4. Experimental results: (a) generation rates 𝐠(𝑡) (continuous lines) and the respective optimal solutions of the primal problem (dashed lines), (b) aggregate rate 𝐑(𝑡) and (c)
network throughput 𝑟1 .

we note a variation in the rates of node at time instant 𝑡 = 30 min Acknowledgement


consequent to the reduction of rate at node 2. This is in accordance
with the results in Fig. 4. In particular, node 3 receives data from nodes This work was supported in part by the EPRSC-ENCORE Network+
4 and 5 and has generation rate 17.5 packet/s. After time instant 𝑡 = 60 project ‘‘Dynamics and Resilience of Multilayer Cyber-Physical So-
min, energy harvested at node 3 decreases to zero. However, because in cial Systems’’ and in part by DIETI departmental projects "Fair" and
this case the control algorithm does not take into account of the energy "CAMDEN TOWN", University of Naples Federico II, Italy.
harvesting changes, node 3 continues to generate packets at a rate of
Declaration of competing interest
17.5 pack/s and manages incoming traffic from nodes 4 and 5. This
yields node 3 to quickly run out of energy and forcing the experiment to
None declared.
stop. Therefore differently to the results in Fig. 4, in this experiment the
network lifetime is reduced at 83 min. On the other hand, the network Appendix A. Proof of Theorem 1
throughput is controlled to the maximum value 𝑐1 = 100 (Fig. 5(c)).
This remarks the feature of the proposed algorithm to trade off between
throughput and network lifetime WSN performance. Proof. It is known that the optimal solution of the dual problem
coincides with the 𝐠∗ of the primal problem with duality gap equal to
7. Conclusion and future work zero if the Karush–Kuhn–Tucker (KKT) conditions are met (Proposition
5.1.5 in Bertsekas, 1997). In Low (2003) authors show that for the
In this paper, a distributed control algorithm for energy harvesting problem in the form (1) KKT conditions are met for a dynamic system
wireless sensor networks to cope with both link congestion and energy in the form (4) provided that dynamic equations (18), (21) and (22)
neutrality condition has been proposed. By applying the primal–dual satisfy the following conditions:
decomposition method, we devised a distributed control algorithm such
1. For all 𝑖 ∈ 𝑁, 𝐹𝑖 and 𝛾𝑖 are non-negative functions so that 𝑔𝑖
that the closed loop system converges to the optimal solution of the and 𝑟𝑖 are non-negative variables. Moreover, equilibrium points
primal problem. Convergence and global asymptotic stability have been of system equations, namely 𝑔̄𝑖 , exist.
analytically proven. As proof of concept of the proposed optimization- 𝛿𝐹
2. For all 𝑖, 𝐹𝑖 are continuously differentiable and 𝛿𝑞 𝑖 ≠ 0 in the set
based distributed algorithm, an experiment validation has been carried 𝑖
{(𝑔𝑖 , 𝑞𝑖 )|𝑔𝑖 > 0; 𝑞𝑖 > 0}. Moreover, 𝑓𝑖 in (20) is a non increasing
out on a small but representative EH-WSN under varying energy source
function of 𝑞𝑖 .
irradiation and rate. It is shown that with our algorithm the link
3. Let (𝜌̄𝑖 , 𝑟̄𝑖 ) denote the penalty and the aggregate rate at node 𝑖 at
capacity is maximized while source rates dynamically are controlled
the equilibrium. If 𝛾𝑖 (𝜌̄𝑖 , 𝑟̄𝑖 ) = 0 then 𝑟̄𝑖 ≤ 𝑐𝑖 with equality if 𝜌̄𝑖 > 0.
according to the available energy from the harvesting supply system.
4. For all 𝑖, 𝑓𝑖 are strictly decreasing functions.
Remarkably, the experimental results are in a good agreement with the
optimal theoretical ones, assessing the effectiveness of the proposed We show that the closed loop system (18), (21) and (22) satisfies
algorithm and its implementation over a real WSN. We have also shown the above hypotheses.
that our algorithm outperforms the lifetime of the network, compared Hypothesis 1. holds since equilibrium points of (18), (21) and (22)
with the case when the energy harvesting constraints are not taken exist as demonstrated and 𝜌𝑖 , 𝑔𝑖 are positive state variables. Addition-
into account at a cost of network throughput reduction. This remarks ally, also hypothesis 3. holds as shown above in the manuscript (Sec.
the feature of the algorithm to trade off between throughput and 4.2).
network lifetime WSN performance. Although our experiment includes Hypotheses 2. and 4. hold as well. Indeed, function 𝐹𝑖 in (18) is by
𝛿𝐹
a small network, results suggest the promising application of the pro- its definition continuously differentiable and 𝛿𝑞 𝑖 ≠ 0 for all (𝑔𝑖 , 𝑞𝑖 ) ∈ 𝐃.
𝑖
posed approach to EH-WSNs. Extension of the experimental validation Moreover, 𝑓𝑖 in (20) is a strictly decreasing function since
and performance evaluation comparison of the proposed approach are 𝑑𝑓𝑖 2𝛽1 ⋅ 𝑔𝑖
=− <0
object of ongoing work. 𝑑𝑔𝑖 (1 + 𝛽1 ⋅ 𝑔 2 )2
𝑖

7
A. Frezzetti and S. Manfredi Control Engineering Practice 92 (2019) 104075

Fig. 5. Experimental results when harvesting node condition is not taken into account in the algorithm implementation: (a) generation rates 𝐠(𝑡) (continuous lines) and the
respective optimal solutions of the primal problem (dashed lines), (b) aggregate rate 𝐑(𝑡) and (c) network throughput 𝑟1 .

for each 𝑔𝑖 > 0, being 𝛽1 a positive constant. Consequently, 𝑈𝑖 defined 𝒱 (𝜂𝜂 , 𝜔 , 𝜌 ) is a non negative function being the sum of non negative
in (23) is strictly concave. The KKT conditions hold and Theorem 1 is terms. Moreover, 𝒱 = 0 at equilibrium (𝜂𝜂 , 𝜔 , 𝜌 ) = (𝜂𝜂 , 𝜔 , 𝜌 ) since:
demonstrated. □ – 𝜂 𝑗 = 0, 𝜌𝑗 ≥ 0, with 𝜌𝑗 > 0 for nodes with ℎ𝑗 − 𝜙𝑟𝑗 = 0.
– 𝜔𝑗 = 0, 𝜌𝑗 ≥ 0, with 𝜌𝑗 > 0 for nodes with 𝑐𝑗 − 𝑟𝑗 = 0.
– 𝛷𝑖 (𝑞 𝑖 ) = 0. Additionally, 𝒱 is radially unbounded, since for each
Appendix B. Proof of Theorem 2
(𝜂𝜂 , 𝜔, 𝜌) such that ‖(𝜂𝜂 , 𝜔, 𝜌)‖ → +∞, 𝒱 → +∞. Finally, 𝒱 is a non
increasing function of time. Therefore, it can be concluded that the
Inspired by the stability analysis provided in Paganini (2001), the provided system is stable in Lyapunov sense. To show that 𝒱 is a
proof is composed of two parts: (1) firstly, we show that the proposed non-increasing function, we derive 𝒱 in respect to time variable. We
dynamic system is stable in the Lyapunov sense; (2) secondly, under have:
the assumption that 𝐋 is a full row rank matrix, the global asymptotic ( ) ( )
∑ ( ) ∑ ( )
stability is demonstrated by the LaSalle invariance principle. ̇
𝒱 = 𝜁2 𝜂𝑗 𝜂̇ 𝑗 + ℎ𝑗 − 𝜙𝑟𝑗 𝜌̇ 𝑗 + 𝜙 𝜁3 𝜔𝑗 𝜔̇ 𝑗 + 𝑐𝑗 − 𝑟𝑗 𝜌̇ 𝑗
𝑗∈𝑁𝑒 𝑗∈𝑁𝑞
( )
Proof. Let preliminary define the following non negative integral ∑( ) ∑( )
function 𝛷𝑖 (𝑞𝑖 ) for all node 𝑖: +𝜙 𝑟𝑗 − 𝑟𝑗 𝜌̇ 𝑗 + 𝑟𝑗 − 𝑟𝑗 𝜌̇ 𝑗
𝑗∈𝑁𝑒 𝑗∈𝑁𝑞
𝑞𝑖 ( ) ∑𝑛
𝛷𝑖 (𝑞𝑖 ) = 𝑔 𝑖 − 𝑧𝑖 (𝜎) 𝑑𝜎 (24) where for the derivative of 𝑖=1 𝛷𝑖 (𝑞𝑖 ) we have used the result in (25).
∫𝑞 𝑖
It follows:
where 𝑔 𝑖 is the steady-state rate at node 𝑖, while 𝑧𝑖 (𝑞𝑖 ) is defined in (17) ( ) ( )
∑ ( ) ∑ ( )
and is a decreasing function of 𝑞𝑖 . 𝛷𝑖 (𝑞𝑖 ) is then non negative for all 𝑞𝑖 ̇
𝒱 = 𝜁2 𝜂𝑗 𝜂̇ 𝑗 + ℎ𝑗 − 𝜙𝑟𝑗 𝜌̇ 𝑗 + 𝜙 𝜁3 𝜔𝑗 𝜔̇ 𝑗 + 𝑐𝑗 − 𝑟𝑗 𝜌̇ 𝑗
and 𝛷𝑖 (𝑞𝑖 ) = 0 if and only if 𝑞𝑖 = 𝑞 𝑖 . 𝑗∈𝑁𝑒 𝑗∈𝑁𝑞
∑ ∑ ∑
Let study the derivative over time of function 𝑖 𝛷𝑖 (𝑞𝑖 ): = 𝑣𝑒(𝑗) + 𝑣𝑞 (𝑗)
( 𝑛 ) ∑ 𝑛 ( ) 𝑛 ( ) ( )𝑇 𝑗∈𝑁𝑒 𝑗∈𝑁𝑞
𝑑 ∑ ∑
𝛷𝑖 (𝑞𝑖 ) = 𝑔 𝑖 − 𝑧𝑖 (𝑞𝑖 ) 𝑞̇ 𝑖 = 𝑔 𝑖 − 𝑧𝑖 (𝑞𝑖 ) 𝑞̇ 𝑖 = 𝐠 − 𝐠 ⋅ 𝐪̇
𝑑𝑞𝑖 𝑖=1 (27)
𝑖=1 𝑖=1

being 𝑇 the transpose of its argument and where the definition 𝑔𝑖 = Now we show that 𝑣𝑒(𝑗) for all 𝑗 ∈ 𝑁𝑒 and 𝑣𝑞 (𝑗) for all 𝑗 ∈ 𝑁𝑞 are non
𝑧𝑖 (𝑞𝑖 ) is used. From (3) it follows: positive functions.
Let analyse 𝑣𝑒(𝑗) by dividing its domain of definition (𝜂𝑗 , 𝜌𝑗 ) into four
( 𝑛 ) ( )𝑇 ( )𝑇 ( )𝑇
𝑑 ∑ regions.
𝛷𝑖 (𝑞𝑖 ) = 𝐠 − 𝐠 ⋅ 𝐪̇ = 𝐠 − 𝐠 ⋅ 𝐋𝑇 𝜌̇ = 𝐑 − 𝐑 ⋅ 𝜌̇ =
𝑑𝑞𝑖 𝑖=1 – Region 1. (𝜂𝑗 = 0 ∧ 𝜌𝑗 > 0). According to model (21) it results:
∑( ) ∑( ) (25)
𝑟𝑗 − 𝑟𝑗 𝜌̇ 𝑗 + 𝑟𝑗 − 𝑟𝑗 𝜌̇ 𝑗 𝑣𝑒(𝑗) = (ℎ𝑗 − 𝜙𝑟𝑗 )𝛽2 (𝜙𝑟𝑗 − ℎ𝑗 ) < 0
𝑗∈𝑁𝑒 𝑗∈𝑁𝑞
– Region 2. (𝜂𝑗 = 0 ∧ 𝜌𝑗 = 0).
Now we propose the following suitable Lyapunov function for the
dynamical system (18), (21) and (22). Our candidate function is: 𝑣𝑒(𝑗) = (ℎ𝑗 − 𝜙𝑟𝑗 )𝛽2 max{0, (𝜙𝑟𝑗 − ℎ𝑗 )}
( ) In this case, if 𝜙𝑟𝑗 < ℎ𝑗 then 𝑣𝑒(𝑗) = 0. On the other hand, if 𝜙𝑟𝑗 > ℎ𝑗 it
∑ 𝜂𝑗2 ( )
𝒱 (𝜂𝜂 , 𝜔 , 𝜌 ) = 𝜁2 + ℎ𝑗 − 𝜙𝑟𝑗 𝜌𝑗 results 𝑣𝑒(𝑗) = −𝛽2 (ℎ𝑗 − 𝜙𝑟𝑗 )2 < 0.
2
𝑗∈𝑁𝑒 – Region 3. (𝜂𝑗 > 0 ∧ 𝜌𝑗 > 0). According to the dynamic model (21)
( ) (26)
∑ 𝜔2𝑗 ( ) ∑𝑛 we obtain:
+𝜙 𝜁3 + 𝑐 𝑗 − 𝑟 𝑗 𝜌𝑗 + 𝜙 𝛷𝑖 (𝑞𝑖 )
𝑗∈𝑁
𝑞
2 𝑖=1 𝑣𝑒(𝑗) = 𝜁2 𝜂𝑗 𝜂̇ 𝑗 + (ℎ𝑗 − 𝜙𝑟𝑗 )𝛽2 (𝛼2 𝜂𝑗 + 𝜙𝑟𝑗 − ℎ𝑗 ) = ⋯ = −𝛽2 (ℎ𝑗 − 𝜙𝑟𝑗 )2 < 0

8
A. Frezzetti and S. Manfredi Control Engineering Practice 92 (2019) 104075

– Region 4. (𝜂𝑗 > 0 ∧ 𝜌𝑗 = 0). We have: Finally from definitions (3), (4), (17) and (30) we have:

𝑣𝑒(𝑗) = 𝜁2 𝜂𝑗 𝜂̇ 𝑗 + (ℎ𝑗 − 𝜙𝑟𝑗 )𝛽2 max{0, (𝛼2 𝜂𝑗 + 𝜙𝑟𝑗 − ℎ𝑗 )} 𝐪(𝑡) = 𝐋𝑇 (𝜌𝜌0 + 𝜌 ′ ⋅ 𝑡)


( )
( ℎ𝑗 −𝜙𝑟𝑗 ) 𝐠(𝑡) = 𝐳 𝐋𝑇 (𝜌𝜌0 + 𝜌 ′ ⋅ 𝑡) (32)
Now, if 𝛼2 𝜂𝑗 + 𝜙𝑟𝑗 − ℎ𝑗 ≤ 0 i.e. 𝜂𝑗 ≤ 𝛼2
i.e. it results: ( )
𝐑(𝑡) = 𝐋 ⋅ 𝐳 𝐋𝑇 (𝜌𝜌0 + 𝜌 ′ ⋅ 𝑡)
𝑣𝑒(𝑗) = 𝛽2 𝛼2 𝜂𝑗 (𝜙𝑟𝑗 − ℎ𝑗 ) ≤ −𝛽2 (ℎ𝑗 − 𝜙𝑟𝑗 )2 < 0
where 𝐳 = 𝑑𝑖𝑎𝑔{𝑧𝑖 (𝑞𝑖 (𝑡))} as in (17). Notice that for all 𝑡 > 0 it results:
On the other hand, if 𝛼2 𝜂𝑗 + 𝜙𝑟𝑗 − ℎ𝑗 < 0 we have: ( )𝑇 () ∑ ′
𝜌 ′ ⋅ 𝐑̇ 𝑡 = 𝜌𝑗 𝑟̇ 𝑗 (𝑡) = 0 (33)
𝑣𝑒(𝑗) = 𝜁2 𝜂𝑗 (𝜙𝑟𝑗 − ℎ𝑗 ) + (ℎ𝑗 − 𝜙𝑟𝑗 )𝛽2 (𝛼2 𝜂𝑗 + 𝜙𝑟𝑗 − ℎ𝑗 ) = −𝛽2 (ℎ𝑗 − 𝜙𝑟𝑗 )2 ≤ 0 𝑗∈𝑁

A similar analysis can be done for 𝑣𝑞 (𝑗) and it will be omitted for Indeed if 𝜌′𝑗
> 0, 𝑟𝑗 (𝑡) is constant since for all 𝑗 ∈ 𝑁𝑒 condition (28).(a)
brevity. In conclusion, 𝑣𝑒(𝑗) and 𝑣𝑞 (𝑗) are non positive definite functions holds and for all 𝑗 ∈ 𝑁𝑞 condition (29).(a) holds as well.
∀𝑗 and this proves that 𝒱 is a non increasing function, since 𝒱̇ ≤ 0. Substituting (32) in (33) we have:
To prove the global asymptotic stability of the system, we will invoke ( )𝑇 ( ) ( )𝑇 ( )
𝜌 ′ ⋅ 𝐑̇ 𝜌 = 𝜌 ′ ⋅ 𝐋 ⋅ 𝐳̇ 𝐋𝑇 (𝜌𝜌0 + 𝜌 ′ ⋅ 𝑡) ⋅ 𝐋𝑇 𝜌 ′ = 0 (34)
the LaSalle invariance principle (Khalil, 2002; Paganini, 2001). To this
end, let define the equilibrium set ℰ = {(𝜂𝑗 , 𝜔𝑗 , 𝜌𝑗 ) ∶ 𝒱̇ = 0}, ℰ𝑒 = Noting that 𝐳̇ = 𝑑𝑖𝑎𝑔{𝑧̇ 𝑖 (𝑞𝑖 (𝑡))} is strictly negative for all of its compo-
{(𝜂𝑗 , 𝜌𝑗 ), 𝑗 ∈ 𝑁𝑒 ∶ 𝑣𝑒(𝑗) = 0} and ℰ𝑞 = {(𝜔𝑗 , 𝜌𝑗 ), 𝑗 ∈ 𝑁𝑞 ∶ 𝑣𝑞 (𝑗) = 0}. nents and the above equation holds if and only if 𝐋𝑇 𝜌 ′ = 0, or, invoking
Since 𝒱̇ can be decomposed as in (27), we have that ℰ = ℰ𝑒 ∪ ℰ𝑞 . the hypothesis of full rank on 𝐋, 𝜌 ′ = 0. This implies that for all 𝑗 ∈ 𝑁𝑒
To show the global asymptotic stability of the system, we have to show 𝜂𝑗 (𝑡) = 0, and for all 𝑗 ∈ 𝑁𝑞 , 𝜔𝑗 (𝑡) = 0 and for all 𝑗 ∈ 𝑁 𝜌𝑗 (𝑡) = 𝜌𝑗 (0). In
that the largest invariant set ℐ contained in the set ℰ coincides with the other terms, we have shown that the largest invariant set ℐ contained
equilibrium point of system (21)–(22), or equivalently that no solution in ℰ is composed of just one point, (𝜂𝜂 , 𝜔 , 𝜌 ) = (𝜂𝜂 , 𝜔 , 𝜌 ) that also coincides
of system can stay identically in ℰ other than the equilibrium point of with the equilibrium point of system (21) (22).
system (21)–(22). Finally, recalling that 𝒱 is radially unbounded as shown above,
Let 𝜔 = [𝜔𝑗 (𝑡)], 𝜂 = [𝜂𝑗 (𝑡)], 𝜌 = [𝜌𝑗 (𝑡)] for all 𝑗 ∈ 𝑁 so that from the LaSalle invariance principle it derives that system (21) (22) is
(𝜔𝜔, 𝜂 , 𝜌 ) ∈ ℰ , 𝜂 0 = 𝜂 (0), and 𝜌 0 = 𝜌 (0), and 𝜂 𝑒 (𝑡) = [𝜂𝑗 (𝑡)], 𝜌 𝑒 = [𝜌𝑗 (𝑡)] globally asymptotically stable. This completes the proof. □
with 𝑗 ∈ 𝑁𝑒 vectors of initial system states so that (𝜂𝜂 𝑒 , 𝜌 𝑒 ) ∈ ℰ𝑒 , and
𝜂 0𝑒 = 𝜂 𝑒 (0), and 𝜌 0𝑒 = 𝜌 𝑒 (0). Similarly, we denote with 𝜔 𝑞 (𝑡) = [𝜔𝑗 (𝑡)], References
𝜌 𝑞 (𝑡) = [𝜌𝑗 (𝑡)] with 𝑗 ∈ 𝑁𝑞 vectors of initial system states so that
Akyildiz, I. F., Su, W., Sankarasubramaniam, Y., & Cayirci, E. (2002). Wireless sensor
(𝜂𝜂 𝑞 , 𝜌 𝑞 ) ∈ ℰ𝑞 . and 𝜂 0𝑞 = 𝜂 𝑞 (0), and 𝜌 0𝑞 = 𝜌 𝑞 (0). Let consider a trajectory in
networks: a survey. Computer Networks, 38(4), 393–422, Elsevier.
the set ℰ𝑒 . In this case, we may have: Anastasi, G., Conti, M., Di Francesco, M., & Passarella, A. (2009). Energy conservation
ℎ𝑗 in wireless sensor networks: A survey. Ad Hoc Networks, 7(3), 537–568, Elsevier.
(𝑎) 𝑟𝑗 = or Babayo, A. A., Anisi, M. H., & Ali, I. (2017). A review on energy management schemes
𝜙 in energy harvesting wireless sensor networks. Renewable and Sustainable Energy
(28)
ℎ𝑗 Reviews, 76, 1176–1184, Elsevier.
(𝑏) 𝑟𝑗 < if 𝜌𝑗 = 𝜂𝑗 = 0 Bertsekas, D. P. (1997). Nonlinear programming. Journal of the Operational Research
𝜙
Society, 48(3), 334–334.
for all 𝑗 ∈ 𝑁𝑒 . Now we show that the system state space belonging to ℰ𝑒 Besbes, H., Smart, G., Buranapanichkit, D., Kloukinas, C., & Andreopoulos, Y. (2013).
eventually evolves along a trajectory given by: 𝜂 𝑒 = 𝜂 0𝑒 and 𝜌 𝑒 = 𝜌 0𝑒 +𝜌𝜌′𝑒 ⋅𝑡, Analytic conditions for energy neutrality in uniformly-formed wireless sensor
networks. IEEE transactions on wireless communications, 12(10), 4916–4931.
where 𝜌 ′𝑒 = 𝜁2 ⋅ [𝜂𝑗 (0)] for all 𝑗 ∈ 𝑁𝑒 . Specifically for each 𝑗 we have:
Chiang, M. (2005). Balancing transport and physical layers in wireless multihop
networks: jointly optimal congestion control and power control. IEEE Journal on
1. if 𝜂𝑗 (0) = 0, from (28) and 𝑟𝑗 (𝑡) ≤ ℎ𝑗 ∕𝜙 it results: 𝜂𝑗 (𝑡) = 0
Selected Areas in Communications, 23(1), 104–116, IEEE.
and 𝜌𝑗 (𝑡) = 𝜌𝑗 (0). Moreover, from system equation (21) it results Como, G., Lovisari, E., & Savla, K. (2015). Throughput optimality and overload behavior
𝜌̇ 𝑗 (𝑡) = 0 for all time 𝑡, and 𝜌𝑗 (𝑡) = 𝜌𝑗 (0). of dynamical flow networks under monotone distributed routing. IEEE Transactions
on Control of Network Systems, 2(1), 57–67.
2. if 𝜂𝑗 (0) > 0 then 𝜂𝑗 (𝑡) = 𝜂𝑗 (0) and 𝜌𝑗 = 𝜌𝑗 (0) + 𝜁2 𝜂𝑗 (0) ⋅ 𝑡. Indeed,
Curry, R. M., & Smith, J. C. (2016). A survey of optimization algorithms for wireless
with 𝜂𝑗 (0) > 0, we fall again to the case (a) of (28). sensor network lifetime maximization. Computers and Industrial Engineering, 101,
145–166, Elsevier.
Similarly for ℰ𝑞 and related condition 𝑣𝑞 (𝑗) = 0 it results: Dehwah, A. H., Shamma, J. S., & Claudel, C. G. (2017). A distributed routing scheme
for energy management in solar powered sensor networks. Ad Hoc Networks, 67,
(𝑎) 𝑟𝑗 = 𝑐𝑗 or 11–23, Elsevier.
(29)
(𝑏) 𝑟𝑗 < 𝑐𝑗 if 𝜌𝑗 = 𝜔 𝑗 = 0 Du, R., Gkatzikis, L., Fischione, C., & Xiao, M. (2017). On maximizing sensor network
lifetime by energy balancing. IEEE Transactions on Control of Network Systems.
for all 𝑗 ∈ 𝑁𝑞 . With similar considerations given above for the set ℰ𝑒 , Du, P., Zou, W., Zhou, Z., Gao, W., Huang, X., & Xin, H. (2011). An active congestion
with ℰ𝑞 we may infer: 𝜔 𝑞 = 𝜔 0𝑞 and 𝜌 𝑞 = 𝜌 0𝑞 + 𝜌 ′𝑞 ⋅ 𝑡, where 𝜌 ′𝑞 = 𝜁3 ⋅ 𝜔 0𝑞 help mechanism in wireless sensor network. In Vehicular technology conference,
2011, IEEE (pp. 1–5). IEEE.
for all 𝑗 ∈ 𝑁𝑞 .
Frezzetti, A., & Manfredi, S. (2015). Evaluation of energy efficiency-reconstruction
Finally, from (27) it results that 𝒱̇ = 0 such that 𝑣𝑒(𝑗) = 0 for error trade-off in the co-design of compressive sensing techniques for wireless lossy
all 𝑗 ∈ 𝑁𝑒 and 𝑣𝑞 (𝑗) = 0 for all 𝑗 ∈ 𝑁𝑞 . Therefore, the state space sensor networks. International Journal of Wireless Information Networks, 22, 386–398,
𝜔0 , 𝜂 0 , 𝜌 0 ) with a trajectory
𝜔, 𝜂 , 𝜌 ) ∈ ℰ evolves from the initial state (𝜔 Springer.
(𝜔
Frezzetti, A., Manfredi, S., & Pagano, M. (2015). A design approach of the solar
staying inside ℰ and characterized by: harvesting control system for wireless sensor node. Control Engineering Practice, 44,
45–54, Elsevier.
⎧𝜔 = 𝜔 0
⎪ Karakus, C., Gurbuz, A. C., & Tavli, B. (2013). Analysis of energy efficiency of
⎨𝜂 = 𝜂 0 (30) compressive sensing in wireless sensor networks. IEEE Sensors Journal, 13(5),
⎪ 0 ′ 1999–2008, IEEE.
⎩𝜌 = 𝜌 + 𝜌 ⋅ 𝑡 Kelly, F. P., Maulloo, A. K., & Tan, D. K. (1998). Rate control for communication net-
works: shadow prices, proportional fairness and stability. Journal of the Operational
where 𝜌 ′ = [𝜌′𝑗 ] is a column vector of 𝑛 angular coefficients given by:
Research Society, JSTOR, 237–252.
{ Khalil, H. K. (2002). Nonlinear systems. Upper Saddle River.
′ 𝜁2 ⋅ 𝜂𝑗 (0) if𝑗 ∈ 𝑁𝑒 Low, S. H. (2000). A duality model of TCP flow controls. In Proceedings of ITC specialist
𝜌𝑗 = (31)
𝜁3 ⋅ 𝜔𝑗 (0) if𝑗 ∈ 𝑁𝑞 seminar on IP traffic measurement, modeling and management, Vol. 2.

9
A. Frezzetti and S. Manfredi Control Engineering Practice 92 (2019) 104075

Low, S. H. (2003). A duality model of tcp and queue management algorithms. Pourazarm, S., & Cassandras, C. G. (2017). Energy-based lifetime maximization and
IEEE/ACM Transactions on Networking, 11(4), 525–536. security of wireless-sensor networks with general nonideal battery models. IEEE
Low, S. H., Paganini, F., & Doyle, J. C. (2002). Internet congestion control. Control Transactions on Control of Network Systems, 4(4), 793–804.
Systems, 22(1), 28–43, IEEE. Rault, T., Bouabdallah, A., & Challal, Y. (2014). Energy efficiency in wireless sensor
Manfredi, S. (2017). Multilayer control of networked cyber-physical systems: application networks: A top-down survey. Computer Networks, 67, 104–122, Elsevier.
to monitoring, autonomous and robot systems. Advances in Industrial Control, Shaikh, F. K., & Zeadally, S. (2016). Energy harvesting in wireless sensor networks: A
Springer. comprehensive review. Renewable and Sustainable Energy Reviews, 55, 1041–1054,
Mathuna, C. O., ODonnell, T., Martinez-Catala, R. V., Rohan, J., & OFlynn, B. (2008).
Elsevier.
Energy scavenging for long-term deployable wireless sensor networks. Talanta,
Zhai, H., & Fang, Y. (2006). Distributed flow control and medium access in multihop
75(3), 613–623, Elsevier.
ad hoc networks. IEEE Transactions on Mobile Computing, 5(11), 1503–1514.
Nakayama, K., Dang, N., Bic, L., Dillencourt, M., Bozorgzadeh, E., & Venkatasubra-
Zhang, D., Chen, Z., Zhou, H., Chen, L., & Shen, X. S. (2016). Energy-balanced
manian, N. (2014). Distributed flow optimization control for energy-harvesting
wireless sensor networks. In 2014 IEEE international conference on communications cooperative transmission based on relay selection and power control in energy
(ICC), 2014, IEEE (pp. 4083–4088). IEEE. harvesting wireless sensor network. Computer Networks, 104, 189–197, Elsevier.
Paganini, F. (2001). On the stability of optimization-based flow control. In Proceedings Zhang, J., & Zheng, D. (2006). A stochastic primal-dual algorithm for joint flow control
of the American control conference, 2001, IEEE, Vol. 6 (pp. 4689–4694). IEEE. and mac design in multi-hop wireless networks. In 40th Annual conference on
Pimentel, D., & Musílek, P. (2010). Power management with energy harvesting devices. information sciences and systems, 2006, IEEE (pp. 339–344). IEEE.
In 23rd Canadian conference on electrical and computer engineering (CCECE), 2010,
IEEE (pp. 1–4). IEEE.

10

You might also like