Online EnerMana for HEV (IEEE Trans,Model,DP,Solution)

download Online EnerMana for HEV (IEEE Trans,Model,DP,Solution)

of 13

Transcript of Online EnerMana for HEV (IEEE Trans,Model,DP,Solution)

  • 8/2/2019 Online EnerMana for HEV (IEEE Trans,Model,DP,Solution)

    1/13

    3428 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 57, NO. 6, NOVEMBER 2008

    Online Energy Management forHybrid Electric Vehicles

    John T. B. A. Kessels, Member, IEEE, Michiel W. T. Koot,Paul P. J. van den Bosch, Member, IEEE, and Daniel B. Kok

    AbstractHybrid electric vehicles (HEVs) are equipped withmultiple power sources for improving the efficiency and perfor-mance of their power supply system. An energy management (EM)strategy is needed to optimize the internal power flows and satisfythe drivers power demand. To achieve maximum fuel profitsfrom EM, many solution methods have been presented. Optimalsolution methods are typically not feasible in an online applica-tion due to their computational demand and their need to havea priori knowledge about future vehicle power demand. In thispaper, an online EM strategy is presented with the ability to mimicthe optimal solution but without using a priori road information.

    Rather than solving a mathematical optimization problem, themethodology concentrates on a physical explanation about whento produce, consume, and store electric power. This immediatelyreveals the vehicle characteristics that are important for EM. Itis shown that this concept applies to many existing HEVs as wellas possible future vehicle configurations. Since the method onlyfocuses on typical vehicle characteristics, the underlying algorithmrequires minor computational effort and can be executed in realtime. Clear directions for online implementation are given in thispaper. A parallel HEV with a 5-kW integrated starter/generator(ISG) is selected to demonstrate the performance of the EM strat-egy. Simulation results indicate that the proposed EM strategyexhibits similar behavior as an optimal solution obtained fromdynamic programming. Profits in fuel economy primarily arisefrom engine stop/start and energy obtained during regenerative

    braking. This latter energy is preferably used for pure electricpropulsion where the internal combustion engine is switched off.

    Index TermsEnergy management (EM), fuel economy, hybridelectric vehicle (HEV), incremental fuel cost.

    I. INTRODUCTION

    OVER the years, the automotive industry has put much

    effort into developing vehicles that satisfy todays high

    standards on safety and comfort and simultaneously com-

    ply with strict environmental regulations. One of the leading

    Manuscript received October 22, 2007; revised January 6, 2008. First

    published March 5, 2008; current version published November 12, 2008. Thereview of this paper was coordinated by Mr. D. Diallo.

    J. T. B. A. Kessels was with the Control Systems Group, Department ofElectrical Engineering, Technische Universiteit Eindhoven, 5600 Eindhoven,The Netherlands. He is now with TNO Science and Industry, 5700 Helmond,The Netherlands (e-mail: [email protected]).

    M. W. T. Koot is with Vision Dynamics, 5652 Eindhoven, The Netherlands(e-mail: [email protected]).

    P. P. J. van den Bosch is with the Control Systems Group, Department ofElectrical Engineering, Technische Universiteit Eindhoven, 5600 Eindhoven,The Netherlands (e-mail: [email protected]).

    D. B. Kok is with the Micro Hybrid Systems and Energy ManagementGroup, Ford Dunton Technical Center, SS15 6EE Basildon, U.K. (e-mail:[email protected]).

    Color versions of one or more of the figures in this paper are available onlineat http://ieeexplore.ieee.org.

    Digital Object Identifier 10.1109/TVT.2008.919988

    technologies on the road to sustainable mobility is the hy-

    brid electric vehicle (HEV) [24]. By applying a secondary

    power source, these vehicles offer a significant improvement in

    fuel economy compared with vehicles with a traditional drive

    train.

    HEVs require an advanced energy management (EM) strat-

    egy to control the power of the primary and secondary power

    source [3]. Many solutions have already been presented in the

    past, covering heuristic approaches [1], [5] as well as advanced

    strategies based on optimization techniques [6], [8], [9], [12],[14], [19], [21], [22]. In particular, the optimization concept

    offers a well-defined solution to the stated EM problem. How-

    ever, online implementation is often not possible due to their

    computational demand. Another obstacle is the need for a priori

    knowledge in terms of exact predictions about the future driving

    cycle, see, e.g., [20]. To circumvent this problem, Guzzella and

    Sciarretta propose the Equivalent Consumption Minimization

    Strategy (ECMS) [8], which is directly derived from an optimal

    solution, but no prediction information is needed. Likewise,

    this concept is also demonstrated by Delprat et al. [6], em-

    phasizing the direct relation with optimal control. Nonetheless,

    owing to the mathematical problem formulation, the important

    vehicle characteristics for EM cannot easily be recognized inthe available numerical solution. As a result, difficulties appear

    when predicting the impact of new vehicle technologies on the

    performance of EM. In addition, the EM system itself remains

    a complex optimization algorithm, and its decision process

    cannot easily be decomposed into elementary decisions.

    Dual to the ECMS method, this paper presents an online

    EM strategy based on physical insight into the EM problem.

    Although the concept still originates from optimal control,

    a detailed analysis leads to a clear understanding of vehicle

    characteristics, which are dominant for EM. This yields direct

    information about when to produce, store, and consume electric

    power, as well as which components are responsible for theseactions. Rather than solving one complex optimization prob-

    lem, insight is generated into the decision process behind EM.

    Adequate rules are defined to cut down the system complexity

    of the EM strategy. This way, a real-time implementation is

    easily satisfied because of the low computational demand; the

    EM strategy only exploits typical vehicle characteristics that are

    relevant for EM.

    An early publication of the proposed EM strategy appeared

    in [12], but then, the research focused on vehicles with a

    traditional drive train. Subsequently, a proof of concept was

    published in [11], showing experimental results with a demon-

    strator vehicle on a roller dynamometer. This paper presents

    0018-9545/$25.00 2008 IEEE

  • 8/2/2019 Online EnerMana for HEV (IEEE Trans,Model,DP,Solution)

    2/13

    KESSELS et al.: ONLINE ENERGY MANAGEMENT FOR HYBRID ELECTRIC VEHICLES 3429

    Fig. 1. Three alternative vehicle topologies (arrows indicate nominal power flow). (a) S-HEV. (b) P-HEV. (c) S/P-HEV.

    a generalized concept for existing HEVs and possible future

    configurations. By means of three popular HEV topologies,

    i.e., a series (S), parallel (P), and series/parallel (S/P) HEV, the

    validity of the proposed concept will be demonstrated. It will

    be shown that the EM strategy achieves a performance close to

    the optimal strategy (based on exact prediction information) but

    without the need for knowledge about the future driving cycle.

    In addition, a formal strategy derivation is presented to point outthat the method is physically and mathematically sound. Clear

    algorithms for online implementation are given as well.

    This paper is organized as follows. An adequate vehicle

    model is presented in Section II. This model is used in

    Section III to explain the physical background of the EM

    decision process. This section also defines different operating

    modes that will be used by the EM algorithm. A formal problem

    definition for the EM strategy is given in Section IV, and

    the mathematical derivation of the strategy can be found in

    Section V. This latter section also presents an exact description

    for online implementation, including a method to estimate the

    driving characteristics online in the vehicle. The simulation re-sults focus on the P-HEV topology and are given in Section VI.

    Finally, conclusions can be found in Section VII.

    II. VEHICLE MODEL

    This paper focuses on three popular HEV topologies. First,

    the S-HEV does not have a mechanical connection between

    the internal combustion engine (ICE) and the wheels, and its

    model is shown in Fig. 1(a). Next, the P-HEV is demonstrated

    in Fig. 1(b). In P-HEV, the ICE and the electric machine can

    both give tractive force to the wheels, offering freedom for EM.

    The third vehicle configuration is the S/P-HEV, and its model is

    visualized in Fig. 1(c). The S/P-HEV has maximal freedom forEM due to its versatile power split device. In these three vehicle

    topologies, one can recognize similar components. These are

    discussed below.

    Rather than focusing on vehicle dynamics, the correspond-

    ing models concentrate on fueling characteristics and energy

    efficiency. As a result, only dynamic aspects necessary for EM

    are considered. A good example of EM including all dynamic

    equations can be found in [15].

    A. Fuel Tank(F)

    The fuel in the tank is the primary energy source in all vehicle

    configurations. In this paper, the chemical energy content of

    fuel is denoted by the lower heating value hf (in joules pergram).

    B. ICE

    In the literature, the fuel consumption of an ICE is usually

    shown by a static nonlinear map, describing the relation be-

    tween the crankshaft torque m (in newton meter), the enginespeed (in radians per second), and the fuel rate f(m, )(in grams per second). Here, it is preferred to define the fuel

    consumption as a function of engine power Pm (in watts) andspeed as

    fuelrate = f(m, ) = f(Pm|) (in grams per second).(1)

    The notation with the conditional operator | is introduced toemphasize the dependency of Pm on . The advantage of thisnotation becomes clear when explaining the control objective

    of an EM strategy.

    Contrary to P-HEV and S/P-HEV, the S-HEV has no me-

    chanical connection between the ICE and the wheels. There-fore, ICE is allowed to run in several operating points (m, )

  • 8/2/2019 Online EnerMana for HEV (IEEE Trans,Model,DP,Solution)

    3/13

    3430 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 57, NO. 6, NOVEMBER 2008

    and still deliver equal power Pm = m. From an efficiencypoint of view, it will be beneficial to operate the engine only

    in those operating points that entail minimum fuel use for a

    given power request. The set of operating points that fulfill this

    criterion is called the e-line (economy line), and for a given

    power request Pm, the corresponding e-line operating point is

    uniquely defined as

    Pe-linem = arg min(m,)Q(Pm)

    f(m, ) (in watts) (2)

    with

    Q(Pm) = {(m, ) | Pm = m} . (3)

    As a result, (1) simplifies for the S-HEV into the following

    description:

    fuelrate = f

    Pe-linem

    (in grams per second). (4)

    C. MG

    An HEV also uses an electric machine for vehicle propulsion.

    Depending on the power flow direction, this machine operates

    in either motor mode or generator mode. Therefore, the electric

    machine will be addressed in this paper as motor/generator

    (MG) as well. It is assumed that appropriate facilities exist

    in the electric power net to actively control the power of all

    electrical components.

    At its mechanical side, the signals of interest are the speed

    MG (in radians per second), the torque MG (in newton me-ters), and the mechanical power Pem = MG MG (in watts).

    The electric side only considers the electric power Pe (in watts)without notion of the voltage and current. When operating ingenerator mode, Pe and Pem take a positive value. On the otherhand, these signals become negative valued during motor mode.

    For each mode, the corresponding efficiency is defined as

    Motor mode:

    mm(MG, MG) =PemPe

    =MG MG

    Pe[] (5)

    Generator mode:

    gm(MG, MG) =Pe

    Pem=

    PeMG MG

    []. (6)

    It is assumed that both efficiencies can be measured and that

    they remain fixed for a given sampling interval. This way, the

    model of the electric machine can be expressed by one single

    equation

    Pem = max

    mmPe,

    1

    gmPe

    (in watts). (7)

    Similar to [20], the model (7) introduces linear losses for the

    MG, although it does not provide a good description of the

    friction losses at zero power. In [13], it is shown how these

    losses can be included as a speed-dependent offset term, but this

    extension is not taken into account here. The power limitations

    of the electric machine are defined at its mechanical side as

    Pemmin Pem Pemmax. (8)

    TABLE IPARAMETER LIST FOR DRIVE TRAIN MODEL

    For the S-HEV, there is another electric machine present, but

    this device only operates in generator mode. The model for this

    generator G is adopted from (7) as

    Pg = gmPgm (in watts). (9)

    D. Drive Train (D)

    The power demand Pd from the drive train is calculated witha quasi-static backward-calculating vehicle model. Depending

    on the selected vehicle topology, this model encompasses dif-

    ferent vehicle components. The vehicle chassis and the final

    drive are present in all vehicle topologies. For a given ve-

    hicle speed v(t) (in meters per second) and road slope (t)(in radians), the corresponding wheel force Fw(t) (in newtons)is equal to

    Fw(t) = mv(t) +1

    2CdAdv(t)

    2

    + mg [sin((t)) + Cr cos((t))] . (10)

    A description of all the parameters is given in Table I. Further-more, this table also provides the parameter values for a mid-

    sized vehicle as used in the simulation environment.

    The drive train torque d(t) (in newton meters) and therotational speed d(t) (in radians per second) in S-HEV andS/P-HEV follow by taking into account the wheel radius wr(in meters) and the final drive ratio fr[] as

    S-HEV and S/P-HEV:

    d(t) =

    frwr

    v(t)d(t) =

    wrfr

    Fw(t). (11)

    Different from these vehicle configurations, the drive train of

    the P-HEV incorporates a gearbox. To that end, the gear ratio

    gr(t)[] is added to the previous equations

    P-HEV:

    d(t) =

    frwr

    gr(t)v(t)

    d(t) =wrfr

    1gr(t)

    Fw(t). (12)

    Finally, the mechanical power request Pd (in watts) from thedrive train is equal to

    Pd(t) = d(t)d(t) = v(t)Fw(t). (13)

    E. Power Split (PS)

    P-HEV and S/P-HEV have more than one power sourceavailable for vehicle propulsion. Therefore, they make use of a

  • 8/2/2019 Online EnerMana for HEV (IEEE Trans,Model,DP,Solution)

    4/13

    KESSELS et al.: ONLINE ENERGY MANAGEMENT FOR HYBRID ELECTRIC VEHICLES 3431

    Fig. 2. Description of planetary gear train.

    mechanical power split to connect the individual power sources

    to the drive train. The power split is assumed to have no energy

    losses and provides the following power balance:

    P-HEV: Pm = Pd + Pem (in watts) (14)

    S/P-HEV: Pm = Pd + Pem1 + Pem2 (in watts). (15)

    In P-HEV, the power split establishes a direct connectionbetween ICE and the drive train, whereas the connection of

    the MG includes an additional gear. On the other hand, the

    power split of the S/P-HEV is built up with an electronically

    controlled Continuously Variable Transmission (eCVT) using

    a planetary gear train (see Fig. 2). Recently, various eCVT

    configurations have been developed for future HEVs, all using

    similar hardware technologies [17]. As a result, they can still

    rely on the same concept for EM.

    The configuration selected here is derived from a Toyota

    Prius, where the ICE is directly coupled to the carrier gear, see,

    e.g., [16]. The first MG (MG1) is linked to the sun gear, whereas

    the second MG (MG2) is linked to the ring gear. In addition, the

    drive train is linked to the ring gear. Therefore, the power at the

    ring gear Pr (in watts) is aggregated from the drive train powerand the power through MG2 as

    Pr = Pd + Pem2. (16)

    The kinematics of a planetary gear train define the relation

    between the rotational speeds of the ring gear r, sun gear s,and carrier gear c. Using the radii of the ring gear Rr (inmeters) and the sun gear Rs (in meters), the rotational speedsin the system are described by

    s + rZ = c(Z+ 1) (17)

    with the basic gear ratio Z[] defined as

    Z =RrRs

    . (18)

    By assuming that there is no inertia present in the planetary gear

    train, and when there is no energy dissipation, the law of power

    conservation describes the static relation between all powers

    applied to the eCVT. Given the speed and torque , as definedin Fig. 2, the following relation emerges:

    Psun + Pc + Pr = 0 (19)

    ss + cc + rr = 0. (20)

    In addition, the torques acting on the sun, carrier, and ring gear

    have to be balanced in all situations as

    s + c + r = 0. (21)

    The relations (17), (20), and (21) provide a complete descrip-

    tion of the planetary gear train. Nevertheless, for practical

    situations, it is convenient to replace (20) and (21) with analternative representation. A description with the basic gear

    ratio follows from the substitution of (17) and (21) into (20).

    This yields two linear equations, i.e.,

    r = Zs (22)

    c = (Z+ 1)s. (23)

    Note that these relations are also found by considering two

    special situations for (17) and (20), i.e., c = 0, and r = 0.

    F. Battery (B)

    The battery model is constructed from two elements, i.e., an

    energy storage buffer and a static efficiency map. The efficiency

    map expresses the relation between the power Pb (in watts) atthe battery terminals and the net internal power Ps (in watts).For convenience, the efficiency map is modeled through an ef-

    ficiency number 0 bat 1. Physically, one can make a dis-tinction between the losses that occur during charging (Ps 0)and discharging (Ps < 0). Denoting these losses through theparameters + 1 and 1, respectively, with bat=

    +

    results in the following parameterization:

    Pb = maxPs, 1+

    Ps . (24)A simple integrator is used to model the energy storage buffer

    of the battery, i.e.,

    Es(t) = Es(0) +

    t0

    Ps(t)dt (in joules). (25)

    The theoretical energy capacity Ecap (in joules) is used tonormalize the stored energy in the battery, and its energy status

    is defined by the state of energy (SOE) as

    SOE(t) =

    Es(t)

    Ecap 100 [%]. (26)

    G. EL

    The electric load (EL) denotes the electric power demand of

    all auxiliaries present in the vehicle. It is assumed that this load

    is power based, and its power demand is represented by PL(in watts).

    H. Electric Power Net

    The power net accumulates the power flow from all electric

    devices. For simplicity, it is assumed that the required powerconverters (including their energy losses) are incorporated into

  • 8/2/2019 Online EnerMana for HEV (IEEE Trans,Model,DP,Solution)

    5/13

    3432 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 57, NO. 6, NOVEMBER 2008

    Fig. 3. Fuel consumption ICE matches piecewise linear function (ICE speed

    constant).

    the electric machine and ELs. Furthermore, no losses are as-

    sumed in the wiring system. This way, the power net reduces to

    the following power balance:

    S-HEV: Pe + Pg = Pb + PL (27)

    P-HEV: Pe = Pb + PL (28)

    S/P-HEV: Pe1 + Pe2 = Pb + PL. (29)

    III. STRATEGY OUTLINE

    From an EM strategy, it is expected that it supplies the total

    (electric) power demand with the highest efficiency and, hence,minimum fuel usage. By controlling the power through the MG,

    it decides upon efficient operating points of many vehicle com-

    ponents. The storage capacity of the battery takes care of the un-

    balance between the generated power and the requested power.

    To obtain insight into the decision process of an EM strategy,

    emphasis will be put on the fuel map of the ICE. Therefore,

    the fuel function f(Pm|) in (1) from a 2.0 gasoline ICEis shown in Fig. 3. In contrast to existing literature, the fuel

    map is deliberately presented as a function of engine power

    Pm using separate curves for different engine speeds . Thisway, it reveals three areas where there exists an almost linear

    (affine) relation between f and Pm. Typically, the slope ofthis curve remains almost constant within each area and equals

    the following definition:

    (Pm|) =f(P|)

    P

    P=Pm

    (in grams per joule). (30)

    Intuitively, it is clear that expresses the additional fuel massflow to produce a small amount of mechanical power given a

    certain power level Pm. Hence, expresses the incremental fuelcost. With the help of this definition, each fuel curve can be

    approximated as a piecewise affine function

    f(Pm|) fi() +i()Pm

    for i = 0, 1, . . . , N f; Pm i (31)

    and with Nf + 1 line segments

    i = {Pm|Pmi Pm Pmi+1}. (32)

    For this particular engine, it is sufficient to take Nf = 2

    f(Pm|)

    0, ifP

    m < Pm1f1 + 1Pm, ifPm1 Pm < Pm2f2 + 2(Pm Pm2), ifPm Pm2

    (33)

    where Pm1 = f1/1 < 0 denotes the fuel cutoff point, andPm2 > 0 indicates a point close to the maximum engine power.Note that all parameters fj , j , and Pmj depend on , although

    the influence on j is limited (jhf = 2.4 4.0[]). Thisimmediately puts a limitation on the possible benefits of any

    EM strategy, since profits in fuel economy arise in the fol-

    lowing ways.

    1) Producing less power with the ICE and thereby requestingenergy from the battery are favorable for large values

    ofj .2) Producing extra power with the engine and storing the

    surplus energy in the battery are favorable for small

    values ofj .3) For a constant ICE power Pm, a reduced engine speed

    decreases the fuel use f1. Moreover, if the ICE is turnedoff, the battery becomes the primary power source, and

    although recharging requires additional fuel, this can still

    economically be attractive since f1 has been eliminated.

    Now suppose that 1, 2, and f1 of ICE are constant. In

    this case, the EM strategy can only benefit from the followingmechanisms to improve the vehicles fuel economy.

    Regenerative braking (R): In a traditional vehicle, the me-

    chanical friction brakes become active when the

    driver wants to decelerate the vehicle. From an en-

    ergy point of view, this is not an economic solu-

    tion, since all kinetic energy is wasted into heat.

    To recover the energy that comes available during

    braking periods, one can operate the electric machine

    in generator mode. This way, the electric machine

    absorbs power from the drive train and stores it in

    the battery. In terms of fuel economy, this is the

    most economical way to charge the battery, sinceit requires no additional fuel. Nevertheless, for a

    correct driveability of the vehicle, it is not allowed

    to recuperate all braking energy solely from the front

    or rear wheels. Therefore, the potential available

    energy reduces for a two-wheel drive. In addition, the

    power ratings of the electric machine and the charge

    acceptance of the battery limit the actual amount of

    stored energy.

    Engine stop/start: In a traditional vehicle, the ICE is kept

    idle when the driver requests no propulsion power.

    A simple method to save fuel is to stop the engine at

    those moments (e.g., waiting for traffic lights). For

    driver comfort, a quick vehicle response is preferredwhen the driver wants to move on, but given the

  • 8/2/2019 Online EnerMana for HEV (IEEE Trans,Model,DP,Solution)

    6/13

    KESSELS et al.: ONLINE ENERGY MANAGEMENT FOR HYBRID ELECTRIC VEHICLES 3433

    fact that a powerful electric machine is present in an

    HEV, smooth engine cranking is no problem.

    Motor only (MO): Except for exclusively applying engine

    stop/start at moments when the vehicle stands still

    (or during R), an HEV can also apply electric drive

    and still keep the ICE off. This situation is referred to

    as the MO mode. In particular, the selection of MOduring vehicle launch turns out to be an excellent

    way to utilize the energy from R.

    All the mentioned methods take place with the ICE turned

    off. Furthermore, their economic profits rely on the absolute

    (constant) value of 1, 2, and f1 rather than variations overtime. This is different when the ICE is turned on. In that case,

    the EM strategy considers the changes in and discriminatesbetween the following modes of operation.

    Baseline (BL): By definition, the BL situation implies that

    the battery is not used so that Pb = 0. Physically, thismeans that the mechanical power delivered by the

    ICE continuously matches the vehicle power demandPd plus PL.

    Motor assist (MA): In the situation of MA, ICE produces

    less mechanical power than requested by the drive

    train and the auxiliary loads. The battery supplies the

    remaining part so that Pb < 0. MA should be appliedat moments when is relative high. In (33), thiscorresponds to the area where 2 becomes active.Typically, these areas are characterized by the fact

    that ICE operates near its maximum power limits

    (see Fig. 3).

    Charging mode (C): In C, ICE produces additional mechan-

    ical power such that the electric machine can charge

    the battery: Pb > 0. Different than in R, this mode re-quires additional fuel consumption, so C is preferred

    when is small. According to (33), this will be in thearea where 1 holds.

    The given strategy outline focuses on HEVs with the ICE as

    the primary energy source. Future vehicle configurations possi-

    bly appear with alternative power sources, e.g., a fuel cell stack

    [3], [24]. Nevertheless, all new technologies currently known

    still exhibit the input/output description from (31) so that a

    similar reasoning for EM is still valid. If variations in i and/orfi are present in (31), the EM system is able to improve theenergy efficiency of the vehicle. The EM strategy decides which

    mode should become active and what power is needed for all

    components. This decision process relies on an optimization

    algorithm and will be introduced in the next section.

    IV. PROBLEM FORMULATION

    From a control point of view, the EM strategy decides upon

    two variables, i.e., the electric machine power Pem and theengine-running signal S {0, 1}. By manipulating these twovariables, the EM strategy is aiming at minimum fuel costs.

    Although an HEV is a complex dynamical system, it is as-

    sumed that there exists a static relation between the manipulated

    variables and the momentary fuel consumption. Therefore, itis possible to formulate the fuel costs by means of algebraic

    relations. The optimal EM strategy follows from the solution of

    a standard optimization problem

    minx

    J(x) subject to G(x) 0 (34)

    where J(x) is the objective function, and G(x) expresses theconstraints on the decision variable x. The function J repre-sents the cumulative fuel use of ICE over an arbitrary driving

    cycle with time length te as

    S-HEV: J(Pem, S) =

    te0

    f

    Pe-linem (t)

    S(t)dt (35)

    P-HEV: J(Pem, S) =

    te0

    f(Pm(t)|(t)) S(t)dt (36)

    S/P-HEV: J(Pem1, Pem2, S) =

    te0

    f(Pm(t)|(t)) S(t)dt. (37)

    Note that the decision variable x covers two different signaltypes, i.e., the power of the electric machine Pem(t) and theengine-running signal S(t). Whereas Pem can take positive andnegative values, the signal S switches between two discretevalues {0, 1}.

    Evaluation of the cost functions (35)(37) is done in the

    following way. By selecting a predefined driving cycle, the

    required drive train power Pd is calculated from (13). Fur-thermore, it is assumed that the EL power PL is also given.Depending on the selected HEV topology, the vehicle model in

    Section II describes the exact relation between Pd and PL andthe corresponding fuel cost f, provided that the decision vari-ables Pem and Sare known. How to calculate these variables isexplained in Section V.

    Except for concentrating on the overall fuel consumption

    of the vehicle, the EM system also has the responsibility of

    satisfying several constrains. For example, the operating range

    of the engine, electric machine, and battery are limited in power.

    Therefore, inequality constraints are introduced to limit the

    minimum and maximum power flow through these compo-

    nents, e.g.,

    Pmmin(t) Pm(t) Pmmax(t) t [0, te] (38)

    Pemmin(t) Pem(t) Pemmax(t) t [0, te] (39)

    Pbmin(t) Pb(t) Pbmax(t) t [0, te]. (40)

    According to the standard optimization problem (34), these

    constraints have to be rewritten in terms of G(x) 0. Again,the relations from the vehicle model in Section II are used.

    In addition to the constraints mentioned above, the EM

    system also has the responsibility of guaranteeing a charge-

    sustaining vehicle. A charge-sustaining strategy claims that the

    battery satisfies a minimum SOElevel at the end of the drivingcycle. This is achieved by including an end-point constraint on

    the energy level of the battery, e.g.,

    Es(te) Es ref Es(0) +

    te0

    Ps(t)dt Es ref (41)

  • 8/2/2019 Online EnerMana for HEV (IEEE Trans,Model,DP,Solution)

    7/13

    3434 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 57, NO. 6, NOVEMBER 2008

    where Es ref is an arbitrarily selected reference value thatshould be satisfied at t = te, e.g., Es ref = Es(0).

    V. ONLINE STRATEGY

    Many methods are known for calculating the optimal solution

    of (34) subject to (38)(41). Probably the most celebratedmethod is dynamic programming (DP) [2]. This optimization

    technique provides an optimal solution to the described EM

    problem, with its limited resolution enforced by the selected

    grid accuracy.

    Typically, EM strategies using DP provide a benchmark

    regarding the maximum potential fuel savings. A real-time

    EM strategy using DP is hampered by two restrictions. First,

    a priori knowledge is needed about the vehicle power demand

    (i.e., Pd(t) and PL(t) need to be known in advance alongthe entire driving cycle). Second, calculations from DP over

    a lengthy driving cycle are computationally demanding. To

    overcome both problems, this paper proposes a novel online

    implementation for EM, which originates from physical insight

    into the EM problem. The underlying strategy is strongly

    related to the ECMS method from [8], although the mathe-

    matical backgrounds originate from [23]. The method is able

    to mimic the optimal solution from DP without the need for

    having prediction information about the future driving cycle.

    Moreover, it does not suffer from complex calculations and is

    generally applicable to present and future HEV configurations.

    The next section describes the EM strategy for S-HEV

    and P-HEV. Details regarding the S/P-HEV are given in

    Section V-B.

    A. EM for S-HEV and P-HEV

    To make the EM strategy generally applicable, the opti-

    mization problem (34) is rewritten with an alternative decision

    variable such that x covers the net battery power Ps, togetherwith the engine-running signal S. This is possible since thebattery efficiency model (24) and the description of the electric

    power net (27)(29) provide a unique relation between Ps andthe controlled variable Pem. For S-HEV and P-HEV, this meansthat (35) and (36) translate into

    J(Ps, S) =

    te

    0(Ps, S|Pd, PL, )dt (42)

    where

    S-HEV: (Ps, S|Pd, PL, )

    = f

    Pe-linem

    S (in grams per second) (43)

    P-HEV: (Ps, S|Pd, PL, )

    = f(Pm|)S (in grams per second). (44)

    According to this description, two subproblems are defined

    since the engine-running signal S takes only two values {0, 1}.Furthermore, the end-point constraint (41) is reformulated

    such that the integral constraint is avoided in the optimizationproblem.

    First, consider the situation with S = 1. This allows the HEVto operate in modes BL, MA, and C. Without the inequality

    constraints G(x) 0, the optimization problem from (34) re-duces to

    minPs

    te

    0

    (Ps|Pd, PL, )dt. (45)

    Next, the end-point constraint (41) will be added. In the case

    where the energy level of the battery at t = te has to be equalto the initial starting value Es(0), the inequality constraintchanges into an equality constraint

    Es(0) +

    te0

    Ps(t)dt = Es(0)

    te0

    Ps(t)dt = 0. (46)

    A new problem definition is formulated from (45) together

    with the equality constraint (46). For convenience, it is written

    in discrete time, although the sampling interval T has beenomitted, i.e.,

    minPs

    Npk=1

    (Ps(k)|Pd(k), PL(k), (k))

    subject to

    Npk=1

    Ps(k) = 0. (47)

    Finding a solution for this optimization problem can be done

    by incorporating the equality constraint into the Lagrangian

    function using a Lagrange multiplier . A similar approach has

    been followed in [8]. The following Lagrangian L is defined:

    L (Ps(1), . . . , P s(Np), )

    =

    Npk=1

    (Ps(k)|Pd(k), PL(k), (k))

    Npk=1

    Ps(k). (48)

    Physically, this new objective function makes sense, because it

    adds the energy exchange with the battery to the fuel consump-

    tion of the ICE. The quantity represents the correspondingfuel cost when energy is stored or taken from the battery. It is

    clear that there exists a strong relation between this quantity

    and the definition of , as given in (30). The minimum value

    for L(Ps(1), . . . , P s(Np), ) can be calculated by solving thefollowing set of equations:

    L (Ps(1), . . . , P s(Np), )

    Ps(k)

    = (Ps(k)|Pd(k), PL(k), (k))

    Ps(k) = 0 (49)

    for 1 k Np, and

    L (Ps(1), . . . , P s(Np), )

    =

    Npk=1

    Ps(k) = 0. (50)

    To guarantee a global optimal solution,(Ps(k)|Pd(k), PL(k), (k)) has to be a convex function at

  • 8/2/2019 Online EnerMana for HEV (IEEE Trans,Model,DP,Solution)

    8/13

    KESSELS et al.: ONLINE ENERGY MANAGEMENT FOR HYBRID ELECTRIC VEHICLES 3435

    each time instant k. This observation is justified by the shapeof the fuel curves (31) in combination with the model for the

    MG (7) and the battery model (24). This convexity implies that

    there exists one unique solution (Ps (1), . . . , P s (Np),

    ) tothe set ofNp + 1 equations in (49) and (50).

    The solution is calculated with information from the entire

    driving cycle, i.e., Pd(k), PL(k), and (k) are assumed to beknown at each time instant k = 1, . . . , N p. However, if

    is

    known, all Np equations from (49) are entirely decoupled. Thismeans that calculating Ps (k) can be done by solving (49) withPd(k), PL(k), and (k) only known at time k. For a rigorousproof, see [23]. As a result, the optimal solution Ps (k) from(47) is also found by minimizing the given criterion at each

    time instant k, i.e.,

    Ps = arg minPs

    {(Ps|Pd, PL, ) Ps} . (51)

    The solution presented so far only considers the cost criterion

    in combination with one equality constraint. Returning to the

    original problem, one has to guarantee that the inequality

    constraints from (38)(40) are not violated. The vehicle model

    can be used to rewrite all constrains in terms of Ps. Next, theseconstraints are combined into one new constraint on Ps with theupper and lower bound Psmax and Psmin, respectively. Finally,a feasible solution for the original problem is obtained through

    saturation ofPs with these new boundaries, i.e.,

    PS1s = sat [Ps ]PsmaxPsmin

    =min(max(Ps , Psmin) , Psmax) . (52)

    The corresponding fuel cost for the power setpoint PS1s is

    fS1

    =

    PS1

    s |Pd, PL,

    PS1

    s . (53)

    It is important to notice that these extra constraints can be added

    without violating the optimality of the calculated solution. After

    all, the function (Ps|Pd, PL, ) was assumed to be a convexfunction. By applying saturation on Ps, this function still pre-serves convexity, and hence, the optimal solution is uniquely

    defined. It is clear that an optimal value for Ps will sometimesappear at its boundary. Therefore, the value for might bedifferent for the situation where saturation is included. How to

    find a suitable value for will be discussed in Section V-C.Now consider the situation with S = 0. Therefore, engine

    is turned off, and MO becomes active. A necessary condition

    for MO is that the electric motor is able to handle the powerrequest Pd and that the battery can supply the power from theelectric motor and PL. Both requirements are satisfied when thefollowing inequalities hold:

    Pd Pemmin Pdmm

    PL Pbmin. (54)

    During braking phases, an engine stop is also preferred,

    whereas the MG stores energy from R in the battery. All vehicle

    configurations considered in this paper are front-wheel driven,

    but for a correct driveability of the vehicle, recuperation of

    free kinetic energy requires interaction between the front and

    rear wheels. Although the distribution of the front and rearwheel braking forces is beyond the scope of this paper, it is

    important to realize that the EM system cannot recover all avail-

    able kinetic energy solely at the front wheels. The parameters

    that determine the ideal braking force distribution are, among

    others, the desired vehicle deceleration and the cargo weight.

    A good introduction to the braking force distribution in HEVs

    is given in [7]. Without loss of generality, it is assumed here

    that the braking force at the front wheels is fixed at 60%, i.e.,rb = 0.60[], whereas the rear wheels take care of the remain-ing part. Now the power from the electric machine during MO

    is equal to

    PMOe = min

    1

    mmPMOem , gmP

    MOem

    (55)

    with PMOem = Pd during vehicle propulsion (Pd 0), andPMOem = rb Pd is available during R (Pd < 0)

    PMOem = sat [max(rb, Pd, Pd)]PemmaxPemmin

    . (56)

    Although the ICE is turned off, it should be noted that P-HEV

    suffers from additional losses at the power split due to the

    engine drag torque. In this paper, these losses are neglected.

    Similar to the electric machine, the battery is also limited

    in power. In particular, during R, the charge acceptance of the

    battery is important, and incorporating these limitations yields

    the following net battery power:

    PMOs = min

    1

    PMOb ,

    +PMOb

    (57)

    with PMOb = sat PMOe PLPbmaxPbmin . (58)ICE is not running during MO, and therefore, the momen-

    tary fuel consumption is zero. The additional fuel costs for

    recharging the battery afterward are estimated using the optimal

    Lagrange multiplier as

    fS0 = PMOs . (59)

    Finally, to select MO instead of the other modes, the following

    condition must be satisfied [calculated from (53) and (59)]:

    fS0 < fS1. (60)

    The final control law for Ps comprises the combination of allpossible situations. This means that MO becomes active when

    both (54) and (60) are satisfied. Discrimination between the

    other modes is done in (51) and (52). This can be numerically

    done by taking a grid for Ps and evaluating the criterion in (51)for all the grid points. Altogether, this yields the following EM

    strategy law for Ps:

    PEMs (Pd, PL, , ) =

    PMOs , if(54) (60)PS1s , elsewhere.

    (61)

    The actual power setpoint for the MG is calculated from theS-HEV or P-HEV model equations in Section II.

  • 8/2/2019 Online EnerMana for HEV (IEEE Trans,Model,DP,Solution)

    9/13

    3436 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 57, NO. 6, NOVEMBER 2008

    TABLE IICALCULATIONS FOR S/P-HEV TO DETERMINE (c, c)

    B. EM for S/P-HEV

    The S/P-HEV topology incorporates two electric machines

    (i.e., MG1 and MG2), which both require a suitable control law

    for EM. Together with the engine stop/start feature, the EM

    strategy decides upon three variables, i.e., Pem1, Pem2, andS. Similar to the configurations from the previous section, Sonly takes two discrete values {0, 1}; therefore, the objective

    function in (37) can be rewritten into

    J(Pem1, Pem2, S) =

    te0

    (Pem1, Pem2, S|d, d, PL)dt (62)

    where

    (Pem1, Pem2, S|d, d, PL) = f(c, c)S [g/s]. (63)

    Contrary to the other HEV configurations, it is preferred to cal-

    culate the fuel use of S/P-HEV in (63) with separate speed and

    torque signals (c, c) rather than power Pd. This is enforced bythe mechanical power split, which uses a planetary gear train for

    connecting MG1, MG2, and ICE to the drive train. According

    to the model description in Section II, ICE is connected to the

    ring gear, and its operating point (c, c) is uniquely defined bythe three decision variables together with the power request of

    the driving cycle. The necessary steps to calculate c and c aresummarized in Table II. Note that the calculations in Table II

    are only valid when d = 0. However, this restriction has nofurther consequences because it is not economically attractive

    to switch on the ICE and charge the battery when the vehicle

    stands still. This observation is justified by the fact that ICE has

    a relatively large fuel offset when it runs idle, and there exist

    enough other driving moments where charging of the battery

    can be done more efficiently. Hence, d = 0 implies S = 0.In addition, different from S-HEV and P-HEV, S/P-HEV

    requires individual constraints to limit the speed and torque

    range of ICE. A single constraint for the power from ICE

    would be too conservative. Given the minimum and maximum

    torque line of the engine, it follows that this constraint is time

    dependent. A similar reasoning also holds for the engine speed,

    leading to the following constraints:

    min(t) c(t) max(t) (64)

    mmin(t) c(t) mmax(t) t [0, te]. (65)

    Altogether, S/P-HEV offers significantly more freedom for

    EM, but it is still possible to apply a similar optimization ap-proach as described in Section V-A. This means that the discrete

    variable S will be eliminated by creating two subproblems.First, the situation with S = 1 is considered so that ICE ispermanently on. For a given drive train request (d, d), theoptimal power setpoints for MG1 and MG2 are calculated by

    PS1em1, P

    S1em2

    = arg min(Pem1,Pem2)Q(d,d)

    {(Pem1, Pem2|d, d, PL)

    Ps} (66)

    where the set Q only covers the feasible setpoints for MG1 andMG2 to satisfy the constraints (38)(40), (64), and (65) as

    Q(d, d) =

    (Pem1, Pem2)|Pem1min Pem1 Pem1max Pem2min Pem2 Pem2max Pbmin Pe1 + Pe2 PL Pbmax mmin c mmax min c max

    . (67)

    The signals c and c are calculated according to the steps listed

    in Table II. The model from MG1 and MG2 in (7) is usedto calculate the electric power Pe1 and Pe2, respectively. Thenet battery power Ps is calculated from the battery model (24).Eventually, the result from (66) is used to estimate the minimum

    expected fuel cost when the engine is turned on, i.e.,

    fS1 =

    PS1em1, PS1em2|d, d, PL

    PS1s . (68)

    Now consider the situation with S = 0 so that the engine isturned off. In S/P-HEV, MG2 takes care of the drive train power,

    whereas the battery must be able to handle the corresponding

    power request and PL. Apparently, this situation turns out tobe similar to S-HEV and P-HEV so that (54)(60) also apply

    here, taking into account that the electric machine correspondsto MG2; therefore, Pem = Pem2.

    Altogether, the EM system defines the control law for Pem1and Pem2. The combination of the individual modes for S = 0and S = 1 leads to the following strategy description:

    PEMem1 (Pd, PL, , ) =

    0, if(54) (60)PS1em1, elsewhere

    (69)

    PEMem2 (Pd, PL, , ) =

    PMOem2, if(54) (60)PS1em2, elsewhere.

    (70)

    The number of MG pairs influences the overall complexity

    of the EM system. Nevertheless, the calculations for S/P-HEVhave a similar complexity as for S-HEV and P-HEV. As a

    result, the numerical calculations are not ready to explode

    for more advanced vehicle configurations. In addition, future

    vehicle configurations, possibly equipped with even more MG

    pairs, will be able to keep the numerical complexity of the EM

    strategy limited. For example, an HEV constructed with in-

    wheel motors for its power train gives rise to an EM system with

    multiple MG pairs. Without considering additional driveability

    issues for safe and stable vehicle operation, the presented EM

    algorithm can still be applied.

    All these strategies have in common that they only rely on

    the present state of the vehicle including . This means that

    knowledge about the future driving cycle is not needed. Thecalculation of will be explained below.

  • 8/2/2019 Online EnerMana for HEV (IEEE Trans,Model,DP,Solution)

    10/13

    KESSELS et al.: ONLINE ENERGY MANAGEMENT FOR HYBRID ELECTRIC VEHICLES 3437

    Fig. 4. Feedback diagram for estimating .

    C. Adaptive EM Strategy

    Every driving cycle requires a different to achieve apreferred energy level in the battery at the end of the driving

    cycle. This corresponds to the situation where the energyneed for Pd and PL is balanced with the energy for batterycharging. It is important to notice that power-requesting modes

    such as MO and MA take less energy from the battery for

    higher values of . Conversely, mode C stores more energyin the battery when increases. As a result, there exists aunique solution for where the energy from R and C equalsthe energy for MO and MA (including the battery losses for

    temporarily storing energy). Then, it follows that the SOEreturns to its initial value at the end of the driving cycle. Thisobservation is also recognized by Delprat et al. [6]. By means

    of simulations, Section VI shows that the EM strategy with

    mimics the optimal solution from DP, provided that there are

    no restrictions in the buffer size from the battery. It follows

    that exactly knowing the future driving cycle is not a strict

    requirement for excellent performance (see also [18]).

    To calculate , accurate information about the vehicle powerdemand is required along the entire known driving cycle. This

    is inconvenient for online implementation because it requires,

    among others, an exact prediction of the vehicle speed, includ-

    ing external disturbances such as wind or road inclination. In-

    stead of focusing on how to obtain this prediction information,one can also rely on the driving behavior from the past. Assum-

    ing that the speed profile from the past provides a good repre-

    sentation of the future driving cycle, there are several methods

    to estimate an appropriate value (see [4], [12], and [19]).The method that has been selected here results in an adaptive

    strategy, as presented in [12]. The basic idea is that the SOEof the battery indicates whether is estimated correctly or not.In the case has been selected too small, the battery becomesdepleted in the end. Conversely, when is selected too high, thebattery becomes fully charged. From a control point of view,

    this corresponds to a leveling control problem where SOE

    should be kept near a nominal value SOEref. A proportionalintegral (PI) controller with a rather small bandwidth fulfills

    this requirement. The block diagram is shown in Fig. 4, with equal to

    (t) = 0 + KPe(t) + KI

    t0

    e()d (71)

    with 0 as an initial guess. Selecting the parameters KP and KIfor a small closed-loop bandwidth allows for a tracking error

    between the actual SOE and SOEref. This freedom is neededby the EM system to temporarily store/retrieve energy into/from

    the battery. On the other hand, a charge-sustaining strategy mustkeep the SOE sufficiently close to SOEref to prevent battery

    depletion or overcharging. Therefore, the bandwidth of the PI

    controller should not be too small.

    A suitable bandwidth should be selected according to the

    power spectrum of upcoming driving cycles. Only the fre-

    quency components outside the bandwidth of the PI controller

    contribute to the fuel profits from the EM system. Frequencies

    inside the bandwidth are counteracted by the PI controllerto establish a charge-sustaining strategy. Consequently, these

    frequency components will not provide any profits in fuel

    economy. Appropriately tuning of the parameters KP and KIis described and analyzed in [10] and [13].

    VI. SIMULATION RESULTS

    Although the EM strategy applies to all three HEV configu-

    rations, the results shown in this section focus on the situation

    with P-HEV. Compared with a vehicle with a traditional drive

    train, only minor changes are necessary to create a P-HEV.

    As will be shown in this section, a P-HEV with a small

    hybridization factor offers significant benefits in fuel economy.

    Examples with S-HEV or S/P-HEV can be found in [10].

    The P-HEV configuration represents a mild HEV equipped

    with a 2.0 SI engine and an ISG. Simulations are done for theNew European Driving Cycle (NEDC) with the corresponding

    drive train power Pd calculated from the backward-facing vehi-cle model in Section II. The clutch is located between the power

    split and the drive train.

    The battery and the ISG can both handle electric power up to

    5 kW. During motor mode and generator mode, the efficiency

    of the ISG is equal to mm = gm = 0.8[]. Therefore, itsmechanical power Pem ranges between 4000 and 6250 W.

    The effective battery capacity is assumed to be Ecap = 2.0 MJ,and its efficiency is equal to bat = 0.85[]. These batteryparameters represent a 36-V battery with a rated capacity of

    Cbat = 30 Ah. Only 50% of its capacity is effectively availablein terms of energy, i.e.,

    Ecap 0.5 3600 Cbat Unombat

    = 0.5 3600 30 36 = 2.0 MJ. (72)

    For this vehicle configuration, seven alternative EM strate-

    gies are evaluated. All strategies start at SOE = 75%. Thefirst simulation (Sim1) only applies mode BL with the engine

    always running and the battery power permanently zero, so thatS = 1, and Pb = 0. The second simulation (Sim2) evaluates acontrol strategy obtained from DP. The DP strategy provides an

    optimal strategy within the selected grid accuracy and serves

    as a benchmark. The third simulation (Sim3) uses the EMstrategy from (61) with taken constant. A careful selectionof guarantees that the SOE at the end of the driving cycleequals 75%. It is important to recognize that Sim3 is equal tothe situation with no feedback (i.e., zero bandwidth) for the

    PI controller. Consequently, Sim3 achieves maximum profitsin fuel economy, whereas restrictions for a charge-sustaining

    strategy are lacking.

    Sim4 applies the same strategy as Sim3, but now, is

    estimated with the PI controller from Section V-C. Finally,the last three simulations (Sim5, Sim6, and Sim7) are added

  • 8/2/2019 Online EnerMana for HEV (IEEE Trans,Model,DP,Solution)

    11/13

    3438 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 57, NO. 6, NOVEMBER 2008

    TABLE IIIFUEL CONSUMPTION OF P-HEV ON NEDC

    to give insight into the performance of the EM strategy and

    how it depends on the individual operating modes. All three

    strategies are derived from Sim3, but by excluding specificmodes, the effects of MA, MO, R, and C are investigated. Sim5demonstrates the added value of MA in combination with C by

    excluding R and MO. Sim6 shows the large impact of R onfuel economy by using its recovered energy solely for MO. In

    this simulation, MA and C are excluded from the EM strategy.

    In addition to R, battery charging with C is also an effective

    method to gather energy for MO. In Sim7, the benefits fromMO in combination with C are elucidated.

    For all the strategies, the corresponding fuel consumption

    is shown in Table III. It should be noted that Sim4 endsthe driving cycle with a slightly different SOE level, and itsfuel consumption has been corrected for this with the average

    value of , see, e.g., [21] for a similar approach. An importantobservation is that Sim3 and Sim4 achieve almost similarimprovements in fuel consumption, although Sim4 does notrely on knowledge about the future driving cycle. The reason

    why Sim3 performs slightly better than Sim2 is because DP isrestricted to the accuracy of the grid, whereas Sim3 does not

    suffer from this limitation.Comparing Sim5 with Sim3, it follows that MA only gains

    1.4% fuel reduction. This relative small benefit is explained in

    two ways. First, the engine map only contains limited variations

    for (see Fig. 3). In particular, in the area where it deliverslow power, the fuel curves are straight and in parallel. As a

    result, C will not be profitable to generate energy for MA

    due to the losses in the battery. Second, the NEDC driving

    cycle has a moderate power request so that large engine powers

    (where the engine map exhibits more variations for ) arebarely observed. Since MA is only profitable if variations for

    exist, its contribution will be small for this particular driving

    cycle.The majority of the fuel savings for EM arise by applying

    MO. By switching off ICE, MO eliminates the fuel offset f1 atmoments when no (or possibly a small amount of) propulsion

    power is needed. During MO, the requested energy is supplied

    by the battery, and to recharge the battery, two mechanisms are

    possible, i.e., R and C. According to these two methods, the

    corresponding fuel savings are illustrated with two simulations

    Sim6 and Sim7, respectively. Sim6 only receives energy fromR to compensate for the energy taken from the battery during

    MO. This yields a fuel reduction of 15.8%. Furthermore, C is

    also a viable method to charge the battery and extend the poten-

    tial benefits for MO. Without R, Sim7 achieves a fuel reduction

    of 17.5%, which is more than Sim6. The reason why Sim7gains more profits than Sim6 is because the energy recovered

    by R is limited through the braking periods prescribed by the

    driving cycle. The energy from C is only limited by power

    constraints from components, and significantly more energy is

    available to apply MO.

    The strategies Sim5, Sim6, and Sim7 are all a subset fromSim3. Nevertheless, it can be seen that the contribution of each

    individual strategy adds up to a higher fuel reduction than theperformance shown by Sim3. This is because the individualmodes of the EM strategy do not independently operate off of

    each other. For example, MO receives energy from R and C.

    The energy from R is produced without extra fuel cost, but the

    total available energy is limited. Therefore, the EM strategy

    decides upon the extra energy from C, but the additional fuel

    benefits will be less.

    From a computational point of view, only limited calcula-

    tions are required for Sim4. This way, an online implementa-tion in the vehicle is always guaranteed. A proof of concept

    is already given in [11], which presents roller dynamometer

    experiments with a similar EM strategy for vehicles with a

    traditional drive train. With a series-production vehicle, a fuel

    reduction of more than 2.5% is achieved without major changes

    to the vehicle hardware.

    Considering the small difference in fuel economy between

    Sim4 and Sim2, it can be concluded that the proposed onlinestrategy is very promising. In addition, the costs for the selected

    vehicle hardware yield a good tradeoff between investment and

    reward. Hybridization with a 5-kW ISG yields approximately

    25% fuel reduction.

    Nevertheless, the simulations do not include a penalty for

    switching the engine on and off. Furthermore, the free energy

    from R during vehicle deceleration reduces due to friction

    losses in the combustion engine, and these losses are not wellcovered in the simulation model. These drawbacks come with

    all the strategies, and therefore, it is reasonable to believe that

    the relative difference between all simulations remains equal.

    The control sequences for Sim2, Sim3, and Sim4 are shownin Fig. 5. From the SOE trajectory, it is clear that Sim2 andSim3 exhibit similar behavior. Because these strategies utilizea priori knowledge from the driving cycle, they allow larger

    deviations for SOE and still return to 75% at the end of thecycle. This is different for Sim4. By construction, this strategykeeps SOE near its reference value, but the main controlactions from Sim2 and Sim3 are still visible. Another point of

    attention is the control setpoint Pem. Due to the piecewise linearbehavior of the engine map, the desired control actions areoften found in extremes, sometimes giving rise to nonsmooth

    switching behavior.

    For this particular driving cycle, the periods where ICE is

    switched off are interrupted by short periods with multiple

    stop/start events (see Fig. 5). To include a penalty for switching

    ICE on and off, an extra inequality constraint can be added

    to (54), which becomes true only if ICE has been on for a

    sufficiently long time. For example, a counter can be used

    to measure the time elapsed since the last cranking event.

    This way, a tradeoff emerges between extra fuel savings for

    switching the ICE off versus frequent cranking of the ICE. The

    engine emissions during cranking, as well as the driveabilityaspects, have to be investigated to optimize this tradeoff.

  • 8/2/2019 Online EnerMana for HEV (IEEE Trans,Model,DP,Solution)

    12/13

    KESSELS et al.: ONLINE ENERGY MANAGEMENT FOR HYBRID ELECTRIC VEHICLES 3439

    Fig. 5. Simulation results for Sim2, Sim3, and Sim4 on NEDC driving cycle.

    More generally, the results presented so far have paid little

    attention to the driveability aspects. However, from an EM

    system, it is expected that it certainly does not harm safety

    aspects or driver comfort. To satisfy all these functionalities,

    the proposed EM system should become part of a supervisory

    control algorithm, whereas dedicated controllers will be needed

    to take care of local control actions. For example, smoothtransitions are desired when changing the power demand be-

    tween ICE and EM. In addition, during braking periods, a local

    controller has to take care of the power distribution between the

    front and real wheels. In addition to mechanical requirements,

    the electric power net also requires local power converters to

    take care of the electric power flow in an efficient way. In

    addition, intelligent strategies for charging and discharging the

    battery are foreseen such that excessive battery wear can be

    prevented. Finally, the EM system should focus on more aspects

    than fuel economy alone (e.g., exhaust gas emissions during a

    cold and warm engine start). Ultimately, an integrated power

    train control is foreseen, where the EM system operates at a

    supervisory level.

    VII. CONCLUSION

    This paper has presented an online concept for EM in

    various HEV configurations. This concept provides not only

    a mathematical solution to the EM problem but the physical

    explanation behind an EM strategy as well. This way, two

    important vehicle characteristics, which directly determine the

    potential fuel benefits of any EM strategy, are recognized,

    i.e., the slope of the fuel map with respect to engine power

    and the speed-dependent fuel use when the engine runs idle.

    Not only existing HEVs but also future vehicle configurationssatisfy the proposed EM concept if their primary power source

    incorporates sufficient variations for both of the aforementioned

    characteristics.

    A practical solution has been presented on how to implement

    this strategy concept in each vehicle configuration. Owing to its

    low computational demand, an online vehicle implementation

    is guaranteed. The charge-sustaining requirement is satisfied by

    means of a feedback mechanism, comparing the actual energyin the battery with a preferred reference value. There is no need

    for prediction information or complex estimators of the vehicle

    status.

    Simulations with a mild P-HEV have demonstrated cor-

    rect control actions for the EM strategy. It turns out that

    the results in fuel economy come very close to the optimal

    strategy calculated with DP. This result has led to the obser-

    vation that adding prediction information to the EM strategy

    yields limited extra profits in fuel economy. Furthermore, it

    is concluded that most fuel benefits originate from engine

    stop/start during MO mode in combination with free energy

    from R. Battery charging is preferably done with energy from

    R, but regular battery charging during driving further extendsthe fuel profits from MO mode. It is concluded that the

    low hybridization factor has an excellent return-on-investment,

    with benefits in fuel economy of up to 25%. The fuel prof-

    its from MA are only a few percent for the studied vehicle

    configuration.

    REFERENCES

    [1] B. M. Baumann, G. Washington, B. C. Glenn, and G. Rizzoni, Mecha-tronic design and control of hybrid electric vehicles, IEEE/ASME Trans.

    Mechatron., vol. 5, no. 1, pp. 5872, Mar. 2000.[2] D. P. Bertsekas, Dynamic Programming and Optimal Control, 2nd ed.

    Belmont, MA: Athena Scientific, 2000.[3] C. C. Chan, The state of the art of electric, hybrid, and fuel cell vehicles,Proc. IEEE, vol. 95, no. 4, pp. 704718, Apr. 2007.

  • 8/2/2019 Online EnerMana for HEV (IEEE Trans,Model,DP,Solution)

    13/13

    3440 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 57, NO. 6, NOVEMBER 2008

    [4] J.-S. Chen and M. Salman, Learning energy management strategy forhybrid electric vehicles, in Proc. IEEE VPPC, Chicago, IL, Sep. 2005,pp. 6873.

    [5] S. R. Cikanek and K. E. Bailey, Regenerative braking system for ahybrid electric vehicle, in Proc. Amer. Control Conf., Anchorage, AK,May 2002, vol. 4, pp. 31293134.

    [6] S. Delprat, J. Lauber, T. M. Guerra, and J. Rimaux, Control of a parallelhybrid powertrain: Optimal control, IEEE Trans. Veh. Technol., vol. 53,

    no. 3, pp. 872881, May 2004.[7] Y. Gao and M. Ehsani, Electronic braking system of EV and HEVIntegration of regenerative braking, automatic braking force control andABS, presented at the SAE Future Transportation Technol. Conf.,Costa Mesa, CA, Aug. 2001, SAE Paper 2001-01-2478.

    [8] L. Guzzella and A. Sciarretta, Vehicle Propulsion SystemsIntroductionto Modeling and Optimization. Berlin, Germany: Springer-Verlag, 2005.

    [9] T. Hofman, M. Steinbuch, R. Van Druten, and A. Serrarens, Rule-basedenergy management strategies for hybrid vehicles, Int. J. Elect. HybridVeh., vol. 1, no. 1, pp. 7194, Jul. 2007.

    [10] J. T. B. A. Kessels, Energy management for automotive powernets, Ph.D. dissertation, Dept. Elect. Eng., Technische Univ. Eindhoven,Eindhoven, The Netherlands, 2007.

    [11] J. T. B. A. Kessels, M. Koot, B. de Jager, P. P. J. van den Bosch,N. P. I. Aneke, and D. B. Kok, Energy management for the electricpowernet in vehicles with a conventional drivetrain, IEEE Trans. ControlSyst. Technol., vol. 15, no. 3, pp. 494505, May 2007.

    [12] M. Koot, J. T. B. A. Kessels, B. de Jager, W. P. M. H. Heemels,P. P. J. van den Bosch, and M. Steinbuch, Energy management strategiesfor vehicular electric power systems, IEEE Trans. Veh. Technol., vol. 54,no. 3, pp. 771782, May 2005.

    [13] M. W. T. Koot, Energy management for vehicular electric power sys-tems, Ph.D. dissertation, Dept. Mech. Eng., Technische Univ. Eindhoven,Eindhoven, The Netherlands, 2006.

    [14] C.-C. Lin, H. Peng, J. W. Grizzle, and J.-M. Kang, Power managementstrategy for a parallel hybrid electric truck, IEEE Trans. Control Syst.Technol., vol. 11, no. 6, pp. 839849, Nov. 2003.

    [15] J. Liu and H. Peng, Control optimization for a power-split hybridvehicle, in Proc. Amer. Control Conf., Minneapolis, MN, Jun. 2006,pp. 466471.

    [16] J. Liu, H. Peng, and Z. Filipi, Modeling and analysis of the Toyota hybridsystem, in Proc. Int. Conf. Advanced Intell. Mechatron., Monterey, CA,Jul. 2005, pp. 134139.

    [17] J. M. Miller, Hybrid electric vehicle propulsion system architectures ofthe e-CVT type, IEEE Trans. Power Electron., vol. 21, no. 3, pp. 756767, May 2006.

    [18] C. Musardo, S. Benedetto, S. Bittanti, Y. Guezennec, L. Guzzella, andG. Rizzoni, An adaptive algorithm for hybrid electric vehicles en-ergy management, in Proc. FISITA World Automotive Conf., Barcelona,Spain, May 2004.

    [19] C. Musardo, G. Rizzoni, andB. Staccia, A-ECMS: An adaptive algorithmfor hybrid electric vehicle energy management, in Proc. Joint 44th IEEEConf. Decision Control, Eur. Control Conf., Seville, Spain, Dec. 2005,pp. 18161823.

    [20] P. Pisu andG. Rizzoni, A comparative study of supervisorycontrol strate-gies for hybrid electric vehicles, IEEE Trans. Control Syst. Technol.,vol. 15, no. 3, pp. 506518, May 2007.

    [21] A. Sciarretta, M. Back, and L. Guzzella, Optimal control of parallelhybrid electric vehicles, IEEE Trans. Control Syst. Technol., vol. 12,no. 3, pp. 352362, May 2004.

    [22] G. Steinmaurer and L. del Re, Optimal energy management for mildhybrid operation of vehicles with an integrated starter generator, pre-sented at the SAE World Congr., Detroit, MI, Apr. 2005, SAE Paper 2005-01-0280.

    [23] P. P. J. van denBosch andF. A. Lootsma, Scheduling of power generationvia large-scale nonlinear optimization, J. Optim. Theory Appl., vol. 55,no. 2, pp. 313326, Nov. 1987.

    [24] J. van Mierlo, G. Maggetto, and P. Lataire, Which energy source forroad transport in the future? A comparison of battery, hybrid and fuelcell vehicles, Energy Convers. Manag., vol. 47, no. 17, pp. 27482760,Oct. 2006.

    John T. B. A. Kessels (S04M08) received theB.S. degree in electrical engineering from FontysHogescholen, Eindhoven, The Netherlands, in 2000and the M.S. and Ph.D. degrees in electrical engi-neering from Technische Universiteit Eindhoven in2003 and 2007, respectively.

    After working as a Postdoctoral Researcher withTechnische Universiteit Eindhoven, in 2007, he

    joined the Automotive Group of TNO Science andIndustry, Helmond, The Netherlands. His researchinterests include energy and emission management

    for light-duty and heavy-duty vehicles, with emphasis on integrated powertraincontrol.

    Michiel W. T. Koot was born in Sittard,The Netherlands, in 1977. He received the M.S.and Ph.D. degrees in mechanical engineering fromTechnische Universiteit Eindhoven, Eindhoven,The Netherlands, in 2001 and 2006, respectively.

    He is currently with Vision Dynamics, Eindhoven.His research interests include dynamics, control, andoptimization applied to mechanical systems and au-tomotive vehicles.

    Paul P. J. van den Bosch (M84) received the M.S.degree (cum laude) in electrical engineering andthe Ph.D. degree in optimization of electric energysystems from Delft University of Technology, Delft,The Netherlands.

    After graduating, he was with the Control SystemsGroup, Delft University of Technology, where hewas appointed Full Professor in control engineeringin 1988. In 1993, he was appointed Full Profes-sor with the Control Systems Group, Departmentof Electrical Engineering, Technische Universiteit

    Eindhoven, Eindhoven, The Netherlands. In 2004, he was a part-time Pro-

    fessor with the Department of Biomedical Engineering. He is currently withTechnische Universiteit Eindhoven. His research interests concern modeling,optimization, and control of dynamicalsystems. In his career, he has cooperatedon many research projects with industry, including automotive applications,large-scale electric systems, advanced electromechanical actuators, and, re-cently, embedded systems and biomedical modeling. In addition to his scientificactivities, he has received several prizes for his educational activities and severalpatents. He has served on the editorial boards of journals (Journal A) andconference committees.

    Daniel B. Kok received the B.S. degree from HogereTechnische School of Automotive Engineering,Apeldoorn, The Netherlands, in 1991 and the M.S.and Ph.D. degrees in mechanical engineering fromTechnische Universiteit Eindhoven, Eindhoven, The

    Netherlands, in 1994 and 1999, respectively. HisPh.D. dissertation was entitled Design Optimisationof a Flywheel Hybrid Vehicle.

    In 1994, he was a Research Associate withTechnische Universiteit Eindhoven. From 1999 to2000, he was a Project Leader on hybrid electric

    vehicles with TNO Automotive, Delft, The Netherlands. In 2000, he was withthe Ford Research Center, Aachen, Germany. Since July 2005, he has been aTechnology Manager with the Micro Hybrid Systems and Energy ManagementGroup, Ford Dunton Technical Center, Basildon, U.K., where he is also aTechnical Leader.