Energy 198 (2020) 117289

Contents lists available at ScienceDirect

Energy
journal homepage: www.elsevier.com/locate/energy

A driving pattern recognition-based energy management for plug-in

hybrid electric bus to counter the noise of stochastic vehicle mass
Hongqiang Guo*, Daizheng Hou, Shangye Du, Ling Zhao, Jian Wu, Ning Yan
School of Mechanical & Automotive Engineering, Liaocheng University, Liaocheng, 252000, China

a r t i c l e i n f o a b s t r a c t

Article history: Because the strong coupling relationship between energy management and required power, the Pon-
Received 23 August 2019 tryagin’s Minimum Principle (PMP)-based energy management should consider the noise of stochastic
Received in revised form vehicle mass for plug-in hybrid electric bus (PHEB). However, if the vehicle mass is evaluated on-line, the
6 February 2020
control complexity will be greatly increased. This paper proposes a driving pattern recognition method
Accepted 29 February 2020
to address the problem. The method is constituted by a look-up table and the K-nearest neighbor al-
Available online 5 March 2020
gorithm (KNN). The look-up table is used to recognize the robust design value (the inverse value of the
robust co-state), where the average velocity at every bus station is taken as input, and the robust design
Keywords:
Plug-in hybrid electric bus
value is taken as output. More importantly, the robust design value is found off-line by Design For Six
Energy management Sigma (DFSS) method, and can counter the noise of stochastic vehicle mass. Because of this, the noise of
Driving pattern recognition the stochastic vehicle mass can be neglected in adaptive energy management control. The Monte Carlo
DFSS Simulation (MCS) and simulation test results show that the proposed method is reasonable, robust and
Stochastic vehicle mass applicable; the fuel economy can be averagely improved by 34.36%, compared to a rule-based energy
management.
© 2020 Elsevier Ltd. All rights reserved.

1. Introduction should be adjusted based on the driving conditions. In contrast, the

optimization-based strategies such as dynamic programming (DP),
The energy saving becomes one of the most the key issues in pontryagin’s minimum principle (PMP) and equivalent consump-
different fields [1e3]. The Plug-in hybrid electric vehicle (PHEV) is tion minimization strategy (ECMS) et al. can realize optimal or
recognized as a promising technology to mitigate the problem of approximately optimal control, once the driving conditions are
energy crisis [4,5]. It can both prolong the driving range and known in a prior. However, they are difficult to be used in stochastic
recharge the battery from power grid, by combining the advantages traffic circumstance [12].
of hybrid electric vehicle (HEV) and pure electric vehicle (PEV) [6]. To realize adaptive control in real world, many predictive
Therefore, it is highly favored all over the world [7]. methods are proposed by deploying optimization-based strategies
Energy management is of great importance for fuel economy [13]. The first is the receding horizon control (RHC) method.
improvement of PHEV, and many methods have been proposed [8]. Ref. [14] proposed a RHC-based energy management by combining
The rule-based strategy has good real-time and robust control a linear state of charge (SOC) reference trajectory, a Markov chain-
performances [9]. However, the thresholds are usually designed as based velocity prediction algorithm and DP. Ref. [15] proposed a
constant values [7]. Therefore, the strategy may have low adapt- control strategy by employing a dynamic SOC plan method, a
ability in stochastic traffic environment, and the fuel economy is Markov chain-based velocity prediction algorithm and PMP.
usually deteriorated. To improve the fuel economy, some rule- Ref. [16] proposed a conventional model predictive control (MPC)-
based strategies strive to make the PEHV work in blended based energy management, by proposing a reference SOC plan
charge-depletion (BCD) mode over the whole trip [10,11]. However, method with dynamic traffic feedback data. Nevertheless, the
this is also a difficult work in practice, because the thresholds reference SOC plan method is indispensable [17]. However, the
optimal reference SOC cannot be completely predicted, due to the
stochastic and unrepeatable characteristics of driving conditions.
Moreover, because the reference SOC is usually taken as the control
* Corresponding author. objective at every receding horizon, the fuel economy may be
E-mail address: [email protected] (H. Guo).

https://doi.org/10.1016/j.energy.2020.117289
0360-5442/© 2020 Elsevier Ltd. All rights reserved.
2 H. Guo et al. / Energy 198 (2020) 117289

thereby constructing a look-up table to realize adaptive energy

Abbreviations management control. Here, the look-up table extraordinarily takes
the average velocity at every bus station as input and the robust
PMP Pontryagin’s minimum principle design value as output. Moreover, a KNN algorithm is proposed to
PHEB Plug-in hybrid electric bus recognize the similar driving cycle from the sampling set. The main
KNN K-nearest neighbor algorithm innovation is that although the stochastic vehicle mass is consid-
DFSS Design For Six Sigma ered in the DFSS, it can be neglected in the driving pattern
MCS Monte Carlo Simulation recognition-based energy management. The main reasons are
PHEV Plug-in hybrid electric vehicle summarized as follows.
HEV Hybrid electric vehicle Firstly, the DFSS can find the “flat” zone of the design variable
PEV Pure electric vehicle (the inverse value of the co-state). That is to say, although the noise
BCD Blended charge-depletion of the stochastic vehicle exists in reality, the robust design value can
DP Dynamic programming always make the terminal SOC of the PHEB locate in the defined
ECMS Equivalent consumption minimization strategy feasible zone. In other words, the robust design value is insensitive
RHC Receding horizon control to the noise of the stochastic vehicle mass.
SOC State of charge of the battery Secondly, because the noise of the stochastic vehicle mass has
MPC Model predictive control weak effect on the robust design variable, it can be neglected in the
AMT Automated mechanical transmission driving pattern recognition-based energy management control,
OPT, LHD Optimal Latin hypercube design thereby reducing the control complexity.
MIGA Multi-island genetic algorithm The remainder of this paper is structured as follows. The com-
CS Charge-sustaining ponents together with the parameters of the PHEB are described in
section 2. The formulation of the driving pattern recognition-based
energy management is detailed in section 3. The results and dis-
cussion of the DFSS are presented in Section 4. The simulation re-
further sacrificed with the non-optimal reference SOC. The second sults and discussion are described in section 5, and the conclusions
is the driving pattern recognition-based energy management. It can are drawn in section 6.
directly identify the similar historical driving cycle, and then pro-
vide feasible control vector. More importantly, it can avoid the
prediction of reference SOC. Ref. [18] proposed an adaptive control
strategy based on K-nearest neighbor (KNN) algorithm. The optimal
control vector was firstly optimized off-line by a series of driving
cycles, and then the similar control vector was executed on-line.
Besides, PMP-based method is also proposed. Ref. [19] proposed a
co-state recognition method by analyzing the past driving patterns.
The simulation results showed that the method was reasonable and
had great promising for practical application.
However, seldom investigations consider the noise of stochastic
vehicle mass. In fact, it is of great importance for the optimal energy
management control, especially for city bus. For example, Ref. [20]
discussed the coupling relationship between the stochastic vehicle
mass and the component sizing. Because the energy management
was taken as the calculation module of the component sizing, the
stochastic vehicle mass might indirectly influence the energy
management. Ref. [17] proposed a RHC-based energy management
for plug-in hybrid electric bus (PHEB) by designing a predictive Fig. 1. The configuration of the PHEB.
model of terminal SOC constraint, where the influence of the sto-
chastic vehicle mass on the energy management was demon-
strated. In addition, Ref. [21] proposed a correctional DP-based
energy management in consideration of this noise, where the sto-
chastic vehicle mass was also taken as a noise for the energy
management.
This paper proposes a driving pattern recognition-based energy
management for a PHEB with PMP algorithm. Because the sto-
chastic vehicle mass has strong effect on the energy management, it
inevitably has significant effect on the recognition of the co-state.
However, if the stochastic vehicle mass is considered in the
driving pattern recognition method, the control complexity will be
increased. Moreover, the evaluation precision of the vehicle mass
should also be well-controlled; otherwise the energy management
control performance may be deteriorated. Therefore, how to design
a robust strategy that makes the co-state has low sensitive to the
stochastic vehicle mass is indispensable.
The main contribution of this paper is that a Design For Six
Sigma (DFSS) method is proposed to find the robust design values
(the inverse values of the co-state) for different driving cycles, Fig. 2. The fuel consumption rate MAP of the engine.
 H. Guo et al. / Energy 198 (2020) 117289 3

Fig. 3. The efficiency MAP of the motor.

the hybrid driving mode or the active charging mode can be real-
ized. The curb weight of the vehicle is assumed as 10,500 kg; the
engine is a six-cylinder diesel engine with the maximum power of
167 kW and the maximum torque of 850Nm; the motor is a per-
manent magnet synchronous motor with the maximum power of

Fig. 4. The Rint battery model.

2. The components together with the parameters of the PHEB

As shown in Fig. 1, an engine, a clutch, a motor and an automated

mechanical transmission (AMT) are mechanically coupled to form
the single-parallel configuration of the PHEB. If the clutch is dis-
engaged, the single-motor driving mode or regenerative braking
mode can be realized; otherwise, the single-engine driving mode, Fig. 6. The road slope.

Fig. 5. The running characteristic of the PHEB.

 Fig. 7. The historical driving cycles.

Fig. 8. The iteration histories of the combined driving cycles: (a) the l; (b) the terminal SOC; (c) the fuel consumption.
 H. Guo et al. / Energy 198 (2020) 117289 5

Fig. 8. (continued).

Table 1 The instantaneous fuel consumption of the engine is indis-

The deterministic l. pensable and can be obtained by
The combined drivng cycle The l The average value

No.1 2932.99 2930.06

No.2 2935.32
Te ,ne be ðTe ; ne Þ
No.3 2930.14 fe ¼ , (1)
No.4 2922.29 9550 3600
No.5 2928.90
No.6 2937.26 where fe denotes the instantaneous fuel consumption of the en-
No.7 2930.22 gine; Te denotes the torque of the engine; ne denotes the speed of
No.8 2923.33
the engine; be ðTe ; ue Þ denotes the fuel consumption rate of the
engine, which can be interpolated by a look-up table formulated by
the fuel consumption rate MAP (Fig. 2).
130 kW and the maximum torque of 850 Nm; the AMT has six The efficiency of the motor can be interpolated by the efficiency
speed ratios: 6.39, 3.97, 2.4, 1.48, 1 and 0.73; the speed ratio of the MAP of the motor (Fig. 3), and the power of the motor can be ob-
final drive is 5.57; The capacity of the battery is 50 Ah. tained by
6 H. Guo et al. / Energy 198 (2020) 117289

Fig. 9. The historical driving cycles.

Fig. 10. The principle the MCS and the descriptive sampling method.

 
C Av2
Tr ¼ mgðsin a þ fr cos aÞ þ D þ dmav ,rw (4)
sgnðTmÞ 21:15
Tm ,nm hm
Pm ¼ (2)
9550 where Tr denotes the torque; m denotes the vehicle mass; a de-
notes the road grade; fr denotes the friction coefficient; CD denotes
where Pm denotes the power of the motor; Tm denotes the torque of
the coefficient of the air resistance; A denotes the windward area; v
the motor; nm denotes the speed of the motor; hm denotes the
denotes the velocity; d denotes the rotating mass conversion factor;
efficiency of the motor, which can be interpolated by a look-up
av denotes the acceleration; rw denotes the wheel radius.
table. Here, the look-up table is formulated by taking the speed
and the torque of the motor as inputs, and the efficiency as output.
As shown in Fig. 4, the battery can be modeled by the Rint 3. The formulation of the driving pattern recognition-based
model, which can be described as energy management

3.1. The formulation of the PMP

8·
> 1
>
> S OC ¼ Q Ib In our previous works, we have defined a PMP-based energy
< b management [22], where the objective function is designed as
 qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi (3)
>
> 1
>
: Ib ¼ Voc  V 2oc  4Rb Pb ðtf
2Rb
J ¼ ½fe ðuðtÞÞ þ hjshiftðtÞjdt (5)
where Voc denotes the open-circuit voltage; Pb denotes the power 0
of the battery; Ib denotes the battery current, Rb denotes the in-
ternal resistance; Qb denotes the battery capacity. where uðtÞ denotes the control vector, which is described as uðtÞ ¼
Based on the driving equation of the vehicle, the equivalent ½throttleðtÞ; shiftðtÞ’ . Here, throttleðtÞ denotes the throttle of the
resistance torque on the wheels can be described as engine, which is ranged from 0 to 1; shiftðtÞ denotes the shifting
 H. Guo et al. / Energy 198 (2020) 117289 7

Fig. 11. The principle of the MCS.

Fig. 12. The probability distribution of the terminal SOC: (a) the terminal SOC of the average method; (b) the terminal SOC of the driving pattern recognizaion method with
deterministic l.

instruction of the AMT, which is designed as 1, 0 and 1 to denote and Te;max , respectively; the boundaries of Tm are denoted by Tm;min
the action of downshift, hold on and upshift, respectively; h de- and Tm;max , respectively.
notes the weighted factor. Moreover, the control vector is sampled In addition, based on the principle of the PMP, the Hamiltonian
by Optimal Latin hypercube design (Opt, LHD) and is taken as a function is described as
compact format (one-dimension). In addition, some constraints are
also defined as ’
HðSOCðtÞ; uðtÞ; l ðtÞ; tÞ ¼ fe ðuðtÞÞ þ hjshiftðtÞj
8 qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
>
> n  ne ðtÞ  ne_max Voc ðSOCðtÞÞ  V 2oc ðSOCðtÞÞ  4Rb ðSOCðtÞÞPb ðuðtÞÞ
< ne_min  n ðtÞ  n ’
m_min m m_max  l ðtÞ
S:t (6) 2Rb ðSOCðtÞÞQnom
> Te;min ðtÞ  Te ðtÞ  Te;max ðtÞ
>
:
Tm;min ðtÞ  Tm ðtÞ  Tm;max ðtÞ (7)

where ne_min , ne_max and nm_min , nm_max denote the corresponding where l’ ðtÞ is the co-state. In practice, the initial co-state is un-
boundaries of ne and nm ; the boundaries of Te are denoted by Te;min known and the co-state is difficult to be solved by analytic method.
8 H. Guo et al. / Energy 198 (2020) 117289

Fig. 15. The average velocities at different bus stations for different combined driving
cycles.

l’ ðtÞ ¼  lðtÞ (8)

where lðtÞ denotes the design variable. It is also named co-state in

our previous work, because it is just the inverse value of the co-
state (l’ ðtÞ). Moreover, the recognition of the co-state can be also
recognized as the recognition of the design variable. In addition, Eq.
Fig. 13. The solutions of the deterministic optimization and DFSS. (7) can be also written as

H’ ðSOCðtÞ; uðtÞ; lðtÞ; tÞ ¼ fe ðuðtÞÞ þ hjshiftðtÞj

Because the variable co-state has weak influence on the fuel con- qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
sumption, the co-state can be taken as constant and be identified by Voc ðSOCðtÞÞ  V 2oc ðSOCðtÞÞ  4Rb ðSOCðtÞÞPb ðuðtÞÞ
recognition method [19]. In our PMP programming, a design vari- þ lðtÞ
2Rb ðSOCðtÞÞQnom
able for the l’ ðtÞ is defined as
(9)
The optimal control vector can be found by minimizing the

Fig. 14. The design stream of the driving pattern recognition-based energy management.
 H. Guo et al. / Energy 198 (2020) 117289 9

Fig. 16. The iteration histories of the DFSS for the combined driving cycles: (a) The Sigma-level (s) of the terminal SOC; (b) the objective value and the l; (c) the partially enlarged
figure of the l; (d) the objective value of the feasible point.

Hamiltonian function, which can be presented by 3.3. The noise of stochastic vehicle mass

’
u* ðtÞ ¼ arg minfHðSOCðtÞ; uðtÞ; l ðtÞ; tÞg (10) As shown in Fig. 5, the mass of the PHEB will show strong sto-
chastic, due to the stochastic distribution of passenger mass. For
where u* ðtÞ denotes the optimal control vector. example, assuming the passenger number in road segment i is 5,
one passenger gets off and 20 passengers get on at the bus station
iþ1. Then, the passenger number in road segment iþ1 becomes 24.
3.2. The KNN algorithm If the average passenger mass is assumed as 70 kg, the vehicle mass
will increase 1330 kg. This will greatly influence the recognition of
The KNN algorithm is one of the most popular and suitable the design variable. Therefore, the noise of stochastic vehicle
classification methods [23]. It is also suitable for the driving pattern should be considered in the driving pattern recognition method.
recognition in this paper. The Euclidian distance between the test
sample and the sampling set is selected and is described as 3.4. The problem of the deterministic design method
 2
dm n ¼ sqrt vn  vcyc n (11) Just as stated above, the stochastic vehicle mass has great effect
cyc ¼ 1; 2; ,,,8; n ¼ 1; 2; ,,,25 on the energy management. However, many energy managements
usually take the vehicle mass as constant. This may make the

where vn denotes the average velocity at bus station n; vcyc n de- adaptive energy management far away from reality. To clarify this
notes the average velocity at bus station n; cyc denotes the index of problem, a verification is carried out by taking the vehicle mass as
the historical driving cycle. It is worth mentioning that only the constant (12600 kg). In specific, the Multi-island genetic algorithm
nearest driving cycle will be selected in the driving pattern recog- (MIGA) is deployed and a series of combined driving cycles are
nition method. constructed. The combined driving cycles are constructed by the
10 H. Guo et al. / Energy 198 (2020) 117289

Fig. 16. (continued).

road slope and a series of historical driving cycles (Figs. 6 and 7). SOC should be redesigned, if the minimum value of the terminal
Moreover, the number of the combined driving cycle is also the SOC is designed as other value. However, the design method is
number of the historical driving cycle. In addition, to avoid the similar.
battery over-discharge, the minimum value of the terminal SOC is The best found design value (l) should make the terminal SOC
designed as 0.25 in this paper. In this case, the lower boundary of satisfy the terminal SOC constraint, whilst making the fuel con-
the terminal SOC in the deterministic optimization should be sumption (L/100 km) reach the minimum value. Based on the
higher than 0.25. The reasons are listed as follows. Firstly, because iteration histories of the combined driving cycles in Fig. 8, the best
the unrepeatable and stochastic characteristics of driving cycles in found design values named deterministic design values can be
real-world, the completely same driving cycles do not exist, and found. It is worth noting that the best found design value usually
then the driving pattern recognition error is inevitable. Therefore, if makes the terminal SOC locate on the constraint boundary, despite
the lower boundary of the terminal SOC in the deterministic opti- of any combined driving cycle. Therefore, the control performance
mization is designed as 0.25, the reliability of the adaptive energy may be easily disturbed once noise occurs. In addition, the detailed
management control may be greatly deteriorated. Secondly, best found design values for all of the combined driving cycles are
although the best found l can be found by the deterministic opti- listed in Table 1.
mization method with the constant vehicle mass, the vehicle mass To verify the reliability of the driving pattern recognition
in real-world is stochastic. This will further reduce the reliability. In method with the deterministic l, other 4 combined driving cycles
addition, the upper boundary of the terminal SOC should also be named from No. 9 to No. 12 are also deployed based on 4 different
limited to ensure the fuel economy. Therefore, the deterministic historical driving cycles (Fig. 9). Moreover, although many data
optimization is formulated by analysis methods have been proposed [24], the Monte Carlo
Simulation (MCS) is deployed to ensure the evaluate precision. As
Min Fu ¼ Fuel shown in Fig. 10, it is a statistical approach, which can evaluate the
S:t: 0:27  SOCT  0:32 (12) probability distribution of the responses by sampling the stochastic
2850  l  3000 variables. Moreover, the descriptive sampling method is selected to
improve the computation efficiency.
where Fuel denotes the fuel consumption; SOCT denotes the ter- As shown in Fig. 11, the driving pattern recognition-based en-
minal SOC. It is worth noting that the feasible zone of the terminal ergy management is deployed as the solving module of the MCS. An
 H. Guo et al. / Energy 198 (2020) 117289 11

Fig. 16. (continued).

experiment matrix will be firstly constructed based on the the calculation of the vehicle mass. Finally, if the simulation num-
descriptive sampling method. Here, the maximum simulation ber is bigger than the maximum simulation number, the verifica-
number is designed as 600; the combined driving cycle and the tion will be terminated; otherwise, the verification will continue.
passenger number at every road segment are designed as stochastic To evaluate the driving pattern recognition method with
variables. In terms of the road segments, they have been defined in deterministic l in Table 1, another method named average method
our previous works to ensure the noise of stochastic vehicle mass in is also proposed, where the average value of the l in Table 1 is
every potential road segment is considered [17,22]. Because the designed as the design value. From Fig. 12, it can be seen that the
stochastic characteristic of driving conditions, the driving distances reliability of the driving pattern recognition method is better than
will slightly different for different driving cycles. Moreover, a road the average method. However, there is still 32% failure probability
segment from the destination location to the department location for the driving pattern recognition method. This implies that the
also exists. Therefore, the road segments are only roughly defined driving pattern recognition method with deterministic l is also
based on the general driving distance distributions of the historical unreliable.
driving cycles. In fact, the rough definition of the road segments has
slightly effect on the investigation of the energy management,
3.5. The formulation of the DFSS method
because the required power is stochastic and the variation of the
stochastic vehicle mass will be changed to the variation of the
From the above analysis, it can be seen that the driving pattern
required power. In addition, the stochastic variable of the passenger
recognition method with deterministic l has low reliability. How-
number can be also recognized as the stochastic variable of pas-
ever, if the l is designed by DFSS method (Fig. 13), it can counter the
senger mass or the vehicle mass. The stochastic vehicle mass is
noise of the stochastic vehicle mass. In this case, the reliability can
calculated by the curb weight of the vehicle added by the passenger
be greatly improved, and the noise of the stochastic vehicle mass
mass, and the passenger mass is calculated by the passenger
can be neglected in the adaptive energy management control.
number multiplied by the average passenger mass (70 kg). The
maximum passenger number is assumed as 60, and the passenger
number is firstly designed as continuous variable in the experiment 3.5.1. The mathematical model of the DFSS
matrix while it will be rounded in the PMP programming to simply To find the robust design value, the terminal SOC constraint is
defined as a soft constraint ranging from 0.27 to 0.33. The objective
12 H. Guo et al. / Energy 198 (2020) 117289

Fig. 16. (continued).

of the DFSS is to minimize the weighed sum of the mean and variation of the design variable; pk denotes the passenger number
standard deviation of fuel consumption (L/100 km). The mathe- in every road segment.
matical model of the DFSS can be formulated by
3.5.2. The design stream of the driving pattern recognition-based
Min Fu ¼ u1 mF þ u2 sF
energy management
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u Different from deterministic optimization, the most important
1X n u 1 X n
mF ¼ Fi ; sF ¼ t ðF  mF Þ2 issue of the DFSS is the reliability analysis for every control factor
n i¼1 n  1 i¼1 i
[25]. In this paper, the reliability is described by s, and the reli-
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ability level is designed as 6s to ensure the reliability.
u
1X n u 1 X n  2 As shown in Fig. 14, the method is constituted by two steps: the
mSOCT ¼ SOCT ðjÞ; sSOCT ¼ t SOCT ðjÞ  mSOCT
n j¼1 n  1 j¼1 DFSS design and adaptive control. In terms of the DFSS design, it is
characterized by double layers. The outer layer is designed as
mðkÞ ¼ 10500 þ 70$pk ; k ¼ 1; 2; 3…25 Hooke-Jeeves algorithm. The inner layer is constituted by MCS al-
gorithm and the PMP-based energy management. The design var-
S:t: 0:27 þ 6sSOCT  mSOCT  0:33  6sSOCT iable is designed as the control factor of the DFSS, and will be
sampled by the designed standard deviation. In addition, the pas-
2850 þ Dl  l  3000  Dl
senger number at every road segment is designed as stochastic
0  pk  60 variable and is assumed as normal distribution. Moreover, the
13 combined driving cycles from No.1 to No.8 are once deployed. The
design stream for every combined driving cycle is listed as follows.
where Fu denotes the objective; mF denotes the mean of the fuel Step 1: Providing a control factor (design variable, l) within the
consumption; sF denotes the standard deviation of the fuel con- design space, based on Hooke-Jeeves algorithm. The initial value is
sumption; u1 and u2 denote the weighted coefficients, where u1 is designed as the average value of the l in Table 1. Moreover, the l
designed as 1 and u2 is designed as 10; mSOCT and sSOCT denotes the will be taken as a mean and a series of points will be sampled based
mean and standard deviation of the terminal SOC; Dl denotes the on the designed standard deviation.
 H. Guo et al. / Energy 198 (2020) 117289 13

Step 2: Modeling the passenger number at every road segment 4. The results and the discussion of the DFSS
as noise (the mean is designed as 30) with normal distribution.
Step 3: Generating the experiment matrix, and then calculating From Fig. 15, it can be seen that the average velocities at bus
the responses (Sigma-level of the terminal SOC, the mean and stations are different for different combined driving cycles. This
standard deviation of the fuel consumption) based on the results implies that the average velocity can be taken as the eigenvalue in
from the PMP-based energy management. the driving pattern recognition-based energy management.
Step 4: The responses are transmitted to the Hooke-Jeeves al- Moreover, the combined driving cycles can be taken as the sam-
gorithm. If the responses satisfy the design objective, stop the pling set.
robust design; otherwise, repeating the steps from Step 1 to Step 4. As described in Eq. (13), the terminal SOC should satisfy the
To construct the look-up table, the corresponding average ve- reliability requirement, meanwhile the objective value should
locities at bus stations of the combined driving cycles are extracted. reach the minimum value. From Fig. 16 (a), it can be seen that not all
Moreover, the robust design values of the combined driving cycles of the design points satisfy the design requirement. From Fig. 16 (b),
will also be implemented into the look-up table. it can be seen that the fluctuation zone of l will be limited once the
In terms of the adaptive control, the driving pattern recognition- approximately robust l is found. Moreover, a slight variation of the
based energy management is deployed and the noise of stochastic l can cause a big fluctuation of the objective value. However, from
vehicle mass is neglected. When the vehicle runs, the average ve- the partially enlarged figure of the l in Fig. 16 (c), it can be seen that
locity will be evaluated at every bus station and the most similar the l always fluctuates, despite it seems to be constant in Fig. 16 (b).
driving cycle in the look-up table will be recognized based on the In addition, the lowest objective values cannot denote the robust
KNN algorithm. Then the corresponding design variable will be design point. For example, the objective value of iteration 2 for the
recognized and transmitted to the PMP-based energy management. combined driving cycle 1 is the lowest, while it is not the robust
design point, because the Sigma-level of the terminal SOC in the
lower boundary is only 0.39s. The similar cases also occur in the
Table 2
DFSS designs for the combined driving cycles of 3, 5, 6 and 7, where
The robust l.
the corresponding iteration numbers are marked by black arrows
The combined drivng cycle The l The average value (Fig. 16 (b)). Fig. 16 (d) shows the feasible design points of the
No.1 2942.41 2938.82 combined driving cycles, where the Sigma-levels of the terminal
No.2 2942.22 SOC at two boundaries are always satisfy the design requirement.
No.3 2940.56 Since the reliability levels of the terminal SOC have been ensured,
No.4 2932.35
No.5 2937.53
the robust l can be selected by the minimum value of the objective
No.6 2947.12 value. Moreover, the robust l can be found for every combined
No.7 2938.31 driving cycle. Based on Fig. 16 (d), the robust design values (l) for all
No.8 2930.06 of the combined driving cycles are list in Table 2.

Fig. 17. The probability distribution of the terminal SOC: (a) the terminal SOC of the average method; (b) the terminal SOC of the driving pattern recognition method with robust l.

Fig. 18. The combined driving cycles.

14 H. Guo et al. / Energy 198 (2020) 117289

Fig. 19. The simulation results of the combined driving cycles: (a) The recognized l and l’ ; (b) The SOC trajectories.

Table 3
The fuel consumptions of the control strategies.

The combined driving cycle A-PMP (L/100 km) Rule-based (L/100 km) A-PMP vs. Rule-based (%)

No.13 9.2523 13.9813 33.82

No.14 9.6151 14.1018 31.82
No.15 9.6084 14.7610 34.91
No.16 9.9249 15.7299 36.90

To evaluate the reliability of the driving pattern recognition from No. 13 to No. 16 are designed to verify the driving pattern
method with robust l, the similar MCS method is once deployed. recognition-based energy management. Different from the above
The only difference is that the deterministic l in Table 1 is changed definition, the combined driving cycles are constituted by the his-
to the robust l in Table 2. Moreover, the average method with the torical driving cycles from No. 9 to No. 12, the road slope and 4
average value in Table 2 is also deployed. From Fig. 17, it can be seen stochastic distributions of passenger mass. Meanwhile, a rule-
that the average method is still not reliable, due to the existing 7% based energy management is also deployed to evaluate the pro-
failure probability. In contrast, the driving pattern recognition posed control strategy.
method with robust l is reliable, because the terminal SOCs are As shown in Fig. 19, the co-state (l’ ) can be easily recognized by
always satisfy the design requirement. the driving pattern recognition method (the strategy is denoted by
A-PMP), despite of any combined driving cycle. The l’ will be
5. The results and discussion of the simulation adjusted with the time moving of the combined driving cycle, while
the adjusting frequency is low. This implies that the combined
As shown in Fig. 18, a series of combined driving cycles named driving cycles in the fixed city bus route have some similarity.
 H. Guo et al. / Energy 198 (2020) 117289 15

Moreover, the SOC trajectories can continuously decrease to the [2] Taner T, Sivrioglu M. Energyeexergy analysis and optimisation of a model
sugar factory in Turkey. Energy 2015;93:641e54.
objective zone, and the charge-sustaining (CS) mode is also inevi-
[3] van Velzen A, Annema JA, van de Kaa G, van Wee B. Proposing a more
table for the A-PMP strategy. In contrast, the rule-based control comprehensive future total cost of ownership estimation framework for
strategy has more obvious CS mode, and has higher fuel con- electric vehicles. Energy Pol 2019;129:1034e46.
sumption than the A-PMP strategy (Table 3). From Table 3, it can be [4] Fathabadi H. Utilizing solar and wind energy in plug-in hybrid electric vehi-
cles. Energy Convers Manag 2018;156:317e28.
seen that the fuel economy can be averagely improved by 34.36%, [5] Yang C, Li L, You S, Yan B, Du X. Cloud computing-based energy optimization
compared to the rule-based control strategy. control framework for plug-in hybrid electric bus. Energy 2017;125:11e26.
[6] Duan B, Wang Q, Zeng X, Gong Y, Song D, Wang J. Calibration methodology for
energy management system of a plug-in hybrid electric vehicle. Energy
6. Conclusions Convers Manag 2017;136:240e8.
[7] Huang Y, Wang H, Khajepour A, Li B, Ji J, Zhao K, Hu C. A review of power
This paper proposes a driving pattern recognition-based energy management strategies and component sizing methods for hybrid vehicles.
Renew Sustain Energy Rev 2018;96:132e44.
management for a PHEB with PMP algorithm. The main innovation [8] Sabri MFM, Danapalasingam KA, Rahmat MF. A review on hybrid electric
is that the noise of stochastic vehicle mass is only considered in the vehicles architecture and energy management strategies. Renew Sustain En-
robust design (off-line), and is in turn neglected in the adaptive ergy Rev 2016;53:1433e42.
[9] Peng J, He H, Xiong R. Rule based energy management strategy for a
control. The main conclusions are summarized as follows. serieseparallel plug-in hybrid electric bus optimized by dynamic program-
ming. Appl Energy 2017;185:1633e43.
(1) Since the energy management is sensitive to the required [10] Duan BM, Wang QN, Wang JN, Li XN, Ba T. Calibration efficiency improvement
of rule-based energy management system for a plug-in hybrid electric vehicle.
power, and the required power is mainly dependent on the
Int J Automot Technol 2017;18(2):335e44.
vehicle mass, the driving pattern recognition method should [11] Padmarajan BV, McGordon A, Jennings PA. Blended rule-based energy man-
consider the noise of stochastic vehicle mass. agement for PHEV: system structure and strategy. IEEE Trans Veh Technol
(2) The DFSS method can well solve the robust design problem 2015;65(10):8757e62.
[12] Huang Y, Wang H, Khajepour A, He H, Ji J. Model predictive control power
of the PMP-based energy management, by taking the design management strategies for HEVs: a review. J Power Sources 2017;341:
variable as control factor and the stochastic vehicle mass as 91e106.
[13] Zhou Y, Ravey A, Pe ra MC. A survey on driving prediction techniques for
noise. The DFSS results demonstrate that the robust design
predictive energy management of plug-in hybrid electric vehicles. J Power
value exists in any driving cycle. Sources 2019;412:480e95.
(3) The MCS and simulation results demonstrate that the pro- [14] Li G, Zhang J, He H. Battery SOC constraint comparison for predictive energy
posed method is reasonable, robust and applicable. The noise management of plug-in hybrid electric bus. Appl Energy 2017;194:578e87.
[15] Xie S, Hu X, Xin Z, Brighton J. Pontryagin’s minimum principle based model
of stochastic vehicle mass can be neglected in the adaptive predictive control of energy management for a plug-in hybrid electric bus.
energy management control. Moreover, the simulation re- Appl Energy 2019;236:893e905.
sults show that the fuel economy can be averagely improved [16] Sun C, Moura SJ, Hu X, Hedrick JK, Sun F. Dynamic traffic feedback data
enabled energy management in plug-in hybrid electric vehicles. IEEE Trans
by 34.36% for the proposed strategy, compared to the rule- Contr Syst Technol 2014;23(3):1075e86.
based control strategy. [17] Guo H, Lu S, Hui H, Bao C, Shangguan J. Receding horizon control-based en-
ergy management for plug-in hybrid electric buses using a predictive model
of terminal SOC constraint in consideration of stochastic vehicle mass. Energy
Declaration of competing interest
2019;176:292e308.
[18] Denis N, Dubois MR, Dube  R, Desrochers A. Blended power management
No conflict of interest exits in the submission of this manuscript. strategy using pattern recognition for a plug-in hybrid electric vehicle. IEEE
We would like to declare that the work described is original Trans Contr Syst Technol 2016;14(2):101e14.
[19] Kim N, Jeong J, Zheng C. Adaptive energy management strategy for plug-in
research that has not been published previously, and not under hybrid electric vehicles with Pontryagin’s minimum principle based on daily
consideration for publication elsewhere, in whole or in part. driving patterns. Int J Precis Eng Manuf-Green Technol 2019;6(3):539e48.
[20] Zhuang W, Zhang X, Peng H, Wang L. Simultaneous optimization of topology
and component sizes for double planetary gear hybrid powertrains. Energies
Acknowledgement 2016;9(6):411.
[21] Li L, Yang C, Zhang Y, Zhang L, Song J. Correctional DP-based energy man-
This work is support by the Natural Science Foundation of agement strategy of plug-in hybrid electric bus for city-bus route. IEEE Trans
Veh Technol 2014;64(7):2792e803.
Shandong Province, China (Grant No. ZR2014EL023, Data of work: [22] Guo H, Liang B, Guo H, Zhang K. A robust co-state predictive model for energy
2014.12 to 2019.5). The authors are also very grateful to the good management of plug-in hybrid electric bus. J Clean Prod 2020;250:119478.
suggestions of the reviewers and editors. [23] Rani P. A review of various KNN techniques. IJRASET 2017;5:1174e9.
[24] Taner T, Sivrioglu M, Topal H, Dalkılıç AS, Wongwises S. A model of energy
management analysis, case study of a sugar factory in Turkey. Sa dhana
References 2018;43(3):42.
[25] Robinette DA. DFSS approach to determine automatic transmission gearing
[1] Taner T. Optimisation processes of energy efficiency for a drying plant: a case content for powertrain-vehicle system integration. SAE Int J Passeng Cars-
of study for Turkey. Appl Therm Eng 2015;80:247e60. Mech Syst 2014;7(2014e01-1774):1138e54.