Control Engineering Practice 98 (2020) 104364

Contents lists available at ScienceDirect

Control Engineering Practice

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

Robust fuzzy model predictive control for energy management systems in

fuel cell vehicles
Di Shen ∗, Cheng-Chew Lim, Peng Shi
School of Electrical and Electronic Engineering, The University of Adelaide, Adelaide, South Australia, 5005, Australia

ARTICLE INFO ABSTRACT

Keywords: Fuel cell vehicle combines the benefits of fuel cell stack and energy storage system to achieve fuel economy and
Energy management zero-emission. The energy management system is vital to the hybrid vehicle systems since it regulates power
Hybrid vehicle flow from the fuel cell stack and energy storage system. In this paper, we design an energy management scheme
Fuzzy control
in fuel cell vehicle systems. By using optimal control principle, we aim to reduce hydrogen consumption while
Model predictive control
maintaining battery state of charge under practical operating constraints and uncertain future power demand.
Fuel cell
Electric vehicle
The Fuzzy modelling approach is employed to describe the nonlinear power plant and a robust model predictive
based control is designed to achieve the desired system performance. Moreover, traffic condition is incorporated
into the energy management controller design to further improve the system performance. The effectiveness
and advantages of the proposed control scheme are illustrated by a simulator developed based on real-world
experimental data.

1. Introduction
To improve fuel economy and use alternative fuel other than
petroleum for transportation, hydrogen-powered fuel cell vehicles
(FCVs) have been an industrial focus in recent years (Burke, 2007;
Ehsani, Gao, Longo, & Ebrahimi, 2018). In an FCV, the power system
comprises a fuel cell stack and energy storage system (such as batteries
or ultracapacitors) to drive the vehicle. The power control unit is vital
to the hybrid system since it determines the distribution of power from
the fuel cell stack and energy storage system.
To design the power control unit, there are several challenging con-
trol system requirements (Hrovat, Di Cairano, Tseng, & Kolmanovsky,
2012; Xue, Yang, Zhang, & Shi, 2017):
• How to guarantee stability and robustness of the nonlinear power
system in FCVs under different driving and weather conditions?
• How to achieve optimality of the power distribution in FCVs
under practical operating constraints and unknown future power
demand?
• How to tackle the nonlinear optimisation problem in real-time by
an embedded controller?
Model predictive control (MPC) has attracted attention for power
control unit design for FCVs and hybrid vehicles as the control tech-
nique is capable of handling a constrained system to achieve optimal
control with certain performance criteria. Model inaccuracy and dis-
turbances, which result in the difference between the plant output and
the one described by the model, can be compensated by taking new
measurements at each time step in MPC.
Most published MPC-based research places emphasis on developing
power control units with a predictive model to achieve high control
performance. For example, in Di Cairano, Bernardini, Bemporad, and
Kolmanovsky (2014), the power management problem is cast in the
form of a stochastic MPC optimisation problem, and power request
is represented as the disturbance in the system and predicted by a
Markov chain to improve optimisation performance. In Zeng and Wang
(2015), a stochastic MPC approach with road information is proposed
to design the power control unit in hybrid vehicles, and the problem
is solved by stochastic dynamic programming. In Payri, Guardiola, Pla,
and Blanco-Rodriguez (2014), the energy management problem is cast
in the form of a finite optimal problem where the future power demand
is estimated to minimise the fuel consumption while maintaining bat-
tery state of charge. In our previous work (Shen, Lim, Shi, & Bujlo,
2018), a Markov chain Monte Carlo sampling-based approximator is
developed to predict long term average power demand, and the energy
management problem is then solved by convex programming. These
MPC-based control approaches have shown significant potential for
FCV energy management system design, but several challenges remain
before their practical use in automotive applications. These practical
challenges include:
• The control system may perform very poorly when the future
power demand is not well described by the model. In FCVs,

∗ Corresponding author.
E-mail addresses: di.shen@adelaide.edu.au (D. Shen), cheng.lim@adelaide.edu.au (C.-C. Lim), peng.shi@adelaide.edu.au (P. Shi).

https://doi.org/10.1016/j.conengprac.2020.104364
Received 23 August 2019; Received in revised form 23 January 2020; Accepted 3 March 2020
Available online 10 March 2020
0967-0661/© 2020 Elsevier Ltd. All rights reserved.
D. Shen, C.-C. Lim and P. Shi Control Engineering Practice 98 (2020) 104364

the future power demand is determined by the driver’s driving

behaviour. From an energy point of view, it is challenging to
predict human’s behaviour even the external driving conditions
are given.
• How to further mitigate the computational burden in the nonlin-
ear optimisation?
• How to blend theoretical development and practical issues to
guarantee stability, robustness and optimality of energy manage-
ment design in FCVs?

Motivated by these practical challenges, we propose a novel robust

fuzzy MPC based approach in this paper to design an energy man-
agement system in FCVs. Unlike previous fuzzy logic based control
strategies (Schouten, Salman, & Kheir, 2003; Sorrentino, Rizzo, & Arsie,
2011), Takagi–Sugeno (T–S) fuzzy approach (Tanaka & Wang, 2004) is
used as a systematic modelling methodology to approximate the nonlin-
ear power system in FCVs. The approach enables the use of most linear
control design tools to address stability analysis, systematic design
Fig. 1. Fuel cell vehicle with parallel hybrid system configuration.
and performance analysis. In addition, the T–S fuzzy approach renders
itself naturally to Lyapunov based system analysis with linear matrix
inequalities (LMIs) (Boyd, El Ghaoui, Feron, & Balakrishnan, 1994;
Lian, He, Zhang, Shi and Wu, 2019; Lian, He, Zhang and Wu, 2019),
where LMIs are used to express constraints in convex optimisation, and
the optimisation problem can be solved in polynomial time.
Furthermore, the robust MPC (RMPC) technique is employed in this
study for the power control unit design. Compared with the stochas-
tic MPC based approaches from the preceding review, advantageous
features of RMPC are twofold: the control technique has capable of
addressing uncertainties and disturbances in the application, and the
optimisation performance of the approach is decoupled from the pre-
dictive model as the optimisation is to minimise the upper bound of the
performance index instead of minimising the objective function over Fig. 2. Power flow in a fuel cell vehicle with parallel energy configuration.

the prediction horizon. In the literature, there have been many research
findings on RMPC and fuzzy systems. A review of RMPC approaches is
given in Bemporad and Morari (1999) and Kothare, Balakrishnan, and 𝑃𝑓 𝑐𝑠 (𝑘) = 𝑖𝑓 𝑐𝑠 (𝑘) ⋅ 𝑉𝑓 𝑐𝑠 (𝑘) (2)
Morari (1996). The features of recent research on robust and stochastic 𝑃𝑏𝑜𝑜𝑠𝑡 (𝑘) = 𝑖𝑑𝑐 (𝑘) ⋅ 𝑉𝑏𝑢𝑠 (𝑘) (3)
MPC are discussed in Mayne (2016). In Lu, Shi, Lam, and Zhao (2015),
Xia, Yang, Shi, and Fu (2010) and Yang, Feng, and Zhang (2014), MPC 𝑃𝑏𝑎𝑡𝑡 (𝑘) = 𝑖𝑏 (𝑘) ⋅ 𝑉𝑏𝑢𝑠 (𝑘) (4)
based approaches with T–S fuzzy modelling framework are discussed. 𝜂𝑏𝑜𝑜𝑠𝑡 (𝑘) = 𝑃𝑏𝑜𝑜𝑠𝑡 (𝑘)∕𝑃𝑓 𝑐𝑠 (𝑘) (5)
The main contribution of this paper is threefold: (i) A robust fuzzy
model predictive based control scheme with the T–S fuzzy modelling 𝜂𝑓 𝑐𝑠 (𝑘) = 𝑃𝑓 𝑐𝑠 (𝑘)∕𝑃ℎ2 (𝑘) (6)
framework that optimises the power distribution in FCVs while main- where 𝑃𝑚𝑜𝑡𝑜𝑟 is the power demand from the motor, 𝑃𝑏𝑜𝑜𝑠𝑡 is the output
taining battery charge-sustaining in the presence of disturbance; (ii) power of the boost converter, 𝑃ℎ2 is the output power of the hydrogen
A theoretical analysis for stability, robustness and performance that is tank, 𝑃𝑏𝑎𝑡𝑡 is the power delivered from the battery pack, 𝑃𝑓 𝑐𝑠 is the
applicable to the energy management system design in FCVs; (iii) An
power delivered from the fuel cell system, 𝑖𝑓 𝑐𝑠 is the output current of
improved robust fuzzy model predictive control by incorporating traffic
the fuel cell stack, 𝑉𝑓 𝑐𝑠 is the terminal voltage of the fuel cell stack,
condition for further improving the system performance in terms of fuel
𝑖𝑑𝑐 is the output current of the boost converter, 𝑖𝑏 is the output current
consumption.
of the battery pack, 𝑉𝑏𝑢𝑠 is the terminal voltage of the battery pack,
The paper is organised as follows. In Section 2, the energy man-
𝜂𝑏𝑜𝑜𝑠𝑡 is the boost converter efficiency, and 𝜂𝑓 𝑐𝑠 is the fuel cell system
agement problem in FCVs is formulated as an optimisation problem
efficiency.
under a fuzzy modelling framework. In Section 3, the robust fuzzy
Hydrogen consumption increases with current and is dependent on
model predictive control based control scheme is presented to design
the number of fuel cells in the stack, which can be calculated by
the power control unit for FCVs. We then extend the control scheme to
incorporate traffic condition for further improving the fuel consump- 𝑚ℎ2 (𝑖𝑓 𝑐 ) = 𝑐𝑜𝑓 𝑐 ⋅ 𝑖𝑓 𝑐 ⋅ 𝑁𝑓 𝑐 (7)
tion. Simulation result is given in Section 4 to illustrate the effectiveness
of the control schemes. Section 5 concludes this paper and discusses where 𝑚ℎ2 is the hydrogen consumption in standard litre per minute,
some directions for future research. 𝑐𝑜𝑓 𝑐 is the corresponding constant coefficient and 𝑁𝑓 𝑐 is the number
of fuel cells.
2. Problem formulation Now, we are ready to present the system dynamic model in this
study.
2.1. System dynamics of the power system in FCVs
2.1.1. Battery model
Fig. 1 shows the typical parallel structure of the power plant in the Fig. 3(a) shows the battery model in this study due to the simplicity
current FCV market (Nonobe, 2017) and Fig. 2 presents the power flow and the focus of energy-efficient driving. 𝑉𝑏𝑢𝑠 is, thus, defined as
in the system. At sampling time 𝑘, we have
𝑖𝑏 (𝑘)
𝑃𝑚𝑜𝑡𝑜𝑟 (𝑘) = 𝑃𝑏𝑎𝑡𝑡 (𝑘) + 𝑃𝑏𝑜𝑜𝑠𝑡 (𝑘) (1) 𝑉𝑏𝑢𝑠 (𝑘) = 𝑛𝑠𝑏 ⋅ (𝐸𝑏𝑎𝑡𝑡 (𝑘) − ⋅ 𝑅𝑏 ) (8)
𝑛𝑝𝑏

2
D. Shen, C.-C. Lim and P. Shi Control Engineering Practice 98 (2020) 104364

Fig. 3. (a) Equivalent circuit model for a battery cell; (b) Equivalent circuit model for
a fuel cell.

where 𝑛𝑠𝑏 is the number of cells in serial and 𝑛𝑝𝑏 is the number of cells
in parallel. Fig. 4. Efficiency map of a fuel cell boost converter.
The battery state of charge (SoC) is estimated by the Coulomb
counting method
𝜂𝑖 ⋅ 𝛥𝑡 2.2. T-S fuzzy modelling framework of the FCV
𝑆𝑏𝑎𝑡𝑡 (𝑘 + 1) = 𝑆𝑏𝑎𝑡𝑡 (𝑘) − 𝑖 (𝑘) (9)
𝐶𝑛 ⋅ 𝑛𝑝𝑏 𝑏
To build a fuzzy modelling framework, we reformulate the energy
where 𝜂𝑖 is the Faraday efficiency of the cell, 𝑆𝑏𝑎𝑡𝑡 is the cell SoC, 𝛥𝑡
management problem in FCVs to minimise the output power from the
is the sampling time period, 𝐶𝑛 is the cell capacity, and 𝑖𝑏 (𝑘) is the
instantaneous current delivered from the battery. hydrogen tank during a driving mission. Consequently, the performance
Furthermore, the battery open-circuit voltage 𝐸𝑏𝑎𝑡𝑡 is assumed to be index 𝐽 (𝑘) is rewritten by
a linear dependence of the SoC with parameters 𝑝𝑎 and 𝑝𝑏 when the 𝑇𝑛 −1
∑
𝑇
battery SoC is between 20% and 80% 𝐽 (𝑘) = 𝑃ℎ2 (𝑘 + 𝑖|𝑘)𝐿ℎ2 𝑃ℎ2 (𝑘 + 𝑖|𝑘)
𝑖=0
𝐸𝑏𝑎𝑡𝑡 (𝑆𝑏𝑎𝑡𝑡 ) = 𝑝𝑎 ⋅ 𝑆𝑏𝑎𝑡𝑡 + 𝑝𝑏 , for 𝑆𝑏𝑎𝑡𝑡 ∈ [0.2, 0.8] (10)
+ 𝜈𝑡 (𝑥(𝑘 + 𝑇𝑛 |𝑘)) (16)
2.1.2. Fuel cell model 𝜈𝑡 (𝑥(𝑘 + 𝑇𝑛 |𝑘)) = 𝑥(𝑘 + 𝑇𝑛 |𝑘) 𝑃 𝑥(𝑘 + 𝑇𝑛 |𝑘)
𝑇
(17)
The fuel cell model is shown in Fig. 3(b) where we assume that
the fuel cell stack operating under normal temperature and pressure. where 𝐿ℎ2 is the weighting scalar, 𝑃 is the weighting matrices, and 𝜈𝑡
In the model, 𝑉𝑓 𝑐 is the terminal voltage of the fuel cell, 𝐸𝑜𝑐 models is the terminal cost on states.
the open-circuit voltage of the fuel cell, 𝑅𝑜ℎ𝑚𝑖𝑐 models the ohmic losses, We define the states, control input, and disturbance acting on the
and 𝑅𝑎𝑐𝑡 models the activation losses. The voltage drop 𝛥𝑉𝑎𝑐𝑡 caused by system as follows.
activation overvoltage is given by
𝑥1 (𝑘) = 𝑆𝑏𝑎𝑡𝑡 (𝑘) − 𝑆𝑏𝑎𝑡𝑡⋅𝑟𝑒𝑓
𝑖𝑓 𝑐
𝛥𝑉𝑎𝑐𝑡 = 𝐴𝑓 𝑐 ⋅ ln( ) (11) 𝑥2 (𝑘) = 𝑉𝑏𝑢𝑠 (𝑘) − 𝑉𝑏𝑢𝑠⋅𝑟𝑒𝑓
𝑖0
𝑢(𝑘) = 𝑃𝑓 𝑐𝑠 (𝑘)
where 𝑖0 is the exchange current density on a proton exchange mem-
brane, 𝐴𝑓 𝑐 is the Tafel slope, and 𝑖𝑓 𝑐 is the current delivered from the 𝑤(𝑘) = 𝑃𝑚𝑜𝑡𝑜𝑟 (𝑘)
fuel cell.
where 𝑆𝑏𝑎𝑡𝑡⋅𝑟𝑒𝑓 and 𝑉𝑏𝑢𝑠⋅𝑟𝑒𝑓 are the reference values of the cell SoC and
To summarise, the voltage response of the fuel cell stack is described
DC-BUS voltage, respectively.
by
Substituting (6) to (16), we have
𝑖𝑓 𝑐 (𝑘)
𝑉𝑓 𝑐𝑠 (𝑘) = 𝑁𝑓 𝑐 ⋅ (𝐸𝑜𝑐 − 𝐴𝑓 𝑐 ⋅ ln( ) − 𝑖𝑓 𝑐 (𝑘) ⋅ 𝑅𝑜ℎ𝑚𝑖𝑐 ) (12) 𝑇𝑛 −1
∑
𝑖0 𝐽 (𝑘) = 𝑢𝑇 (𝑘 + 𝑖|𝑘)𝐿𝑢 𝑢(𝑘 + 𝑖|𝑘)
𝑖=0

+ 𝜈𝑡 (𝑥(𝑘 + 𝑇𝑛 |𝑘)) (18)

2.1.3. Fuel cell boost converter model
Since we focus on energy-efficient driving, the fuel cell boost con- where 𝐿𝑢 = (1∕𝜂𝑓 𝑐𝑠 )2 ⋅ 𝐿ℎ2 .
verter is modelled by its efficiency map. Based on (Hegazy, Van Mierlo, The objective 𝐽 (𝑘) consists of finding the control 𝑢(𝑘) that leads to
& Lataire, 2012; Pahlevaninezhad, Drobnik, Jain, & Bakhshai, 2011), a the minimisation of the fuel consumed over the driving mission while
typical efficiency map of a fuel cell boost converter can be described maintaining the battery SoC within certain admissible range at the
by Fig. 4. The boost converter efficiency is given by
terminal point 𝑇𝑛 .
𝜂𝑏𝑜𝑜𝑠𝑡 (𝑘) = 𝑓 (𝑃𝑏𝑜𝑜𝑠𝑡 (𝑘)) (13)
Remark 1. From an energy point of view, the power demand is
where 𝑓 (𝑃𝑏𝑜𝑜𝑠𝑡 ) is the relationship function between the load and the
determined by the driver’s driving behaviour. Poor control performance
boost converter efficiency.
may occur if the future power demand is not well predicted. However,
The optimality of the power distribution in FCVs is defined by min-
it is challenging to predict human behaviour even if the external driving
imising hydrogen consumption, which is represented by the following
route is given. A compromise is to use a fixed reference value for the
performance index 𝐽 (𝑘)
final states to deal with all driving scenarios. It will lead to sub-optimal
𝑇𝑛
∑ results, but the setting increases the controller’s applicability to handle
𝐽 (𝑘) = 𝑚ℎ2 (𝑖𝑓 𝑐 (𝑘)) (14)
all driving conditions found in the real application.
𝑘=0
where 𝑇𝑛 is the travel time. Combining (1)–(13), we have the following system dynamic model
The stability of the power plant is defined by maintaining the
𝜂𝑖 ⋅ 𝛥𝑡 𝑤(𝑘) − 𝜂𝑏𝑜𝑜𝑠𝑡 𝑢(𝑘)
battery SoC at a certain level over the trip 𝑥1 (𝑘 + 1) = 𝑥1 (𝑘) − ⋅ (19)
𝐶𝑛 ⋅ 𝑛𝑝𝑏 𝑥2 (𝑘) + 𝑉𝑏𝑢𝑠⋅𝑟𝑒𝑓
𝑆𝑏𝑎𝑡𝑡 (𝑘 + 𝑇𝑛 |𝑘) ∈ 𝛺𝑏 (15) 𝑅 𝑤(𝑘) − 𝜂𝑏𝑜𝑜𝑠𝑡 𝑢(𝑘)
𝑥2 (𝑘 + 1) = 𝑛𝑠𝑏 (𝑝𝑎 𝑥1 (𝑘) − 𝑏 ⋅ ) (20)
where 𝛺𝑏 is the terminal constraints set of the battery SoC. 𝑛𝑝𝑏 𝑥2 (𝑘) + 𝑉𝑏𝑢𝑠⋅𝑟𝑒𝑓

3
D. Shen, C.-C. Lim and P. Shi Control Engineering Practice 98 (2020) 104364

IF 𝑧(𝑘) is ‘‘Small’’
𝑥(𝑘 + 1) = 𝐴𝑥(𝑘) + 𝐵1 𝑢(𝑘) + 𝐸1 𝑤(𝑘) (24)
RULE B:
IF 𝑧(𝑘) is ‘‘Large’’
𝑥(𝑘 + 1) = 𝐴𝑥(𝑘) + 𝐵2 𝑢(𝑘) + 𝐸2 𝑤(𝑘) (25)

where
[ ]
1 0
𝐴=
𝑛𝑠𝑏 𝑝𝑎 0

⎡ ⎤ ⎡ ⎤
𝜂𝑖 𝛥𝑡𝜂𝑏𝑜𝑜𝑠𝑡 𝜂𝑖 𝛥𝑡𝜂𝑏𝑜𝑜𝑠𝑡
⎢ 𝐶𝑛 𝑛𝑝𝑏 (𝑥2⋅𝑚𝑎𝑥 +𝑉𝑏𝑢𝑠⋅𝑟𝑒𝑓 ) ⎥ ⎢ ⎥
⎥ 𝐵2 = ⎢⎢ ⎥
𝐶 𝑛 𝑛 𝑝𝑏 (𝑥2⋅𝑚𝑖𝑛 +𝑉 𝑏𝑢𝑠⋅𝑟𝑒𝑓 )
𝐵1 = ⎢
Fig. 5. Local sector nonlinearity. ⎢ 𝑅𝑏 𝜂𝑏𝑜𝑜𝑠𝑡
⎥ 𝑅𝑏 𝜂𝑏𝑜𝑜𝑠𝑡
⎥
⎣ 𝑝𝑏 2⋅𝑚𝑎𝑥 𝑏𝑢𝑠⋅𝑟𝑒𝑓 ⎦
𝑛 (𝑥 +𝑉 ) ⎢ ⎥
⎣ 𝑝𝑏 2⋅𝑚𝑖𝑛 𝑏𝑢𝑠⋅𝑟𝑒𝑓 ⎦
𝑛 (𝑥 +𝑉 )

⎡ ⎤ ⎡ ⎤
−𝜂𝑖 𝛥𝑡 −𝜂𝑖 𝛥𝑡
In this study, the fuzzy model is constructed by the approach of ⎢ 𝐶𝑛 𝑛𝑝𝑏 (𝑥2⋅𝑚𝑎𝑥 +𝑉𝑏𝑢𝑠⋅𝑟𝑒𝑓 ) ⎥ ⎢ ⎥
𝐸1 = ⎢ ⎢
⎥ 𝐸2 = ⎢
𝐶𝑛 𝑛𝑝𝑏 (𝑥2⋅𝑚𝑖𝑛 +𝑉𝑏𝑢𝑠⋅𝑟𝑒𝑓 ) ⎥
the local sector nonlinearity in fuzzy partition spaces (Tanaka & Wang,
⎢ 𝑅𝑏
⎥ 𝑅𝑏
⎥
2004). Shown in Fig. 5 is a local sector nonlinearity. To find a local ⎣ 𝑛𝑝𝑏 (𝑥2⋅𝑚𝑎𝑥 +𝑉𝑏𝑢𝑠⋅𝑟𝑒𝑓 ) ⎦ ⎢ ⎥
⎣ 𝑛𝑝𝑏 (𝑥2⋅𝑚𝑖𝑛 +𝑉𝑏𝑢𝑠⋅𝑟𝑒𝑓 ) ⎦
sector for a nonlinear system
[ ] From Fig. 4, we observe that the boost converter operates efficiently
𝑥(𝑘 + 1) = 𝑓 (𝑥(𝑘)) ∈ 𝑎1 𝑎2 𝑥(𝑘) when the output power of the boost converter is higher than a certain
[ ] level 𝜀𝑝1 (around 10% of its maximum power). In this situation, it is
where 𝑥(𝑘) ∈ 𝑑1 𝑑2 , we approximate the nonlinear function by the
local sector using fuzzy membership functions. reasonable to assume the boost converter efficiency 𝜂𝑏𝑜𝑜𝑠𝑡 a constant.
The steps to construct the fuzzy model is as follows. In addition, a fuel cell system (Gemmen & Johnson, 2006; Staunton,
Eqs. (19) and (20) can be written as Ayers, Marlino, Chiasson, & Burress, 2006; Wipke et al., 2012) shows
the similar efficiency characteristic, that is, a fuel cell system operates
[ ] ⎡ 𝜂𝑖 𝛥𝑡𝜂𝑏𝑜𝑜𝑠𝑡 ⎤
1 0 ⎢ 𝐶𝑛 𝑛𝑝𝑏 (𝑥2 (𝑘)+𝑉𝑏𝑢𝑠⋅𝑟𝑒𝑓 ) ⎥ efficiently and 𝜂𝑓 𝑐𝑠 can be assumed as a constant when the power
𝑥(𝑘 + 1) = 𝑥(𝑘) + ⎢ ⎥ 𝑢(𝑘) delivered from the fuel cell system is higher than a certain value 𝜀𝑝2 .
𝑛𝑠𝑏 𝑝𝑎 0 ⎢ 𝑅𝑏 𝜂𝑏𝑜𝑜𝑠𝑡
⎥
⎣ 𝑛𝑝𝑏 (𝑥2 (𝑘)+𝑉𝑏𝑢𝑠⋅𝑟𝑒𝑓 ) ⎦ Therefore, we introduce a bounded uncertainty 𝛥𝑢 ≤ max(𝜀𝑝1 , 𝜀𝑝2 )
⎡ 𝜂𝑖 𝛥𝑡 ⎤ on the input part to guarantee that the fuel cell system and the boost
⎢ 𝐶 𝑛 (𝑥 (𝑘)+𝑉𝑏𝑢𝑠⋅𝑟𝑒𝑓 ) ⎥ converter operate efficiently. By considering the uncertainty 𝛥𝑢 in the
− ⎢ 𝑛 𝑝𝑏 2 ⎥ 𝑤(𝑘) (21)
⎢ 𝑅𝑏
⎥ system, we can enforce 𝑃𝑏𝑜𝑜𝑠𝑡 to 𝜀𝑝1 when 𝑃𝑏𝑜𝑜𝑠𝑡 < 𝜀𝑝1 and 𝑃𝑓 𝑐𝑠 to 𝜀𝑝2
⎣ 𝑛𝑝𝑏 (𝑥2 (𝑘)+𝑉𝑏𝑢𝑠⋅𝑟𝑒𝑓 ) ⎦ when 𝑃𝑓 𝑐𝑠 < 𝜀𝑝2 . We rewrite the T–S fuzzy model in (24)–(25) by
where 𝑥(𝑘) = [𝑥1 (𝑘), 𝑥2 (𝑘)]𝑇 .
𝑥(𝑘 + 1)
Since 𝑛𝑠𝑏 , 𝑝𝑎 , 𝜂𝑖 , 𝛥𝑡, 𝐶𝑛 , 𝑛𝑝𝑏 , 𝑅𝑏 are known constants, the only nonlin-
ear term in the system is 1∕(𝑥2 (𝑘) + 𝑉𝑏𝑢𝑠⋅𝑟𝑒𝑓 ) if 𝜂𝑏𝑜𝑜𝑠𝑡 can be assumed as ∑
𝑟
= ℎ𝑖 (𝑧)(𝐴𝑥(𝑘) + 𝐵𝑖 (𝑢(𝑘) + 𝛥𝑢) + 𝐸𝑖 𝑤(𝑘)) (26)
a constant. For the nonlinear term, define 𝑧(𝑘) = 1∕(𝑥2 (𝑘) + 𝑉𝑏𝑢𝑠⋅𝑟𝑒𝑓 ). 𝑖=1
Then, we have
where 𝑟 is the number of fuzzy rules (𝑟 = 2 in our case) and
[ ] ⎡ 𝜂𝑖 𝛥𝑡𝜂𝑏𝑜𝑜𝑠𝑡 ⎤
1 0 ⎢ 𝐶𝑛 𝑛𝑝𝑏 𝑧(𝑘)⎥ ℎ1 (𝑧) = 𝑀1 (𝑧(𝑘)) ℎ2 (𝑧) = 𝑀2 (𝑧(𝑘))
𝑥(𝑘 + 1) = 𝑥(𝑘) + ⎢ ⎥ 𝑢(𝑘)
𝑛𝑠𝑏 𝑝𝑎 0 𝑅 𝜂
⎢ 𝑏𝑛 𝑏𝑜𝑜𝑠𝑡 𝑧(𝑘) ⎥
⎣ 𝑝𝑏 ⎦ In addition, the constraint on control input is considered
⎡ 𝜂𝑖 𝛥𝑡 𝑧(𝑘)⎤ 𝑢(𝑘 + 𝑖|𝑘) ∈ U, 𝑖 = 0, 1, … , 𝑇𝑛 − 1 (27)
−⎢ 𝑅 ⎥ 𝑤(𝑘)
𝐶𝑛 𝑛𝑝𝑏
(22)
⎢ 𝑏 𝑧(𝑘) ⎥ where U is a constraint set of the power delivered from the fuel cell
⎣ 𝑛𝑝𝑏 ⎦
stack.
Next, we calculate the maximum and minimum values of 𝑧(𝑘) under
𝑥2 (𝑘) ∈ [𝑥2⋅𝑚𝑖𝑛 , 𝑥2⋅𝑚𝑎𝑥 ], that is,
Remark 2. To employ T–S fuzzy approach as the approximator for
𝑧𝑚𝑎𝑥 = 1∕(𝑥2⋅𝑚𝑖𝑛 + 𝑉𝑏𝑢𝑠⋅𝑟𝑒𝑓 ) the power system in FCVs, an intuitive solution is to take 𝑖𝑓 𝑐 as the
𝑧𝑚𝑖𝑛 = 1∕(𝑥2⋅𝑚𝑎𝑥 + 𝑉𝑏𝑢𝑠⋅𝑟𝑒𝑓 ) control input, 𝑆𝑏𝑎𝑡𝑡 as the state, and then linearise the control system
under a T–S fuzzy modelling framework. However, the solution suffers
For their maximum and minimum values, 𝑧(𝑘) is represented by from computational burden since the premise variables have to be
𝑧(𝑘) = 1∕(𝑥2 (𝑘) + 𝑉𝑏𝑢𝑠⋅𝑟𝑒𝑓 ) the functions of the input variables which introduces a complicated
defuzzification process of fuzzy controllers (Tanaka & Wang, 2004).
= 𝑀1 (𝑧(𝑘)) ⋅ 𝑧𝑚𝑖𝑛 + 𝑀2 (𝑧(𝑘)) ⋅ 𝑧𝑚𝑎𝑥 (23)
Therefore, we adopt the alternative approach (24)–(26) to build the
where 𝑀1 (𝑧(𝑘)) + 𝑀2 (𝑧(𝑘)) = 1. fuzzy modelling framework.
From (23), the membership functions can be defined as
Remark 3. Battery SoC is normally estimated and monitored by a
𝑀1 (𝑧(𝑘)) = (𝑧(𝑘) − 𝑧𝑚𝑎𝑥 )∕(𝑧𝑚𝑖𝑛 − 𝑧𝑚𝑎𝑥 )
battery management system in an FCV. Its estimated state is sent to the
𝑀2 (𝑧(𝑘)) = (𝑧(𝑘) − 𝑧𝑚𝑖𝑛 )∕(𝑧𝑚𝑎𝑥 − 𝑧𝑚𝑖𝑛 ) power control unit via CAN bus. In this study, we follow this concept
We denote the membership functions ‘‘Large’’ and ‘‘Small’’. Then, the and assume the battery SoC is observable.
nonlinear system (21) is represented by the following 2-rule fuzzy
In summary, the energy management problem in this study is to
model
minimise the cost function 𝐽 (𝑘) in (18), subject to the system dynamic
RULE A: model in (26) and constraints in (15) and (27).

4
D. Shen, C.-C. Lim and P. Shi Control Engineering Practice 98 (2020) 104364

3. Control scheme of robust fuzzy model predictive control Proof. See Appendix A □

In this section, we describe our proposed control scheme to address Remark 4. The feedback matrices 𝐾𝑖 computed from Theorem 1 are
the energy management problem posed in Section 2. A robust MPC capable of ensuring that the candidate invariant set 𝛶 is an RPI set for
based approach is proposed to minimise the upper bound of the cost the system in (26). According to Definition 1, the path constraints on
function 𝐽 (𝑘) in (18) with a state-feedback controller the states are imposed by the RPI set.
∑
𝑟
𝑢(𝑘) = ℎ𝑖 (𝑧(𝑘))𝐾𝑖 𝑥(𝑘) (28) Moreover, we impose the following conditions on the terminal cost
𝑖=1 𝜈𝑡 to achieve stability:

3.1. Recursive feasibility 𝜈𝑡 (𝑥(𝑘 + 1 + 𝑖|𝑘)) − 𝜈𝑡 (𝑥(𝑘 + 𝑖|𝑘))

≤ 𝛼 2 (𝑤(𝑘 + 𝑖|𝑘)𝑇 𝑤(𝑘 + 𝑖|𝑘) + 𝛥𝑢(𝑘 + 𝑖|𝑘)𝑇 𝛥𝑢(𝑘 + 𝑖|𝑘))
The first issue that we address is the recursive feasibility of the
control scheme in the presence of disturbance and uncertainties. In − 𝑢𝑇 (𝑘 + 𝑖|𝑘)𝐿𝑢 𝑢(𝑘 + 𝑖|𝑘) (37)
this study, we adopt the concept of robust positively invariant (RPI)
where 𝛼 is a positive scalar. The stability and optimality analysis of the
set (Blanchini, 1999; Mayne, Raković, Findeisen, & Allgöwer, 2006)
control system will be given in the following sections.
to design the control law, which provides sufficient conditions to
guarantee the recursive feasibility.
3.2. Optimality and constraints
Definition 1 (RPI for Discrete-time Systems Blanchini, 1999). For the
system 𝑥(𝑘 + 1) = 𝑓 (𝑥(𝑘), 𝑤(𝑘)), a set 𝛶 ⊂ R𝑛 is robust positively Summing (37) from 𝑖 = 0 to 𝑖 = 𝑇𝑛 − 1,
invariant if for all 𝑥(0) ∈ 𝛶 and all 𝑤(𝑘) ∈ W, the solution is such 𝑇𝑛 −1
∑
that 𝑥(𝑘) ∈ 𝛶 for 𝑘 > 0. 𝐽 (𝑘) = 𝑢𝑇 (𝑘 + 𝑖|𝑘)𝐿𝑢 𝑢(𝑘 + 𝑖|𝑘) + 𝜈𝑡 (𝑥(𝑘 + 𝑇𝑛 |𝑘))
𝑖=0
For the fuzzy system in (26), we define the candidate invariant set 𝑇𝑛 −1
𝛶 as follows. ∑
≤ 𝜈𝑡 (𝑥(𝑘|𝑘)) + 𝛼 2 (𝑤(𝑘 + 𝑖|𝑘)𝑇 𝑤(𝑘 + 𝑖|𝑘)
𝛶 = {𝑥 ∈ R𝑛 ∶ 𝑥𝑇 𝑃 𝑥 ≤ 𝜉} (29) 𝑖=0

+ 𝛥𝑢(𝑘 + 𝑖|𝑘)𝑇 𝛥𝑢(𝑘 + 𝑖|𝑘)) (38)

Derived from the results in Alessandri, Baglietto, and Battistelli (2004)
and Yang et al. (2014), the set in (29) is an RPI set if From (32) and (33), we have
1 𝑤(𝑘)𝑇 𝑤(𝑘) + 𝛥𝑢(𝑘)𝑇 𝛥𝑢(𝑘) 𝑇𝑛 −1
∑
𝑥(𝑘)𝑇 𝑃 𝑥(𝑘) ≥ (30)
𝜉 𝛾 2 + 𝛿2 (𝑤(𝑘 + 𝑖|𝑘)𝑇 𝑤(𝑘 + 𝑖|𝑘) + 𝛥𝑢(𝑘 + 𝑖|𝑘)𝑇 𝛥𝑢(𝑘 + 𝑖|𝑘))
𝑖=0
implies
≤ 𝑇𝑛 (𝛾 2 + 𝛿 2 ) (39)
1 1
𝑥(𝑘 + 1)𝑇 𝑃 𝑥(𝑘 + 1) ≤ 𝑥(𝑘)𝑇 𝑃 𝑥(𝑘) (31)
𝜉 𝜉 Thus, we obtain
where 𝛾 and 𝛿 are known constants that, for all 𝑘 ≥ 0,
𝐽 (𝑘) ≤ 𝜈𝑡 (𝑥(𝑘|𝑘)) + 𝛼 2 𝑇𝑛 (𝛾 2 + 𝛿 2 ) (40)
𝑤(𝑘)𝑇 𝑤(𝑘) ≤ 𝛾 2 (32)
Consider
𝛥𝑢(𝑘)𝑇 𝛥𝑢(𝑘) ≤ 𝛿 2 (33)
𝜈𝑡 (𝑥(𝑘)) = 𝑥(𝑘)𝑇 𝑃 𝑥(𝑘) ≤ 𝜉 (41)
Using the S-procedure technique (Yakubovich, 1992), (31) holds if
there exists 𝜆 ∈ (0, 1] such that the upper bound of the cost function 𝐽 (𝑘) is minimised by solving
1 min 𝜉, subject to 𝜈𝑡 (𝑥(𝑘)) ≤ 𝜉
(𝑥(𝑘 + 1)𝑇 𝑃 𝑥(𝑘 + 1) − 𝑥(𝑘)𝑇 𝑃 𝑥(𝑘)) (42)
𝜉
𝑤(𝑘)𝑇 𝑤(𝑘) + 𝛥𝑢(𝑘)𝑇 𝛥𝑢(𝑘) 1 and the input constrain is defined as follows.
− 𝜆( − 𝑥(𝑘)𝑇 𝑃 𝑥(𝑘))
𝛾 2 + 𝛿2 𝜉
𝑢𝑇 (𝑘)𝑢(𝑘) ≤ 𝜇 2 (43)
≤0 (34)
The following theorem gives us LMIs conditions for existence of the
The following theorem shows that the conditions in (34) can be cast in appropriate 𝑃 satisfying (37) and input constraint (43), and the corre-
the form of LMIs to guarantee 𝛶 is a robust positively invariant set for sponding state feedback matrices 𝐾𝑖 .
the system in (26).
Theorem 2. The state feedback matrices 𝐾𝑖 in (28) that minimise the
Theorem 1. The state feedback matrices 𝐾𝑖 in (28) to ensure 𝛶 being an upper bound of the performance function 𝐽 (𝑘|𝑘) at time step 𝑘 and subject
RPI for the system in (26) are given by to the constraint 𝑢𝑇 (𝑘)𝑢(𝑘) ≤ 𝜇 2 at all time 𝑘 ≥ 0 are given by (35), where
𝐾𝑖 = 𝑀𝑖 𝑄−1 (35) 𝑄 > 0 and 𝑀𝑖 are obtained by solving the following optimisation problem

where 𝜆 ∈ (0, 1], 𝑄 > 0 and 𝑀𝑖 are obtained from the following matrix min 𝜉
inequality subject to
⎡ ⎤ ⎡
⎢ (1 − 𝜆)𝑄 ⋆ ⋆ ⋆⎥
⎢ 𝑄 ⋆ ⋆ ⋆ ⋆ ⎤⎥
⎢ ⋆ ⎥⎥ ⎢
⎢ 0 𝜆∕(𝛾 2 + 𝛿2 ) ⋆
⎢ 0 𝜉𝛼 2 ⋆ ⋆ ⋆ ⎥⎥
≥0
⎢ ⋆ ⎥⎥ ⎢
⎢ 0 0 𝜆∕(𝛾 2 + 𝛿2 )
⎢ 0 0 𝜉𝛼 2 ⋆ ⋆ ⎥⎥ ≥ 0
⎢𝐴𝑄 + 𝐵𝑖 𝑀𝑗 +𝐵𝑗 𝑀𝑖 𝑄⎥⎦ ⎢𝐴𝑄 + 𝐵𝑖 𝑀𝑗 +𝐵𝑗 𝑀𝑖 ⋆ ⎥⎥
𝐵𝑖 +𝐵𝑗 𝐸𝑖 +𝐸𝑗 𝐵𝑖 +𝐵𝑗 𝐸𝑖 +𝐸𝑗
⎣ 2 2 2 ⎢ 2
𝜉 2
𝜉 2
𝑄
⎢ 𝑀𝑖 +𝑀𝑗 ⎥
for 𝑖 ≤ 𝑗 subject to ℎ𝑖 ∩ ℎ𝑗 ≠ ∅ (36) ⎣ 2
𝐿 𝑢 0 0 0 𝜉𝐿𝑢 ⎦
where ⋆ denotes terms readily inferred from symmetry. for 𝑖 ≤ 𝑗 subject to ℎ𝑖 ∩ ℎ𝑗 ≠ ∅ (44)

5
D. Shen, C.-C. Lim and P. Shi Control Engineering Practice 98 (2020) 104364

Lemma 1 (Input-to-state Practical Stable Limón et al., 2006). If there

exists an input-to-state practical stability Lyapunov function 𝑉 (𝑘) that sat-
isfies (49) and (50) for system (48), then the system is called input-to-state
practical stable.

Lemma 2. If there exists a feasible solution of the optimisation problem

in (47) at sampling time 𝑡 = 𝑘, then the solution is also feasible for all time
𝑡 > 𝑘.

Proof. Since (46) is the only matrix inequality in the optimisation

problem that depends on the state, we only need to prove that (46)
is feasible for all time 𝑡 > 𝑘.
The feasibility of (46) implies 𝑥(𝑘|𝑘)𝑇 𝑃 𝑥(𝑘|𝑘) ≤ 𝜉. Thus, for any time
𝑡 > 𝑘, we must prove
Fig. 6. Block diagram of the robust fuzzy model predictive control based control
scheme. 𝑥(𝑘 + 𝑖|𝑘)𝑇 𝑃 𝑥(𝑘 + 𝑖|𝑘) ≤ 𝜉 for all 𝑖 > 0 (51)

As shown in (31), the matrix inequality in (36) implies

[ ]
𝑄 𝑀𝑖𝑇
2 ≥0 (45) 𝑥(𝑘 + 𝑖|𝑘)𝑇 𝑃 𝑥(𝑘 + 𝑖|𝑘) ≤ 𝑥(𝑘|𝑘)𝑇 𝑃 𝑥(𝑘|𝑘) for all 𝑖 > 0
𝑀𝑖 𝜇 𝐼
and Thus, (51) is guaranteed for any time 𝑡 > 𝑘 and the proof
[ ] is completed. □
1 𝑥𝑇 (𝑘)
≥0 (46)
𝑥(𝑘) 𝑄
Theorem 3. Provided that a feasible solution exists for the optimisation
problem in (47) at time 𝑘 = 0, then the system in (26) is input-to-state
Proof. See Appendix B □
practical stable.
Fig. 6 shows the block diagram of the RMPC based control scheme,
and the procedure for the control scheme is shown as follows. Proof. First, from Lemma 2, we can guarantee the recursive feasibility
of the control law. Then, to guarantee input-to-state practical stable
• Step 1: At time step 𝑘, control law (28) is yielded by the solution
of the following optimisation problem with the fuzzy membership of the system in (26), we only need to find an input-to-state practical
values ℎ𝑖 : stability Lyapunov function 𝑉 (𝑘) that satisfies (49) and (50).
Consider optimal control gain 𝐾 ∗ (𝑘) and 𝑃 ∗ (𝑘) at sampling time 𝑘,
min 𝜉 (47) we have
subject to (36), (44), (45), (46)
𝜈𝑡∗ = 𝑥𝑇 (𝑘)𝑃 ∗ (𝑘)𝑥(𝑘) (52)
• Step 2: At time step 𝑘 + 1, the new measurement 𝑥(𝑘 + 1) is taken, 2
𝜆𝑚𝑖𝑛 ‖𝑥‖ ≤ 𝜈𝑡∗ ≤ 𝜆𝑚𝑎𝑥 ‖𝑥‖2
(53)
and Step 1 is repeated.
where 𝜆𝑚𝑖𝑛 and 𝜆𝑚𝑎𝑥 represent minimal and maximal eigenvalues of
Remark 5. The range of the RPI set 𝛶 can be calculated off-line from 𝑃 ∗ (𝑘), respectively.
(47). Since the path constraints on the states are imposed by 𝛶 , if the Moreover, from (37), we obtain
maximum range of 𝛶 does not exceed the battery SoC limitation, then
𝜈𝑡∗ (𝑥(𝑘 + 1|𝑘)) − 𝜈𝑡∗ (𝑥(𝑘|𝑘))
the battery will not be fully depleted or overloaded over the trip.
≤ 𝛼 2 (𝑤(𝑘|𝑘)𝑇 𝑤(𝑘|𝑘) + 𝛥𝑢(𝑘|𝑘)𝑇 𝛥𝑢(𝑘|𝑘))

3.3. Stability analysis − 𝑥𝑇 (𝑘|𝑘)𝐾 ∗ (𝑘|𝑘)𝐿𝑢 𝐾 ∗ (𝑘|𝑘)𝑥(𝑘|𝑘) (54)

which satisfies (50).

In this section, we adopt the input-to-state stability framework
(Jiang & Wang, 2001; Limón, Alamo, Salas, & Camacho, 2006; Magni,
Raimondo, & Scattolini, 2006; Sontag & Wang, 1995) for analysing the
stability of the proposed RMPC based control scheme. First, we give the
following definitions:
Thus, 𝜈𝑡 is an input-to-state practical stability Lyapunov function of
the system, and the proof is completed. □
3.4. Controller design by incorporating traffic condition

Definition 2. If a function 𝜅 ∶ R ≥ 0 → R ≥ 0 is continuous, strictly
increasing, it is called a 𝒦 function; If a function 𝜅 ∶ R ≥ 0 → R ≥ 0 is
a 𝒦 function and 𝜅(𝑘) = ∞ as 𝑘 → ∞, it is called a 𝒦∞ function.
Definition 3. Consider the discrete-time system
𝑥(𝑘 + 1) = 𝑓 (𝑥(𝑘), 𝑢(𝑘), 𝑤(𝑘))
Provided that the disturbance is bounded by 𝑤 ∈ W. A function 𝑉 (∗) ∶
R𝑛 → R ≥ 0 is an input-to-state practical stability Lyapunov function if
it satisfies
𝜃1 (‖𝑥‖) ≤ 𝑉 (𝑘) ≤ 𝜃2 (‖𝑥‖)
𝑉 (𝑥(𝑘 + 1)) − 𝑉 (𝑥(𝑘)) ≤ −𝜃3 (‖𝑥‖) + 𝜙(‖𝑤‖)
where 𝜃1 , 𝜃2 , 𝜃3 are 𝒦∞ functions, 𝜙 is a 𝒦 function, and ‖𝑥‖ and ‖𝑤‖
are norms of given vectors 𝑥 and 𝑤, respectively.
The major disadvantage of using the above RMPC based control
scheme for energy management systems is its conservatism since the
control law intents to minimise the upper bound of the performance
index in (40) where 𝛾 defined in (32) should cover the maximum power
that the vehicle can provide.
In some driving scenarios, the upper bound of the power demand is
(48)
much smaller than the maximum power that the vehicle can provide.
For example, in the scenario when the vehicle is stuck in a traffic jam,
the upper bound of the power demand is passively limited by the poor
traffic condition.
(49) Motivated by the potential feasibility to relax the conditions of
the RMPC based control scheme, we incorporate traffic condition into
(50) the energy management controller design to further improve the fuel
consumption.
To process, we make the following assumption.

6
D. Shen, C.-C. Lim and P. Shi Control Engineering Practice 98 (2020) 104364

• Step 2: At time step 𝑘 + 1, new measurement 𝑥(𝑘 + 1) and

measurements 𝛾 ∗ (𝑘 + 1) are taken, and go to Step 1 again.

To prove recursive feasibility and stability of the new control scheme,

the following useful lemma is given.

Lemma 3. For the system in (26), if there exists a solution to the

optimisation problem in (47) at sampling time 𝑡 = 𝑘, the solution of the
optimisation problem in (57) is also feasible for all 𝑡 ≥ 𝑘.

Proof. Suppose that the solution of the optimisation problem in (47)

is feasible, condition (34) must be satisfied. Since 𝛾 is defined as the
upper bound of 𝑤(𝑘) for all 𝑘 ≥ 0, we must have

𝛾 ∗ (𝑘)2 + 𝛿 2 ≤ 𝛾 2 + 𝛿 2 (58)
Fig. 7. Block diagram of the robust fuzzy model predictive control with driving
scenario recognition. Substituting (58) into (34) yields
1
[𝑥(𝑘 + 1)𝑇 𝑃 𝑥(𝑘 + 1) − 𝑥(𝑘)𝑇 𝑃 𝑥(𝑘)]
𝜉
Assumption 1. There exists a power upper bound mapping system 𝑤(𝑘)𝑇 𝑤(𝑘) + 𝛥𝑢(𝑘)𝑇 𝛥𝑢(𝑘) 1
≤ 𝜆[ − 𝑥(𝑘)𝑇 𝑃 𝑥(𝑘)]
such that traffic conditions are capable of mapping to the different 𝛾 2 + 𝛿2 𝜉
upper bound of the vehicle power demand by several associative rules. 𝑤(𝑘)𝑇 𝑤(𝑘) + 𝛥𝑢(𝑘)𝑇 𝛥𝑢(𝑘) 1
≤ 𝜆[ − 𝑥(𝑘)𝑇 𝑃 𝑥(𝑘)] (59)
An example of the power upper bound mapping from a set of rules 𝛾 ∗ (𝑘)2 + 𝛿 2 𝜉
is as follows. As shown in Theorem 1, the condition in (59) can be cast in the form
of LMIs (56) to guarantee that 𝛶 in (29) is a robust positively invariant
• Rule1: If traffic condition is ‘excellent’, then 𝛾(𝑘) = 𝜃1
set for the system. The rest of proof is similar to that of Lemma 2 and
• Rule2: If traffic condition is ‘good’, then 𝛾(𝑘) = 𝜃2 thus is omitted. □
• Rule3: If traffic condition is ‘bad’, then 𝛾(𝑘) = 𝜃3
• Rule4: If traffic condition is ‘poor’, then 𝛾(𝑘) = 𝜃4 Theorem 4. The feasible control law obtained by solving the optimisation
Given that the mapping between the traffic conditions and the upper problem in (57) robustly stabilises the T–S fuzzy system in (26).
bound of the vehicle power demand can be analysed off-line, and that
live traffic information service is widely available in many countries, it Proof. Lemma 3 states the recursive feasibility of the optimisation
is reasonable and practical to make Assumption 1. problem (57). The rest of proof is similar to that of Theorem 3 and
Extended from Fig. 6, the block diagram of the new control scheme thus is omitted. □
is shown in Fig. 7. Since the upper bound 𝛾 is adjustable based on dif-
ferent traffic conditions, the performance index 𝐽 (𝑘) in (38) becomes: Remark 6. Since (59) implies that (56) relaxes the condition of
(36), the RMPC with varied 𝛾 shows less conservative LMI conditions
of the close-loop system than the RMPC based control scheme with
𝐽 (𝑘) ≤ 𝜈𝑡 (𝑥(𝑘|𝑘)) + 𝛼 2 (𝛾 ∗ (𝑘)2 + 𝛿 2 ) (55)
constant 𝛾. Therefore, by incorporating traffic condition, the RMPC
where 𝛾 ∗ (𝑘)
is the upper bound of the power demand at sampling step based energy management controller is capable of achieving better
𝑘 based on the traffic condition. control performance with regard to minimising the performance index
Intuitively, since 𝛾 ∗ (𝑘) ≤ 𝛾, the new control scheme is capable 𝐽 (𝑘) in (18).
of achieving better control performance in terms of the performance
expressed in (38) than the RMPC based control scheme with constant 4. Case study
𝛾, but how could we guarantee robust stability and feasibility in the
new control scheme? 4.1. Simulation environment
We first give the procedure for the design of the new control
scheme: To illustrate the effectiveness of the proposed control schemes,
Since the upper bounds 𝛾 and 𝛿 only appear in the matrix inequality MATLAB Powertrain BlocksetTM is used in the simulation studies. Since
in (36) of the optimisation problem in (47) in the RMPC based control the software has not supported LMI toolbox (Boyd et al., 1994) and pro-
scheme, we only need to rewrite the matrix inequality in (36) by vided fuel cell models, it is used mainly to simulate the power demand
⎡ from the driver and then develop a power plant model based on real-
(1 − 𝜆)𝑄 ⋆ ⋆ ⋆⎤
⎢ ⎥ world experimental data for verifying the effectiveness of our control
⎢ 0 𝜆
⋆ ⋆⎥ schemes. The data acquisition and processing time resolution (Wei,
⎢ 𝛾 ∗ (𝑘)2 +𝛿 2 ⎥≥0 (56)
⎢ 0 0 𝜆
⋆ ⎥⎥
Qiao, & Liu, 2018) is set to one second.
⎢ 𝛾 ∗ (𝑘)2 +𝛿 2 In order to build a high-fidelity model of the fuel cell vehicle system,
⎢ 𝐵𝑖 𝑀𝑗 +𝐵𝑗 𝑀𝑖 𝐵𝑖 +𝐵𝑗 𝐸𝑖 +𝐸𝑗 ⎥
⎣𝐴𝑄 + 2 2 2
𝑄⎦ the following experiments are conducted to build the battery model and
fuel cell model in the simulator:
Therefore, the procedure for the new control scheme is summarised as
follows. The battery in the vehicle simulator is modelled by a two RC-branch
equivalent circuit as shown in Fig. 8(b). A 20 Ah prototype lithium-ion
• Step 1: At sampling step 𝑘, control law (28) is obtained by solving Polymer battery, as shown in Fig. 8(a), is used to estimate and verify
the following optimisation problem with the fuzzy membership the parameters of the model. Discharge current pulses from 40 A down
values ℎ𝑖 : to 0 A are applied to the prototype, and the corresponding terminal
voltage response data is then collected.
min 𝜉 (57)
The fuel cell is modelled by the equivalent circuit model in
subject to (44), (45), (46), (56) Fig. 10(b), by assuming fuel cells operate at optimal condition. A

7
D. Shen, C.-C. Lim and P. Shi
Fig. 8. Prototype battery cell for experiment (top) and the two RC-branch equivalent
circuit model in the vehicle simulator (bottom).
Fig. 9. Parameters estimation and verification of the battery model.
Ballard FCvelocity-9SSL fuel cell stack, as shown in Fig. 10(a), is
used for model verification. Similar to the battery experiment, load
current from 0 A to 100 A are applied to the fuel cell stack, and the
corresponding terminal voltage response data is collected.
The results of simulated data from the models and measured data
from the prototypes are shown in Figs. 9 and 11, respectively. We
observe that the battery model and the fuel cell model are capable of
capturing the dynamic characteristics of two prototypes well.
The YALMIP toolbox (Lofberg, 2004) with SDPT3 solver (Tütüncü,
Toh, & Todd, 2003) is used to solve the optimisation problems. In the
power plant, The battery pack contains 78 cells (20 Ah) in series then
5 strings in parallel and the fuel cell stack includes 210 cells in series.
In the RMPC based control scheme, the term 𝛾 represents the upper
bound of the maximum power of the vehicle propulsion system. The
term 𝛿 represents the upper bound of the uncertainty 𝛥𝑢 on the control
input to enforce the fuel cell system and the boost converter operate
efficiently. Based on the physical constraints of the system, we set
𝛾 = 60 kW and 𝛿 = 2 kW.
Moreover, the term 𝜆 is the parameter in Theorem 1 to ensure 𝛶
is an RPI set. The term 𝐿𝑢 is the weighting scalar on control input
to balance the control effort (fuel consumption) and the terminal cost
on the state (battery charge-sustaining). The term 𝛼 2 is a parameter
before the disturbance part in (37) which plays a role to determine the
disturbance rejection level of the control system. The primary principle
to select 𝜆, 𝐿𝑢 and 𝛼 2 is to ensure that LMIs (44)–(46) have a feasible
solution while 𝜆 is in (0, 1] and 𝛼 2 is desired to be small to maintain
high-disturbance rejection level. In the case study, we set 𝛼 2 = 9𝑒−5,
𝐿𝑢 = 1.2𝑒−9 and 𝜆 = 0.06.
To assess the performance of the proposed control strategy over
different driving scenarios, we set a combined driving cycle consisting
8
Control Engineering Practice 98 (2020) 104364
Fig. 10. Prototype fuel cell stack for experiment (top) and the equivalent circuit model
in the vehicle simulator (bottom).
Fig. 11. Parameters estimation and verification of the fuel cell model.
of one Worldwide Harmonised Light Vehicles Test Procedure (WLTP)
Class 3b that was developed by the United Nations Economic Com-
mission for Europe, one Federal Test Procedure 75 (FTP-75) that was
defined by US Environmental Protection Agency (EPA), one EPA High-
way Fuel Economy Driving Schedule (HWFET), and one EPA New York
City Cycle (NYCC). By using the vehicle dynamic model in MATLAB
Powertrain BlocksetTM , the power demand of the vehicle is simulated.
Fig. 12 shows the driving velocity and the power demand of the vehicle
over a duration of 5641 s covering the total distance of 59.44 km. For
benchmarking the effectiveness of the proposed control strategies, a
dynamic programming (DP) based algorithm, in which the driving cycle
is assumed to be a priori known, is used.
4.2. Effectiveness of the fuzzy model
To verify the accuracy of the fuzzy model, we compare the battery
SoC response (corresponding to state 𝑥1 ) and DC-BUS voltage response
(corresponding to state 𝑥2 ) between the fuzzy model and the simulator
over the driving cycle in the battery-driven mode. Since MPC has the
ability to take new measurements (𝑥1 and 𝑥2 ) at each sampling time,
we further compare the effectiveness of the off-line fuzzy model and
fuzzy MPC model in the case study.
The results of the fuzzy model performance are shown in Fig. 13.
We observe that both the off-line fuzzy model and the fuzzy MPC model
work effectively to approximate the nonlinear system. The fuzzy MPC
D. Shen, C.-C. Lim and P. Shi
Fig. 12. Driving cycle used in the case study (top) and power demand of the vehicle
in the simulation (bottom).
Fig. 13. Battery SoC output comparison between the off-line fuzzy model and the
simulator (top-left), battery SoC output comparison between the fuzzy MPC model and
the simulator (top-right), and DC-Bus voltage output comparison between the fuzzy
MPC model and the simulator (bottom).
model offers better control performance than that of the off-line fuzzy
model. The reason being that new measurements (𝑥1 and 𝑥2 ) taken
at each sampling time in the fuzzy MPC model enables further model
inaccuracy correction and disturbance compensation, where else there
is no similar capability in the off-line fuzzy model.
4.3. Control performance
Shown in Fig. 14(a) are two scenarios considered in the case study:
(a) RMPC with constant 𝛾 = 60 (called ‘‘RMPC’’), and (b) RMPC with
varied 𝛾.
9
Control Engineering Practice 98 (2020) 104364
Fig. 14. Power demand of the vehicle in the case study and upper bound estimation
of the power demand under RMPC and RMPC with varied 𝛾 (top) and battery state of
charge response under RMPC, RMPC with varied 𝛾, and DP (bottom).
Fig. 14(b) shows the battery SoC response of the system in the
presence of disturbance (power demand). We observe that the battery
SoC is maintained to the reference value at the end of the driving cycle
under all control schemes.
To assess the optimisation performance of the control schemes, we
use miles per gallon gasoline equivalent (MPGe) that defined by US EPA
to measure the average distance travelled per unit of energy consumed.
The formulation to calculate MPGe is
(total mile driven) × (energy of one gallon of gasoline)
total energy consumed
Fig. 15 shows the result of the optimisation performance. We ob-
serve that the proposed control schemes are capable of achieving good
performance compared to the global optimal solution that DP provided.
While RMPC with varied 𝛾 demonstrates the ability to outperform
RMPC, which verifies the statement given in Remark 6.
Furthermore, the numerical results of battery SoC at the terminal
points and MPGe of the FCV under three control schemes over the
driving cycle are shown in Table 1. Since the optimisation horizon is
the whole driving cycle under the DP based approach, it is clear that
battery SoC is maintained at the terminal point for the whole driving
cycle, but not at the terminal point of each separate driving cycle.
When comparing with MPGe of DP, we observe that the proposed
controllers achieve sub-optimal results. However, the results in Table 1
show that the controllers are applicable for all driving cycles to main-
tain battery SoC within certain admissible range near the scheduled
reference SoC with competitive MPGe performance.
4.4. Computational time
Computational burden at each sampling step is a critical limiting
factor for a real-time EMS design. Fig. 16 shows the computational time
of RMPC with varied 𝛾 required to calculate the control law 𝑢(𝑘) at
each sampling time over the driving cycle on a PC with Intel® i7-6700
3.40 GHz CPU processor, 16 GB RAM, using Simulink R2019a. We
observe that the average time to solve the LMI optimisation problem
at each sampling step is 0.2621 s and the maximum time is 0.3877 s.
The low computational cost shows its applicability in practice.
D. Shen, C.-C. Lim and P. Shi Control Engineering Practice 98 (2020) 104364

The approach is formulated by the T–S fuzzy modelling framework

and robust MPC technique. The optimisation problem is cast in the
form LMIs and solved online. At each sampling step, the state-feedback
control law is obtained by minimising the upper bound of the cost
function, subject to system constraints. Furthermore, we incorporate
traffic condition into the controller design to further improve optimi-
sation performance. A MATLAB/Simulink based simulation serves to
illustrate the effectiveness of the control scheme. Although the energy
management design is based on fuel cell vehicles in this study, the
approach is extensible to traditional hybrid vehicle applications.
Our future work includes the implementation of the proposed con-
trol scheme into an embedded controller and the improvement of the
control scheme by considering the influence of fuel cell dynamics.
Fig. 15. Energy consumption comparison among RMPC, RMPC with varied 𝛾, and DP.

Declaration of competing interest

The authors declare that they have no known competing finan-

cial interests or personal relationships that could have appeared to
influence the work reported in this paper.

Acknowledgments

Fig. 16. Computational time of the RMPC (varied 𝛾) based control scheme in the
case study.
The authors would like to thank HySA Systems Integration & Tech-
nology Validation Competence Centre for providing the experimental
data used in this work. This research was supported in part by the
Australian Research Council under Grant DP170102644.
Appendix A. Proof of Theorem 1

Table 1 For notational convenience, we denote

Numerical results of three control schemes over the driving cycle.
∑
𝑟
WLTC ℎ𝑟𝑖 = ℎ𝑖 (𝑧(𝑘))
Control scheme Terminal SoC [%] MPGe 𝑖=1

DP 49.00 61.16 From (29)–(34), if there exists 𝜆 ∈ (0, 1] such that (34) is guaranteed,
RMPC 50.11 54.33 then 𝛶 defined in (29) is a robust positively invariant for system (26).
RMPC (varied 𝛾) 50.07 54.66
Substituting (26) and (28) into (34) gives:
WLTC+FTP−75
1 1
Control scheme Terminal SoC [%] MPGe [𝑥(𝑘 + 1)𝑇 𝑃 𝑥(𝑘 + 1) − 𝑥(𝑘)𝑇 𝑃 𝑥(𝑘)] + 𝜆[ 𝑥(𝑘)𝑇 𝑃 𝑥(𝑘)
𝜉 𝜉
DP 50.78 57.46 1 𝑇 1 𝑇
RMPC 49.99 58.27 − 𝑤(𝑘) 𝑤(𝑘) − 𝛥𝑢(𝑘) 𝛥𝑢(𝑘)]
𝛾 2 + 𝛿2 𝛾 2 + 𝛿2
RMPC (varied 𝛾) 49.99 58.55
1
WLTC+FTP−75+HWFET = {ℎ𝑟𝑖 ℎ𝑟𝑗 [(𝐴𝑖 + 𝐵𝑖 𝐾𝑗 )𝑥(𝑘) + 𝐵𝑖 𝛥𝑢(𝑘) + 𝐸𝑖 𝑤(𝑘)]}𝑇
𝜉
Control scheme Terminal SoC [%] MPGe
× 𝑃 {ℎ𝑟𝑖 ℎ𝑟𝑗 [(𝐴𝑖 + 𝐵𝑖 𝐾𝑗 )𝑥(𝑘) + 𝐵𝑖 𝛥𝑢(𝑘) + 𝐸𝑖 𝑤(𝑘)]}
DP 49.29 62.13
1 1
RMPC 50.08 59.13 − 𝑥(𝑘)𝑇 𝑃 𝑥(𝑘) + 𝜆[ 𝑥(𝑘)𝑇 𝑃 𝑥(𝑘)
RMPC (varied 𝛾) 50.06 59.38 𝜉 𝜉
1 𝑇 1
WLTC+FTP−75+HWFET+NYCC − 𝑤(𝑘) 𝑤(𝑘) − 𝛥𝑢(𝑘)𝑇 𝛥𝑢(𝑘)]
𝛾 2 + 𝛿2 𝛾 2 + 𝛿2
Control scheme Terminal SoC [%] MPGe
1 [ ]
DP 50.00 59.79
= [𝑥𝑇 (𝑘) 𝛥𝑢𝑇 (𝑘) 𝑤𝑇 (𝑘)]{ℎ𝑟𝑖 ℎ𝑟𝑗 𝐴𝑖 + 𝐵𝑖 𝐾𝑗 𝐵𝑖 𝐸𝑖 }𝑇
𝜉
RMPC 50.00 58.57
] ⎡
𝑥(𝑘) ⎤
RMPC (varied 𝛾) 50.00 58.79 [
× 𝑃 {ℎ𝑟𝑖 ℎ𝑟𝑗 𝐴𝑖 + 𝐵𝑖 𝐾𝑗 𝐵𝑖 𝐸𝑖 } ⎢𝛥𝑢(𝑘)⎥
⎢ ⎥
⎣ 𝑤(𝑘) ⎦
− [𝑥𝑇 (𝑘) 𝛥𝑢𝑇 (𝑘) 𝑤𝑇 (𝑘)]
For the dynamic programming based approach, it takes 83.3116 s to
solve the optimisation problem since the whole driving cycle is required ⎡ 1−𝜆 𝑃 ⋆ ⋆ ⎤ ⎡ 𝑥(𝑘) ⎤
⎢ 𝜉 ⎥
to consider in the computation. Although the dynamic programming ×⎢ 0
𝜆
⋆ ⎥ ⎢𝛥𝑢(𝑘)⎥
based approach is capable of providing the global optimal solution, ⎢ 0
𝛾 2 +𝛿 2 ⎢ ⎥
𝜆 ⎥ ⎣ 𝑤(𝑘) ⎦
⎣ 0 2
𝛾 +𝛿2⎦
the heavy computational burden and the requirements of the prior
information of the future power demand limit its ability in a real ≤0
application. Using the Schur complement yields
1−𝜆
5. Conclusion ⎡ 𝜉
𝑃 ⋆ ⋆ ⋆ ⎤
⎢ ⎥
⎢ 0 𝜆∕(𝛾 2 + 𝛿 2 ) ⋆ ⋆ ⎥
In this study, we blend the theoretical analysis and practical issue to ⎢ 0 0 𝜆∕(𝛾 2 + 𝛿 2 ) ⋆ ⎥
develop a new approach for energy management in fuel cell vehicles. ⎢ℎ𝑟 ℎ𝑟 (𝐴 + 𝐵 𝐾 ) ℎ𝑟𝑖 𝐵𝑖 ℎ𝑟𝑖 𝐸𝑖 ( 1𝜉 𝑃 )−1 ⎥⎦
⎣ 𝑖 𝑗 𝑖 𝑖 𝑗

10
D. Shen, C.-C. Lim and P. Shi Control Engineering Practice 98 (2020) 104364

= ℎ𝑟𝑖 ℎ𝑟𝑗 ⎡ 𝑃 ⋆ ⋆ ⋆ ⋆⎤
⎢ 0 𝛼2 ⋆ ⋆ ⋆ ⎥⎥
⎡
1−𝜆
𝑃 ⋆ ⋆ ⋆ ⎤ ⎢
𝜉 ×⎢ 0 0 𝛼2 ⋆ ⋆⎥
⎢ 0 𝜆∕(𝛾 2 + 𝛿 2 ) ⋆ ⋆ ⎥ ⎢ 𝐵𝑖 𝐾𝑗 +𝐵𝑗 𝐾𝑖 𝐵𝑖 +𝐵𝑗 𝐸𝑖 +𝐸𝑗 ⎥
⎢ ⎥ ⎢𝐴 + 𝑃 −1 ⋆⎥
⎢ 0 0 𝜆∕(𝛾 2 + 𝛿 2 ) ⋆ ⎥ ⎢ 𝐾𝑖 +𝐾𝑗 𝐿
2 2 2
⎢ 𝐵𝑖 𝐾𝑗 +𝐵𝑗 𝐾𝑖 𝐵𝑖 +𝐵𝑗 𝐸𝑖 +𝐸𝑗 1 ⎥ ⎣ 𝑢 0 0 0 𝐿𝑢 ⎥⎦
⎣𝐴 + 2 2 2
( 𝜉 𝑃 )−1 ⎦ 2
≥0
≥ 0
Therefore,
Therefore,
1−𝜆
⎡ 𝑃 ⋆ ⋆ ⋆ ⋆⎤
⎡ 𝜉
𝑃 ⋆ ⋆ ⋆ ⎤ ⎢ 0 𝛼2 ⋆ ⋆ ⋆ ⎥⎥
⎢ ⎢
⎢ 0 𝜆∕(𝛾 2 + 𝛿 2 ) ⋆ ⋆ ⎥⎥ ⎢ 0 0 𝛼2 ⋆ ⋆⎥ ≥ 0
⎢ 0 0 𝜆∕(𝛾 2 + 𝛿 2 ) ⋆ ⎥ ⎢ 𝐵𝑖 𝐾𝑗 +𝐵𝑗 𝐾𝑖 𝐵𝑖 +𝐵𝑗 𝐸𝑖 +𝐸𝑗 ⎥
⎢ 𝐵𝑖 𝐾𝑗 +𝐵𝑗 𝐾𝑖 𝐵𝑖 +𝐵𝑗 𝐸𝑖 +𝐸𝑗 ⎥ ⎢𝐴 + 2 2 2
𝑃 −1 ⋆⎥
⎣𝐴 + ( 1𝜉 𝑃 )−1 ⎦ ⎢ 𝐾𝑖 +𝐾𝑗 𝐿
2 2 2
⎣ 2 𝑢 0 0 0 𝐿𝑢 ⎥⎦
≥ 0
for 𝑖 ≤ 𝑗 subject to ℎ𝑖 ∩ ℎ𝑗 ≠ ∅ (61)
for 𝑖 ≤ 𝑗 subject to ℎ𝑖 ∩ ℎ𝑗 ≠ ∅ (60)
Substituting 𝑃 = 𝜉𝑄−1 , 𝑄 > 0, and 𝑀𝑖 = 𝐾𝑖 𝑄 into (61), and then pre-
[ ]
Substituting 𝑃 = 𝜉𝑄−1 , 𝑄 > 0, 𝑀𝑖 = 𝐾𝑖 𝑄 into (60) and then multiplying and post-multiplying by block-diag 𝑄 𝐼 𝐼 𝐼 on (61), we have
[ ]
block-diag 𝑄 𝐼 𝐼 𝐼 on both sides of (60), we have (36) and the (44).
proof is completed. From (42), we can readily obtain (46). Now, It remains to prove
(45). Substituting (28) into (43) gives
Appendix B. Proof of Theorem 2 1 𝑟 𝑟 𝑇
ℎ ℎ 𝑥 (𝑘)𝐾𝑖𝑇 𝐾𝑗 𝑥(𝑘) ≤ 1
𝜇2 𝑖 𝑗
For convenience, we denote 𝑥𝑘 (𝑖) = 𝑥(𝑘 + 𝑖|𝑘) 𝛥𝑢𝑘 (𝑖) = 𝛥𝑢(𝑘 + 𝑖|𝑘)
From (41), we have
𝑤𝑘 (𝑖) = 𝑤(𝑘 + 𝑖|𝑘) here.
The terminal cost function 𝜈𝑡 on state is required to satisfy (37). 1 𝑇
𝑥 (𝑘)𝑃 𝑥(𝑘) ≤ 1
Substituting (26) and (28) into inequality (37) yields 𝜉
If
{ℎ𝑟𝑖 ℎ𝑟𝑗 [(𝐴 + 𝐵𝑖 𝐾𝑗 )𝑥𝑘 (𝑖) + 𝐵𝑖 𝛥𝑢𝑘 (𝑖) + 𝐸𝑖 𝑤𝑘 (𝑖)]}𝑇 𝑃
1 𝑟 𝑟 𝑇 1
ℎ ℎ 𝑥 (𝑘)𝐾𝑖𝑇 𝐾𝑗 𝑥(𝑘) ≤ 𝑥𝑇 (𝑘)𝑃 𝑥(𝑘) ≤ 1
× {ℎ𝑟𝑖 ℎ𝑟𝑗 [(𝐴 + 𝐵𝑖 𝐾𝑗 )𝑥𝑘 (𝑖) + 𝐵𝑖 𝛥𝑢𝑘 (𝑖) + 𝐸𝑖 𝑤𝑘 (𝑖)]} 𝜇2 𝑖 𝑗 𝜉
then (43) holds. Moreover, since
− 𝑥𝑇𝑘 (𝑖)𝑃 𝑥𝑘 (𝑖) − ℎ𝑟𝑖 ℎ𝑟𝑗 [𝑥𝑇𝑘 (𝑖)𝐾𝑖𝑇 𝐿𝑢 𝐾𝑗 𝑥𝑘 (𝑖)]
ℎ𝑟𝑖 ℎ𝑟𝑗 𝑥𝑇 (𝑘)𝐾𝑖𝑇 𝐾𝑗 𝑥(𝑘) ≤ ℎ𝑟𝑖 𝑥𝑇 (𝑘)𝐾𝑖𝑇 𝐾𝑖 𝑥(𝑘)
− 𝛼 2 [𝑤𝑘 (𝑖)𝑇 𝑤𝑘 (𝑖) + 𝛥𝑢𝑘 (𝑖)𝑇 𝛥𝑢𝑘 (𝑖)]
if
≥ {ℎ𝑟𝑖 ℎ𝑟𝑗 [(𝐴 + 𝐵𝑖 𝐾𝑗 )𝑥𝑘 (𝑖) + 𝐵𝑖 𝛥𝑢𝑘 (𝑖) + 𝐸𝑖 𝑤𝑘 (𝑖)]}𝑇 𝑃 1 𝑇 1
ℎ𝑟𝑖 𝑥𝑇 (𝑘)[ 𝐾 𝐾𝑖 − 𝑃 ]𝑥(𝑘) ≤ 0 (62)
𝜇2 𝑖 𝜉
× {ℎ𝑟𝑖 ℎ𝑟𝑗 [(𝐴 + 𝐵𝑖 𝐾𝑗 )𝑥𝑘 (𝑖) + 𝐵𝑖 𝛥𝑢𝑘 (𝑖) + 𝐸𝑖 𝑤𝑘 (𝑖)]}
(43) holds. Substituting 𝑃 = 𝜉𝑄−1 , 𝑄 > 0, and 𝑀𝑖 = 𝐾𝑖 𝑄 into (62) gives
− 𝑥𝑇𝑘 (𝑖)𝑃 𝑥𝑘 (𝑖) − ℎ𝑟𝑖 [𝑥𝑇𝑘 (𝑖)𝐾𝑖𝑇 𝐿𝑢 𝐾𝑖 𝑥𝑘 (𝑖)] (45) by the Schur complement, and thus the proof is completed.
− 𝛼 2 [𝑤𝑘 (𝑖)𝑇 𝑤𝑘 (𝑖) + 𝛥𝑢𝑘 (𝑖)𝑇 𝛥𝑢𝑘 (𝑖)]
References
[ ]
= [𝑥𝑇 (𝑘) 𝛥𝑢𝑇 (𝑘) 𝑤𝑇 (𝑘)]{ℎ𝑟𝑖 ℎ𝑟𝑗 𝐴 + 𝐵𝑖 𝐾𝑗 𝐵𝑖 𝐸𝑖 }𝑇
Alessandri, A., Baglietto, M., & Battistelli, G. (2004). On estimation error bounds
for receding-horizon filters using quadratic boundedness. IEEE Transactions on
] ⎡
𝑥(𝑘) ⎤
[
𝐸𝑖 } ⎢𝛥𝑢(𝑘)⎥
Automatic Control, 49(8), 1350–1355.
× 𝑃 {ℎ𝑟𝑖 ℎ𝑟𝑗 𝐴 + 𝐵𝑖 𝐾𝑗 𝐵𝑖
⎢ ⎥ Bemporad, A., & Morari, M. (1999). Robust model predictive control: A survey. In
⎣ 𝑤(𝑘) ⎦ Robustness in identification and control (pp. 207–226). Springer.
Blanchini, F. (1999). Set invariance in control. Automatica, 35(11), 1747–1767.
− [𝑥𝑇 (𝑘) 𝛥𝑢𝑇 (𝑘) 𝑤𝑇 (𝑘)] Boyd, S., El Ghaoui, L., Feron, E., & Balakrishnan, V. (1994). Linear matrix inequalities
in system and control theory, Vol. 15. SIAM.
⎡𝑃 − ℎ𝑟𝑖 (𝐾𝑖𝑇 𝐿𝑢 𝐾𝑖 ) ⋆ ⋆ ⎤ ⎡ 𝑥(𝑘) ⎤ Burke, A. F. (2007). Batteries and ultracapacitors for electric, hybrid, and fuel cell
×⎢ 0 𝛼2 ⋆ ⎥ ⎢𝛥𝑢(𝑘)⎥ vehicles. Proceedings of the IEEE, 95(4), 806–820.
⎢ ⎥⎢ ⎥
⎣ 0 0 𝛼 2 ⎦ ⎣ 𝑤(𝑘) ⎦ Di Cairano, S., Bernardini, D., Bemporad, A., & Kolmanovsky, I. V. (2014). Stochastic
MPC with learning for driver-predictive vehicle control and its application to
≥0 HEV energy management. IEEE Transactions on Control Systems Technology, 22(3),
1018–1031.
From the Schur complement, we have Ehsani, M., Gao, Y., Longo, S., & Ebrahimi, K. (2018). Modern electric, hybrid electric,
and fuel cell vehicles. CRC press.
⎡𝑃 − ℎ𝑟𝑖 (𝐾𝑖𝑇 𝐿𝑢 𝐾𝑖 ) ⋆ ⋆ ⋆ ⎤ Gemmen, R., & Johnson, C. (2006). Evaluation of fuel cell system efficiency and degra-
⎢ 0 𝛼2 ⋆ ⋆ ⎥⎥ dation at development and during commercialization. Journal of Power Sources,
⎢ 159(1), 646–655.
⎢ 0 0 𝛼2 ⋆ ⎥
⎢ ℎ𝑟 ℎ𝑟 (𝐴 + 𝐵 𝐾 ) Hegazy, O., Van Mierlo, J., & Lataire, P. (2012). Analysis, modeling, and implemen-
⎣ 𝑖 𝑗 𝑖 𝑖 𝑗 ℎ𝑟𝑖 𝐵𝑖 ℎ𝑟𝑖 𝐸𝑖 𝑃 −1 ⎥⎦ tation of a multidevice interleaved DC/DC converter for fuel cell hybrid electric
vehicles. IEEE Transactions on Power Electronics, 27(11), 4445–4458.
⎡ 𝑃 ⋆ ⋆ ⋆ ⋆⎤
Hrovat, D., Di Cairano, S., Tseng, H. E., & Kolmanovsky, I. V. (2012). The development
⎢ 0 𝛼2 ⋆ ⋆ ⋆⎥
⎢ ⎥ of model predictive control in automotive industry: A survey. In 2012 IEEE
= ⎢ 0 0 𝛼2 ⋆ ⋆⎥ international conference on control applications (pp. 295–302). IEEE.
⎢ℎ𝑟 ℎ𝑟 (𝐴 + 𝐵𝑖 𝐾𝑗 ) ℎ𝑟𝑖 𝐵𝑖 ℎ𝑟𝑖 𝐸𝑖 𝑃 −1 ⋆⎥ Jiang, Z.-P., & Wang, Y. (2001). Input-to-state stability for discrete-time nonlinear
⎢ 𝑖 𝑗 𝑟 ⎥ systems. Automatica, 37(6), 857–869.
⎣ ℎ𝑖 𝐾𝑖 𝐿𝑢 0 0 0 𝐿𝑢 ⎦
Kothare, M. V., Balakrishnan, V., & Morari, M. (1996). Robust constrained model
= ℎ𝑟𝑖 ℎ𝑟𝑗 predictive control using linear matrix inequalities. Automatica, 32(10), 1361–1379.

11
D. Shen, C.-C. Lim and P. Shi
Lian, Z., He, Y., Zhang, C.-K., Shi, P., & Wu, M. (2019). Robust 𝐻∞ control for T-S fuzzy
systems with state and input time-varying delays via delay-product-type functional
method. IEEE Transactions on Fuzzy Systems.
Lian, Z., He, Y., Zhang, C.-K., & Wu, M. (2019). Stability and stabilization of TS fuzzy
systems with time-varying delays via delay-product-type functional method. IEEE
Transactions on Cybernetics.
Limón, D., Alamo, T., Salas, F., & Camacho, E. F. (2006). Input to state stability of min–
max MPC controllers for nonlinear systems with bounded uncertainties. Automatica,
42(5), 797–803.
Lofberg, J. (2004). YALMIP: A toolbox for modeling and optimization in MATLAB.
In 2004 IEEE international conference on robotics and automation (IEEE Cat. No.
04CH37508) (pp. 284–289). IEEE.
Lu, Q., Shi, P., Lam, H.-K., & Zhao, Y. (2015). Interval type-2 fuzzy model predictive
control of nonlinear networked control systems. IEEE Transactions on Fuzzy Systems,
23(6), 2317–2328.
Magni, L., Raimondo, D. M., & Scattolini, R. (2006). Regional input-to-state stability for
nonlinear model predictive control. IEEE Transactions on Automatic Control, 51(9),
1548–1553.
Mayne, D. (2016). Robust and stochastic model predictive control: Are we going in the
right direction?. Annual Reviews in Control, 41, 184–192.
Mayne, D. Q., Raković, S., Findeisen, R., & Allgöwer, F. (2006). Robust output
feedback model predictive control of constrained linear systems. Automatica, 42(7),
1217–1222.
Nonobe, Y. (2017). Development of the fuel cell vehicle Mirai. IEEJ Transactions on
Electrical and Electronic Engineering, 12(1), 5–9.
Pahlevaninezhad, M., Drobnik, J., Jain, P. K., & Bakhshai, A. (2011). A load adaptive
control approach for a zero-voltage-switching DC/DC converter used for electric
vehicles. IEEE Transactions on Industrial Electronics, 59(2), 920–933.
Payri, F., Guardiola, C., Pla, B., & Blanco-Rodriguez, D. (2014). A stochastic method
for the energy management in hybrid electric vehicles. Control Engineering Practice,
29, 257–265.
Schouten, N. J., Salman, M. A., & Kheir, N. A. (2003). Energy management strategies
for parallel hybrid vehicles using fuzzy logic. Control Engineering Practice, 11(2),
171–177.
12
Control Engineering Practice 98 (2020) 104364
Shen, D., Lim, C. C., Shi, P., & Bujlo, P. (2018). Energy management of fuel cell hybrid
vehicle based on partially observable Markov decision process. IEEE Transactions
on Control Systems Technology.
Sontag, E. D., & Wang, Y. (1995). On characterizations of the input-to-state stability
property. Systems & Control Letters, 24(5), 351–359.
Sorrentino, M., Rizzo, G., & Arsie, I. (2011). Analysis of a rule-based control strategy for
on-board energy management of series hybrid vehicles. Control Engineering Practice,
19(12), 1433–1441.
Staunton, R. H., Ayers, C. W., Marlino, L., Chiasson, J., & Burress, B. (2006). Evaluation
of 2004 Toyota Prius hybrid electric drive system: Tech. Rep., Oak Ridge, TN (United
States): Oak Ridge National Lab.(ORNL).
Tanaka, K., & Wang, H. O. (2004). Fuzzy control systems design and analysis: a linear
matrix inequality approach. John Wiley & Sons.
Tütüncü, R. H., Toh, K.-C., & Todd, M. J. (2003). Solving semidefinite-quadratic-linear
programs using SDPT3. Mathematical Programming, 95(2), 189–217.
Wei, L., Qiao, J., & Liu, Y. (2018). Data acquisition and processing of driving operation
system for vehicle dynamics simulation. ICIC Express Letters, Part B: Applications, 9,
819–826.
Wipke, K., Sprik, S., Kurtz, J., Ramsden, T., Ainscough, C., & Saur, G. (2012). National
fuel cell electric vehicle learning demonstration final report: Tech. Rep., NREL.
Xia, Y., Yang, H., Shi, P., & Fu, M. (2010). Constrained infinite-horizon model predictive
control for fuzzy-discrete-time systems. IEEE Transactions on Fuzzy Systems, 18(2),
429–436.
Xue, H., Yang, J., Zhang, L., & Shi, C. (2017). Optimization of vehicle scheduling
problem with potential demand based on CS-GA. ICIC Express Letters, 11, 451–456.
Yakubovich, V. (1992). Nonconvex optimization problem: The infinite-horizon linear-
quadratic control problem with quadratic constraints. Systems & Control Letters,
19(1), 13–22.
Yang, W., Feng, G., & Zhang, T. (2014). Robust model predictive control for discrete-
time Takagi–Sugeno fuzzy systems with structured uncertainties and persistent
disturbances. IEEE Transactions on Fuzzy Systems, 22(5), 1213–1228.
Zeng, X., & Wang, J. (2015). A parallel hybrid electric vehicle energy management
strategy using stochastic model predictive control with road grade preview. IEEE
Transactions on Control Systems Technology, 23(6), 2416–2423.