You are on page 1of 12

Discussion of Energy

Management Strategies for


Parallel Hybrid Electric
Vehicles
An In-Depth Search Through Various
Optimization Methods

Author: Karinne Bernanke


Submitted To: Dr. Abishek Gupta

Department of Electrical and Computer Engineering


The Ohio State University
December 9, 2021
Contents
1 Introduction 1

2 Problem Formulation 1

3 Optimization Method Discussion 3


3.1 Dynamic Programming . . . . . . . . . . . . . . . . . . . . . . 3
3.2 Pontryagin’s Minimization Principle . . . . . . . . . . . . . . . 4
3.3 Equivalent Fuel Consumption Minimization Strategy . . . . . 5

4 Conclusion 9

5 References 10

0
1 Introduction
There is a rising demand within the transportation industry to improve ve-
hicle energy efficiency. In the United States, the Corporate Average Fuel
Economy (CAFE) standards state that by 2025, the average fuel economy of
passenger cars should be 47.0 miles per gallon (mpg) [1]. The CAFE stan-
dards have modeled fuel economy of passenger vehicles to increase by close
to 1% every year from 2017 to 2025. Similar standards have been enacted
in other countries as well. For example India is working towards 49.3 mpg
by 2022 and the EU is working towards 64.4 mpg by 2025[1]. The race to
meet these standards has led to a shift towards more sustainable vehicles
that reduce fuel consumption, like Hybrid Electric Vehicles (HEVs), Battery
Electric Vehicles (BEVs), and Fuel Cell Vehicles (FCs). HEVs are the best
short term solution for increasing fuel economy and decreasing emissions.
HEVs include a battery and electric motor along with the combustion en-
gine that can power auxiliary loads and minimize idling losses[3]. This extra
source of energy allows for a downsized engine and the electric motor en-
ables the engine to operate at its peak efficiency. Finally, regeneration can
occur during cruising or braking where the energy of the vehicle can be cap-
tured and used to charge the battery[3]. This report will focus on designing
an optimal torque split strategy for an HEV in an effort to minimize the
overall fuel consumption of the vehicle, a problem referred to as the ’eco-
driving problem’. Section II will discuss the configuration of an HEV and
formulate the optimization problem to improve the fuel efficiency. Section III
will discuss the advantages and disadvantages of using different optimization
methods- including Dynamic Programming, Equivalent Fuel Consumption
Minimization (ECMS), and Pontryagin Minimization Principle (PMP)- for
implementation in an HEV.

2 Problem Formulation
An HEV architecture can either be parallel, series, or powersplit. A paral-
lel HEV splits mechanical and electrical power path which can either work
individually or together to propel the vehicle, yet the motor must be able
to provide power, forward to the wheels and backward to the battery. A
series HEV is powered only by the motor and the engine is used to charge
the battery, which brings in a lower efficiency connected solely to the motor

1
efficiency. A power split HEV combines the two configurations and allows
a vehicle to be both parallel and series with a well developed controller[5].
Below is the general power flow of a powersplit HEV.

Figure 1: Powersplit HEV Drivetrain Configuration [5]

In order to minimize fuel consumption, the power to the wheels from the
engine and from the motor must controlled. This torque split between the
engine and the motor can modulate how much fuel the engine uses, when the
motor will be used, and when regeneration will occur. This must be done
instantaneously as the drivers’ torque request changes throughout a drive
cycle and should be simple enough to implement in a controller, require a
small amount of memory, and be computationally cheap [7]. The factor that
determines the influence of the motor on the demanded torque is the state
of charge of the battery and the control input is the torque split between the
engine and the motor.
The problem will be formulated as a receding horizon optimal control
problem where the optimization is applied over “T” time steps. The state of
the control problem is
xt = [SOCt , vt ]

2
where xt ∈ R2 refers to each time step along the route, SOCt is the state
of charge at time step t, and vt is the vehicle velocity at time step t. The
controller input is
ut = [Teng,t , Tem,t ]
where xt ∈ Rnxm , Teng,t is the engine torque at time step t, and Tem,t is the
electric motor torque at time step t. Next, the cost function can be modeled
as
T −1
X

J (xt ) = minut CT (xt ) + CT (xt , ut )
t=0

where CT (xt ) is the terminal cost calculated by

CT (xt ) = (xref − xT )2

where xref is the reference SOC. CT (xt , ut ) is the running cost calculated by

CT (xt ) = mf (ut )

where mf is the fuel consumption as a function of the control variables, which


is the torque split. This optimal control problem has the following constraints

xt ϵ[xmin , xmax ]
min max
Teng,t ϵ[Teng,t , Teng,t ]
min max
Tem,t ϵ[Tem,t , Tem,t ]
x 1 = xN
where xmin , xmax is the minimum and maximum state of charge respectively,
min max
Teng,t , Teng,t is the minimum and and maximum engine torque respectively,
min max
Tem,t , Tem,t is the minimum and maximum torque for the electric motor re-
spectively, and N is the last time step in the route. The goal is to solve for
the optimal control sequence u∗t0 , u∗T that minimizes J ∗ .

3 Optimization Method Discussion


3.1 Dynamic Programming
The equation above is formulated for Dynamic Programming. DP is com-
monly used and can give sufficient conditions for the global optimal solution,

3
yet it relies on knowledge of the entire drive cycle before computing. There-
fore the method is carried out offline and is not implementable for instanta-
neous calculation of the global minimum fuel consumption.There have been
proposed variations of DP using a rule-based strategy as well as a way to
create a real-time controller based on results from an offline run [6].In this
case, the switching mode from a parallel configuration to a series configura-
tion is determined by the position of a clutch. The state remains the same
as the original problem formulation, yet the controls are the following:

uk = [Tem,1 , Tem,2 , Ck , Ek ]

where Tem,1 is the torque of the motor, Tem,2 is the torque of the generator,
Ck is whether the clutch is open or closed, and Ek is whether the engine is on
or off [6]. The clutch and engine control variable determine the mode of the
vehicle, and the electric motor torque determine the torque split of the two
electric machines. These two controller levels make up the rule based strategy
which is formulated by running many simulations on various drive cycles to
find patterns for each mode. In parallel mode, the torque of the engine can be
computed with the subtraction of motor torque from the torque demanded.
In series, one motor propels the vehicle and the other motor and engine work
to provide electrical power, therefore the engine is mostly able to operate
at optimal efficiency. Still with this improvement of DP formulation, the
computation must be done offline with a priori information and is therefore
not reasonable for implementation in a real time controller. In the instance
that higher connectivity was implemented in the HEV, time would need to
be added as a state. Therefore, DP is not scalable for electrification and
exponentially increases in computation time and expense with more states
and inputs.

3.2 Pontryagin’s Minimization Principle


Pontryagin’s Minimization Principle (PMP) finds the necessary conditions
for a global optimum solution in a constrained minimization problem [7].
PMP is faster than DP and adaptive versions can be implemented on line
in a vehicle using vehicle distance and average velocity as the only a priori
information, which can be found from a GPS when the driver sets a destina-
tion. The cost function is still the instantaneous fuel consumption rate, yet

4
is written in a different form [7]:
Z tf
J= ṁf (u(t))dt
t0

where the control variable is the power of the battery


u(t) = Pbatt
Having the power of the battery as the control is advantageous because the
engine torque and speed of the vehicle ultimately determine fuel consumption
and are luckily both related to the power of the battery. Therefore with PMP,
there are less states and less control variables in the problem formulation.
To solve the optimization problem with PMP, the Hamiltonian function is
constructed to be:
˙
H(SOC(t), u(t), λ(t)) = ṁf (u(t) + λ(t)SOC(t)
where the state is
˙ ∂H
SOC(t) = = f (SOC(t), Pbatt (t)
∂λ
and the co-state is
−∂H ˙
∂ SOC
λ̇(t) = = −λ
∂SOC ∂SOC
In the problem formulation, the co-state varies throughout the drive cycle,
when in fact the variation is almost negligible. This means that λ(0) =
λ(1 : N ) where N is the final time step. Because the co-state now does not
need to be recalculated at each time step, the PMP problem can be solved
online and implemented on a vehicle controller in real time, although with
sub-optimal results. Adaptive PMP methods follow the same formulation,
yet have different ways to optimally tune the co-state using variable that
are already known in the problem, making the calculation computationally
cheap.

3.3 Equivalent Fuel Consumption Minimization Strat-


egy
The Equivalent Fuel Consumption Minimization Strategy (ECMS) is an in-
stantaneous method derived from PMP that optimizes for fuel without veloc-
ity, taking only the assumed drive cycle into account[3]. In this method, the

5
fuel consumption is calculated as a sum of the actual fuel consumption and
the product of the fuel consumption and a penalty factor[4]. This penalty
factor or equivalent factor can be constant or variable, the latter is an indi-
cation an offshoot method of ECMS called Adaptive ECMS (A-ECMS). The
fuel consumption using ECMS is formulated below:
des

Pbatt,k (xk , uk )
ṁ = minṁf,k (xk , uk ) + λfpen,k (xk )
Qlhv
[8] where xk is the state which is the same (SOC and velocity) as the above
methods, uk is the control input, λ is the equivalence factor, fpen,k is the
des
penalty function,Pbatt,k is the desired power of the battery, and Qlhv is the
lower heating value of the fuel. The equivalence factor accounts for the
torque split of electrical and mechanical power and the penalty function is
associated with the SOC. To formulate an equivalence factor suitable for the
entire trip, drive cycle information must be known a priori and the equivalent
factor would then be calculated offline. In efforts to eliminate this need for a
priori information, A-ECMS changes the equivalence factor continuously with
variables already known or easily accessible, therefore if the SOC is close to
its upper bound or lower bound, an increased equivalence factor would cause
a very high instantaneous fuel consumption option that would avoid being a
global minimum. The equivalence factor could rely on engine efficiency of the
output torque curve, equivalent power to the wheels, SOC, or a combination
of different factors[3]. This is a very good method for implementing SOC
limits to mitigate battery degradation.
In one A-ECMS, the SOC is predicted, compared to a reference SOC,
and tuned using proportional control. The equivalence factor would be re-
calculated every T seconds using the following formula [9]:

λ(k + 1) = 0.5(λ(k) + λ(k − 1)) + Kp (SOC0 − SOCT )

where Kp is the proportional gain given by the feedback controller, SOC0 is


the reference SOC, and SOCT is the SOC at the current time step. λ(k + 1)
is the recalculated equivalence factor for the next time step at time T+1.
Finding the average between the equivalence factor in the current time step
and the previous time step gives the formula stability giving the vehicle more
information in case the vehicle is experiencing quick acceleration changes
or rapid regeneration. The equivalence factor is very sensitive, therefore
it is rare that the factor would drastically change between time steps, yet

6
the calculation of the equivalence factor in real time without future velocity
information is sub-optimal, therefore stability will improve the computations
immensely. This A-ECMS method, and generally other A-ECMS methods,
are computationally cheap, do not require any a priori information, and have
been implemented and proven to work in simulations and the industry[3].
Most of the work done on ECMS and A-ECMS to optimize fuel usage is
done for only powertrain optimization and does not take velocity planning
into account. A new method combining approximate DP and ECMS that
combines the advantages of each method, where the computation can be
performed online in real time and the equivalence factor can be computed
optimally. The method requires simplistic route information such as speed
limits and the slope of the road that can be easily acquired via GPS[8]. The
architecture for this method is shown in Figure 2 below.

Figure 2: DP-ECMS Eco-driving Architecture [8]

The general formulation of this method is related to DP, yet includes the
equivalence factor from ECMS and is outlined below:
T −1
X
Jt∗ (xt , λt ) = gT (xT ) + gT (xt , ut , λt )
t=1

where the state variable is


xt = [vt2 , SOCt ]
the square of the velocity vector points at the kinetic energy of the vehicle.
The control variable is
ut = Tptt
where Tptt is the powertrain torque at time t. First the ECMS portion of the
method determines a constant equivalence factor using GPS information and

7
calculates an optimal velocity as a function of the control input. The optimal
velocity is fed through a cost to go recursion that defines the cost function
over the entire route. The method starts at the terminal cost function.

JT (xT ) = gT (xT )

[8]
Jt (xt ) = minut (xt ) Jt+1 (ft (ut , vt∗ ◦ ut )) + gt (xt , ut , vt∗ ◦ ut )
[8] This second portion of the calculation finds the optimal torque split.
Finally, the product of the equivalence factor and penalty function may need
some tuning in order to be constant and optimal for the entire route, therefore
there is a simple tuning equation that outputs an optimal equivalence factor
and penalty function for the route.

λfpent (SOCt ) = λ0 + tan(−(SOCt − SOCdes )λ1 )

where λ0 and λ1 are constants. With this new method, optimal velocity
trajectory is calculated instantaneously, the equivalence factor is calculated
on line and tuned to fit the route, and the powertrain torque needed to
optimize fuel consumption is calculated close to optimally. The final results
for the DP-ECMS online approach versus the DP offline approach are plotted
below [8].

Figure 3: DP-ECMS and DP Pareto Curve Comparison [8]

8
Although the DP-ECMS method is suboptimal, it only varies from the
DP optimal results by 2% [8]. The DP-ECMS curve closely resembles the
DP optimal curve and with such a small error, this DP-ECMS method seems
to be the most promising method for implementing the eco-driving equation
in a real time controller.

4 Conclusion
The above discussion differentiating optimization methods for vehicle fuel
minimization deals with an HEV, yet such a method could be tweaked for
the controllers of other sustainable vehicles such as PHEVs. The largest dif-
ference between an HEV and a PHEV is that PHEVs would need the addition
of a switching mode. This switching mode would determine whether to turn
off the engine for full EV mode or turn on the engine for HEV mode[7]. Addi-
tionally, with recent advances in vehicle connectivity and autonomy, the con-
straint of needing an a priori drive cycle to compute offline optimal solutions
may become obsolete as instantaneous detailed road information becomes
available for the controller to use. For now, every optimizaton method that
can be computed instantaneously is sub-optimal, yet the DP-ECMS method
seems to be the closest to optimal with the integration of GPS provided sim-
ple road characteristics and velocity planning. As many existing methods
have explored powertrain optimality, there is a lack of methods involving ve-
locity optimization. Therefore with more research into velocity planning and
the hopeful advancements in vehicle connectivity, the eco-driving problem
can in the future be optimal and instantaneous.

9
5 References
[1] “Chart library: Passenger vehicle fuel economy,” The international Coun-
cil on Clean Transportation. [Online]. Available: https://theicct.org/chart-
library-passenger-vehicle.fuel-economy.
[2] J. Oncken and B. Chen, “Real-Time Model Predictive Powertrain Con-
trol for a Connected Plug-In Hybrid Electric Vehicle,” IEEE Transactions on
Vehicular Technology, Vol.69, No.8, 2020.
[3] Dr.Rizzoni’s lecture 16
[4] J. Wu, J. Ruan, N. Zhang, and P. Walker, ”An Optimized Real-Time En-
ergy Management Strategy for the Power-Split Hybrid Electric Vehicles,”IEEE
Transactions on Control Systems Technology No. 1063-6536, 2018.
[5] J. Liu, H. Peng, ”Modeling and Control of a Power-Split Hybrid Vehicle,”
IEEE Transactions on Control Systems Technology, Vol.16, No.6, 2008.
[6] D.Bianchi et. al.,”A Rule-Based Strategy for a Series/Parallel Hybrid
Electric Vehicle: An Approach Based on Dynamic Programming,” ASME
2010 Dynamic Systems and Control Conference, No. DSCC2010-6455, 2010.
[7] S. Onori, L. Tribioli,”Adaptive Pontryagin’s Minimum Principle Super-
visory Controller Design for the Plug-In Hybrid GM Volt,” Applied Energy,
No. 147, 2015.
[8] S. Deshpande, D. Jung, and M. Canova,”Integrated Approximate Dy-
namic Programming and Equivalent Consumption Minimization Strategy
for Eco-Driving in a Connected and Automated Vehicle,”preprint(under re-
view),2020.
[9] S. Onori, L. Serrao, G. Rizzoni,”Adaptive Equivalent Consumption Min-
imization Strategy for Hybrid Electric Vehicles,”ASME 2010 Dynamic Sys-
tems and Controls Conference, No.6433, 2010.

10

You might also like