Battery Fast Charging Strategy Based on Model Predictive Control

Department

Jingyu Yan , Guoqing Xu , Huihuan Qian ,Yangsheng Xu

of Mechanical and Automation Engineering, The Chinese University of Hong Kong, Shatin, Hong Kong Shenzhen Institutes of Advanced Technology, Shenzhen, Guangdong, China Email: {jyyan,gqxu,hhqian,ysxu}@mae.cuhk.edu.hk method has simple circuit structure and very low cost so that it has been adopted in most electronic products for many years. However, the charging time is extremely long due to the very small current and it usually works overnight. To reduce the charging time, an easy method is to enlarge the constant current (CC). Charging with 1C current can fully ll an empty battery in 1 hour. However, this method requires a very precise state of charge (SoC) estimator to determine whether the battery is full charged and stop charging process in time. Another disadvantage is that CC charging can not avoid the overvoltage to battery. A battery can be simply modeled by an open circuit voltage (OCV) source and a seriesconnected resistor. The voltage on resistor during CC charging will cause the terminal voltage always higher than OCV and nally lead to overvoltage when OCV approaches to its full value. Taking Li-ion batter as an example, the overvoltage during charging will degrade the crystallographic structure of cathode and cause oxidative decomposition of the electrolytic solvents [2]. To overcome the disadvantages of CC charging, the constant current constant voltage (CCCV) method has been proposed. In the following CV period, after charging voltage reaches a predetermined value in CC period, charging voltage is xed to cutoff value and charging current will be automatically reduced along with the increase of SoC. However, the CV period usually triples the charging time of CC period to ll the left 20%-25% capacity and largely extends the overall charging period [3]. Various combinations of CC and CV periods are also proposed to improve the charging performance, such as CCCVCV [4], (CCCV)2 [5] and so forth. In recent year, microprocessor control units (MCUs) are applied in battery management system and allow the monitoring and estimation of kinds of battery states, based on which intelligent methods can be implemented to control the charging process. The rst kind of intelligent charger is to control charging process based on the detection of tuning points. The tuning points are selected as the threshold points, stationary points and inection points of battery voltage, temperature and lapsed time [6] [7] [8]. These tuning points based methods only change the charging behaviors at tuning points and the periods between tuning points are still in open-loop control. Fuzzy control has been also utilized to solve charging control problem. In [3], the fuzzy controller replaces the usual CV mode in CCCV charging method. Neural network and genetic

AbstractBattery fast charging is a crucial issue in both research and application to realize and promote the mass commercialization of electric vehicles, especially pure electric vehicles. However, due to the strong nonlinear properties of batteries, the charging process should take into consideration various factors such as state of charge (SoC), temperature, and charging current, so as to assure the safety, reduce charging time, and enhance charging efciency. In this paper, we propose a fast charging strategy under the model predictive control framework. Two models are employed to predict SoC and temperature under a sequence of future charging currents. SoC predictor is based on RC equivalent circuit and temperature predictor is based on thermal conduction and convection. The prediction of battery future states allows optimization of the control sequence, with the objectives to follow a predetermined SoC trajectory and to minimize battery temperature rising. Genetic algorithm is introduced to solve the constrained multi-objective optimization problem. The results using Advisor platform demonstrate the availability and efcacy of the proposed framework and prove that it has the ability to reduce charging time and heat generation simultaneously.

I. I NTRODUCTION Battery, as a movable energy storage system, has been broadly used in kinds of electronic devices for many years, such as mobile, shaver, walkman, laptop and so forth. Recently, to improve the energy efciency and reduce emissions, it was also applied to drive vehicles, giving the birth of electric vehicles (EVs). The undoubted trend that battery will be more powerful and more frequently operated raises stricter requirements to battery management methodologies. Charging control is one of the most signicant issues in battery management system, aiming at feeding external energy into batteries by a fast, safe, and efcient way. The fast charing is helpful to reduce the out-of-service time and promote the commercialization of EVs. The safe charging not only assure the safety of users by avoiding battery burning and explosion during charging process, but also prolong the battery life by keeping the overcharge and overheating damages away. The efcient charging can convert the electric energy from charger to electrochemical energy stored in battery as much as possible and enhance energy efciency. A. Literature Review The charging methods have been studied since the birth of rechargeable batteries. The earliest and simplest method is to charge using a constant trick current (CTC) [1]. This

978-1-4244-3574-6/10/$25.00 2010 IEEE

Fig. 1.

Fast charging control framework based on model predictive control.

algorithm are also introduced to design membership functions and inference rules of fuzzy controller [9] [10] [11]. Optimization method, including ant colony algorithm and evolutionary algorithm, is also introduced to optimize a best multi-stage CC charging prole [12] [13]. The optimization aims at nding a best charging prole covering the whole charging process and does not realtime regulate charging current according to instant system measurements. Therefore, its performance seriously depends on the accuracy of the model used in optimization and its robustness is weak, suffered from system noises and timevariant properties of battery. Grey-predicted control system is employed to increase the charging speed without taking temperature rising into consideration [14]. In addition, only one-step prediction is utilized in this study. The smallest predictive horizon weakens the prediction of future system behavior. B. Overview of Proposed Charging Controller In recent years, many efforts have been devoted to establish battery models and some of them are demonstrated to be accurate and efcient to model the behaviors of batteries [15] [16] [17] [18] [19]. These models make it possible to apply model predictive control (MPC) to manage battery charging process. MPC is a typical advanced control method and widely utilized in many elds [20] [21]. Its principle is to predict the system responses to possible future control inputs through system models and nd the best future control sequence which optimizes the user-dened objective function. Only the rst step of the control sequence is implemented, then the system states are sampled again and the calculations are repeated at the next control time. The prediction horizon keeps being shifted forward. Since receding horizon strategy updates prediction based on instant measurements of system inputs and outputs at each control time, its robustness has been demonstrated to be strong [22]. Depending on the linearity of system models and constraints, MPC is categorized as linear MPC and nonlinear MPC. For linear MPC, the best control

sequence can be analytically obtained by solving Diophantine equations. However, when it comes to the nonlinear MPC, the best control sequence can only be numerically calculated under an optimization framework in general [23]. As shown in Fig.1, the fast charging control framework proposed in this work consists of the following components: 1) a SoC predictor, predicting the SoC of battery when feeded by a sequence of future charging current; 2) a temperature predictor, predicting the future battery temperature under the sequence of charging current; 3) a tness evaluator, evaluating the performance of the sequence of charging current; 4) an optimizer, nding the best sequence of charging current using genetic algorithm (GA). In section 2, we rstly illustrate the battery SoC and temperature models. To overcome the model inaccuracy and enhance the robustness of the predictor, a H lter is applied to predict SoC. In section 3, we introduce the basic principle of MPC and formulate the control problem as an optimization problem. The tness evaluation function has been proposed, II. S O C AND T EMPERATURE P REDICTORS To utilize predictive control, battery models are necessary to predict future states exerted by a sequence of future charging current. The concerned battery states in this study are SoC and temperature. The former one indicates the charging speed and whether overcharge damage is possible. The latter one represents the safety in charging process and the amount of lost energy converted to heat. Charging control basically is to (1) maximize charging speed, (2) minimize the rising of temperature and (3) strictly restrict temperature into a safe range. A. RC model for SoC prediction From the denition of SoC, it can be simply calculated by counting the output/input charges, known as columbic counting method. However, this method essentially provides an open-loop estimation of SoC, without self-correcting ability

B. Temperature Model

Vo i
Fig. 2.

Ro Rs
Cs

Rb
Cb

Battery RC Model.

A simple, single-node lumped-parameter thermal model has been established in [26]. It models the thermal process in three stages. In the rst stage, the heat is generated in the battery by Joule effect. In the second, heat of battery conducts and convects to its surrounding air. Last, the surrounding air exchanges heat with the ambient. In RC model, the heat generation is expressed by
2 2 2 (k )Ro (k ) + Is (k )Rs (k ) + Ib (k )Rb (k ) Qg (k ) = Ts Io

to accumulated errors [24]. Model based SoC estimation methods are well studied in recent years and have been veried to outperform columbic counting method. RC model is an appropriate and applicable model and has been applied in the well-know electric vehicle simulator Advisor [25]. As shown in Fig.2, RC model consists of a bulk capacitor Cb and a surface capacitor Cs , which simulate energy storage and dynamic property of the battery respectively. Output resistance Ro , surface resistance Rs and bulk resistance Rb are used to model the internal resistance of battery. Since battery is a typical time-variant plant, these model parameters are also variables depending on SoC and temperature. By denoting the time index as k and selecting state vector as the voltages of bulk and surface capacitors x(k ) = [Vb (k ), Vs (k )] , system input as bus current uk = Ik , output as terminal voltage y (k ) = Vo (k ), and sampling time as Ts , discrete time-variant state space equations of RC model is expressed by x(k + 1) = A(k )x(k ) + B (k )u(k ) y (k ) = C (k )x(k ) + D(k )u(k ) where A(k ) = B (k ) = C (k ) = D(k ) =
Ts Ts Cb (k)(Rb (k)+Rs (k)) Cb (k)(Rb (k)+Rs (k)) Ts s 1 Cs (k)(RbT Cs (k)(Rb (k)+Rs (k)) (k)+Rs (k)) Rs (k)Ts Cb (k)(Rb (k)+Rs (k)) Rb (k)Ts Cs (k)(Rb (k)+Rs (k)) Rs (k) Rb (k) Rb (k)+Rs (k) Rb (k)+Rs (k) Rb (k)Rs (k) Ro (k ) R b (k )+Rs (k )

(6)

where Io (k ) = I (k ) Vs (k)+Rs (k)I (k) Ib (k ) = Vb (k) Rs (k)+Rb (k) Is (k ) = Ib (k ) I (k ) (7)

Meanwhile, the passing heat from battery to surrounding air is expressed by Qp (k ) = T (k 1) Tair (k 1) Ref f (8)

where the Tair is the effective temperature of surrounding air and Ref f is the effective thermal resistance. Based on the assumption that that 50% of the heat from the battery is spent to warm the air, Tair is expressed by Tair (k 1) = Tamb 0.5 Qp (k 1) m air Cair (9)

(1)

(2) SoC in the RC model is estimated using the voltages of the two capacitors based on the relationship between SoC and OCV. Since Cb represents the bulk energy in the battery, it contributes the majority of SoC, as expressed in below: SoC (k ) = where SoCCb (k ) = FOCV SoC (Vb (k )) SoCCs (k ) = FOCV SoC (Vs (k )) (4) (5) 1 20SoCCb (k ) + SoCCs (k ) 21 (3)

where Tamb is the ambient temperature, Qp (k 1) is the passing heat in previous step, m air is the airow rate, and Cair is its heat capacity. Strictly speaking, Ref f is a variable depending on the thermal control method. For example, it will become smaller if the cooling fans are open. However, in this study, we suppose the charging process is conducted in the environment only with natural convection and no thermal control device is equipped. In this case, Ref f is xed as a constant with the value calculated by Ref f = t 1 + hA kA (10)

where h is the heat transfer coefcient in natural environment, A is the total module surface area exposed to air, t is the thickness of module case, and k is thermal conductivity of module case material. Qg (k ) indicates the heat generation which heats the battery while Qp (k ) represents the heat lost which cools the battery by heating surrounding air. Therefore, the battery temperature T (k ) can be calculated by T (k ) = T (k 1) + Qg (k ) Qp (k ) mbat Cbat (11)

and FOCV SoC () is the function mapping OCV to SoC. It usually is predetermined by manufactorys datasheet or experimental testing data.

where mbat is the battery mass and Cbat is its heat capacity.

C. Model based Prediction Besides the above models, to predict battery states under a given sequence of future charging current still requires to solve two problems: (1) how to initialize the prediction at time k and (2) how to realize the multi-step prediction using the one-step predictive model. To initialize the prediction at time k , RC model requires the last system states x(k 1), last input u(k 1) and model parameters at time k . Since u(k 1) and y (k 1) both are available from current sensor and voltage measurement respectively, x(k 1) can be estimated by closed-loop estimators such as extended Kalman lter [27], sigma-point Kalman lter [28] [29] and H lter [30]. The closed-loop lters have the ability to eliminate accumulated errors and estimation result will converge to the real value gradually or into a limited error range. The model parameters are stored in a look-up table indexed by SoC and temperature. The slow change of both SoC and temperature gives the reason to update the model parameters at time k according to SoC (k 1) and battery temperature T (k 1), especially when the Ts is comparatively small. SoC (k 1) is estimated by x(k 1) according to equation (3) and T (k 1) is measured directly from temperature sensor xed on the surface of battery. To predict T (k ), it is also necessary to know the last temperature of surrounding air Tair (k 1). One way is to additionally x a temperature sensor which closes to but do not contact the battery surface. On the other hand, in both practice and simulation, Tair actually is very closed to the ambient temperature and can be roughly treated as a constant. Therefore, the Tair (k 1) in equation (8) is replaced by Tamb and the model is greatly simplied. In this paper, to facilitate the simulation in Advisor environment, we assume a sensor is added to measure the temperature of surrounding air. The above preparation allows one-step prediction of SoC (k ) and T (k ) given the value of I (k ). However, a singlestep prediction usually is insufcient to predict the system behavior in a long process. Generally the prediction horizon in MPC is more than one step. Therefore, a multi-step predictor is necessary to predict system states exerted by a sequence of future charging current. A simple way to realize the multi-step predictor is iterative prediction by using the predictive future system states at time k + j as the initial states in the next time k + j + 1. In a summery, at control time k , the SoC model and temperature model can iteratively map a sequence of future charging current [I (k ), I (k + 1), , I (k + P 1)] to future battery states [SoC (k ), SoC (k +1), , SoC (k + P 1)] and [T (k ), T (k + 1), , T (k + P 1)]. The multi-step prediction is using an open-loop manner in which each step suffers from the prediction error of its previous step and nally remounts to the errors in the initial values at time k 1. Fortunately, these initial values are estimated in a closed-loop manner so that initial errors are limited to be small. Actually, any battery management system requires such an estimator to obtain the realtime SoC and measure battery temperature. The results can

be used as the initial values. III. F ORMALIZATION OF C HARGING C ONTROL UNDER MPC F RAMEWORK Given a series of control inputs, the length of which is denoted as the prediction horizon P , the future system states can be predicted based on dynamic system model. Then the future system behavior under the sequence of inputs can be evaluated based on a performance index. At each control time k , the basic idea of MPC is to nd an optimal sequence of control inputs [ u(k ), u (k + 1), , u (k + P 1)], which optimizes the performance index, and apply the rst element of the input sequence u (k ) to the system as the current control variable. In the charging control problem, MPC is to optimize a se(k ), I (k +1), , I (k + P quence of future charing current [I 1)] which has the best performance index evaluated based on the predicted battery states [SoC (k ), SoC (k +1), , Soc(k + (k + P 1)]. (k ), T (k + 1), , T P 1)] and [T A. Performance Indexes Performance index reects control objectives. The rst objective is to maximize charging speed. However, since it is impossible to directly predict the total charging duration in a limited prediction horizon, we turn to the tracking of an user-dened SoC trajectory instead. A fast rising SoC trajectory equally requires a fast charging speed while a at one represents a slow charging process. The expected SoC trajectory can directly copy from any real charging trajectory controlled by any charging scheme. In addition, the expected trajectory can be set as a real trajectory with revisions based on some special considerations. For example, for a CCCV trajectory, we may slow down the charging speed when SoC is very low to avoid high temperature rising and low energy efciency caused by the comparatively large internal resistance and fasten the charging speed when SoC is in the middle and internal resistance reduces to the smallest. Besides, the expected trajectory can be completely designed based on the demands of charging system, with no reference. To track the expected trajectory, a part of the performance index J1 is expressed by J1 = |SoC (k + P 1) SoC (k + P 1)| (12) where SoC is the expected SoC trajectory. Obviously, we evaluate the tracking performance of SoC only based on the nal prediction state because we simply require that SoC can achieve the expected point in the nal step. The process how to achieve that is not very important from the aspect of SoC. The second objective in charging control is to minimize the rising of temperature, which partially reects energy efciency and the system safety. Assume that the expected trajectory is copied from a CCCV charging process. Without the consideration on temperature, the best sequence of charging current is just the sequence applied in CCCV process. However, if we additionally evaluate the performance based on temperature as

expressed in J2 , the MPC will try to achieve the expected SoC point in the nal step with the process in which the temperature rising is minimized. (k + j ) T (k 1)|j = 0, 1, , P 1} (13) J2 = max{T B. Constraints The constraints in charging process is to strictly forbid the charging sequence which may cause damage to battery or lead to dangerous events. The rst constraint is that SoC can not exceed 100% to avoid overcharge damage. In practice, to reserve some tolerance, 98% SoC is treated as the full charging state. The second constraint is that temperature must be kept in a user-dened range to avoid overheating caused by overcharge or by a large charging current which exceeds the instant charging acceptance level. The two constraints are expressed by C1 : and C2 : SoC (k + j ) 98% (k + j ) T (k + j ) T (14) (15)
Fig. 4. Initialization of one special individual by introducing the best control sequence optimized in last step into the present step.

Fig. 3.

Scheme of the standard GA.

is the user dened safe where j = 0, 1, , P 1 and T range which is designed as either a constant indicating the highest temperature during the whole charging process or a time indexed function specifying the temperature limitation along with charging duration. IV. O PTIMIZATION OF F UTURE C URRENT S EQUENCE U SING G ENETIC A LGORITHM The MPC charging control problem is formulized to minimize the performance indexes J1 and J2 subjected to the constraints C1 and C2 . Essentially, the control problem is transformed to a constrained multi-objective optimization problem. Generally, multi-objective optimization problem can be either solved by multi-objective optimizers directly or rstly transformed to a single-objective problem and then solved by single-objective optimizers [31]. In this work, we apply the latter method because SoC tracking is more important than temperature management in charging process so that the two objectives can be merged into one index by summing with different weights. To solve the optimization problem, genetic algorithm (GA) is applied due to its strong global searching ability without the requirements on derivative information of objective function. Since the two constraints should be strictly satised in charging process to assure safety and the health of battery, the solutions which destroy any one of the constraints will be assigned the worst tness. Based on the above discussion, the tness function to be minimized in GA is expressed by F = 1 J 1 + 2 J 2 + (C1 and C2 are held) (otherwise) (16)

where 1 and 2 are the weights of J1 and J2 The minimization problem is solved by standard GA. the scheme of which is illustrated in Fig. 3 and briey described as the following steps.

Coding. The standard GA generally codes a candidate solution as a string of characters which are usually binary digits, named as a chromosome. The candidate solution is also termed as an individual. Accordingly, the set made up of a number of individuals is termed as population. In this chapter, we apply real-valued coding method, which codes a candidate solution as a set of oat decision variables. The real-valued coding method is proved to have superior performance against the binary-coded coding method in control optimization problems [32]. Initialization. The standard GA starts with the generation of an initial population. Usually, individuals in the initial population are produced randomly. In MPC, the initialization process is executed at each control point to start GA. Since the best control sequence optimized at time k contains good candidates from k + 1 to k + P 1, as shown in Fig. 4, one initial individual is specially designed by shifting it one time step and lling the last charging current with the same value (k + P 1). This individual introduces historical best as I perdition into current optimization process, thus it is helpful to improve the optimization performance, at least almost same as the previous prediction. Fitness evaluation. To each individual in each generation, we can evaluate its tness according to the equation (16). The smaller the tness is, the better the individual is. However, to facilitate the following selection steps, the raw tness usually is scaled to assigned suitable selection pressure to each individual. In this work, the scale function is expressed as Fscale = 1 1 + Fr (17)

TABLE I S IMULATION SETTINGS FOR CONSTANT PARAMETERS .

Symbol C mbat Cbat Ref f Tamb m air Cair Ts P 1 2 M axGen P opSize r R Pm Description Battery nominal capacity Battery mass Battery heat capacity Effective thermal resistance Ambient temperature Airow rate Air heat capacity Control period Prediction horizon Weight of SoC tracking J1 Weight of temperature rising J2 Maximum generation number Population size Power of raw tness in scaling The range of crossover coefcient Mutation probability Variance of Gaussian mutation Value 7 0.37824 795 7.8146 20 5.8333 1009 30 5 100 1 30 50 2 [0.1, 0.9] 0. 2 1 Unit Ah kg J/kgK K/W o C g/s J/kgK s Battery System

Fig. 5. Battery time-variant properties. Taking OCV and Ro as examples. Data source: Advisor.

Remark: The typical value of J1 is around 0.08 while that of J2 is around 2. Therefore, the real weight ratio of J1 to J2 is around 4:1.

where r is power of raw tness. A large r will enlarge the selection pressure quickly to worse individual, accelerate the convergence speed, however, and increase the risk of premature especially for multi-peak landscape, and vice versa. Selection. Individuals are selected from the last generation into the next generation based on the scaled tness Fscale following the survival of the ttest rule. Many selection methods are developed to avoid genetic drift and premature phenomena. In the work, the roulette wheel selection method is adopted [33]. The elitism strategy is also applied in selection to assure that the best solution will never be lost. Crossover. In the crossover step, the standard GA exchanges the information between two parent individuals and produces two child individuals. In this work, the arithmetical crossover method is used. Given two parents x1 and x2 , the children y1 and y2 are produced by linear combinations of parents with a random coefcient R : y1 = x1 + (1 )x2 , y2 = x2 + (1 )x1 . (18)

constant parameters in simulation environment are given in Tab. I. B. Evaluation Method Fast charging speed and low heat generation both are objectives in charging control. However they are conicted with each other. Fast charging essentially requires large current and hence leads to high heat generation, and vice versa. From the view point of multi-objective optimization, the conicted objectives are usually evaluated by Pareto curve. For charging control, x-axis is set as the charging duration and y-axis is set as the nal temperature rising T (end) T (0). As shown in Fig. 6, the curve with circle marks represents the Pareto front of CCCV family, where CC periods applies 1.5C to 6C current for fast charging respectively. Obviously, the CCCV Pareto front splits the objective space into two parts. Any charging result located in the right-up part is worse than CCCV family in both objectives while any one in the left-bottom part is better than CCCV in both objectives. Therefore, we evaluate the performance of charging controller according to the location of results. C. Simulation Results To facilitate the comparison with the conventional CCCV family, in this paper, we simple set the expected SoC trajectory by multiplying 1.05. The new trajectory is try to accelerate the charging speed. Since the working temperature of Li-ion batteries, especially for vehicular batteries, is from 20o C to 40o C, the ambient temperature is set as 20o C and the safe range of temperature is simple set as under 40o C. The search space of charging current is xed from 0C to 6C. The MPC is applied to track the modied SoC trajectories from 1.5C to 4.5C. Faster charging speed can not keep the rising temperature in the safe range. The Pareto front of MPC is illustrated in Fig. 6 by the curve marked with stars. Obviously, results of MPC all dominate those of CCCV family. Only the result of trajectory revised from 1.5C has the similar performance with CCCV. The reason is that 1.5C current is comparatively so small that it limits the applicable current

Mutation. After the crossover step, a subset of individuals is selected with a mutation probability Pm . Aiming at exploring the search space, we use Gaussian mutation in this work, which adds a random value from a Gaussian distribution with variance to each item of the selected individual. Termination. Many terminating conditions have proposed to stop the iteration process. For example, the distances among individuals are smaller than a predetermined value, an individual satises minimum criteria, or the maximum number of generations is reached. we simply apply the last one here. V. R ESULTS A. Simulation Settings To evaluate the proposed MPC based charging control strategy, simulation experiments are conducted based on a well established 7 Ah Saft Lithium Ion battery provided in Advisor. RC model parameters and OCV all are time-variant variables depending on SoC and temperature. The curves of OCV and Ro are shown in Fig. 5 as examples. Interpolating method is applied to create the continuous values space. The

GA

MPC

Fig. 6. Pareto fronts of CCCV and MPC charging methods. The expected trajectories of MPC are modied from CCCV by multiplying 1.05.

VI. C ONCLUSIONS Aiming at accelerating charging speed, reducing temperature rising, and keeping batteries away from damages, we introduce MPC framework to control charging process. Two conventional battery models are employed to predict SoC and temperature. In addition, standard genetic algorithm is applied to optimize the charging current under the multi objectives and constrains. Simulation results compared with CCCV family strongly demonstrate the availability and efcacy of MPC. We can reach the conclusion that the performance of MPC dominate CCCV. This work mainly introduces MPC concept and framework to solve battery charging problem. Many further revisions can greatly improve the results given in this paper. Firstly, to handle the inaccuracy and uncertainty of model, a Kalman predictor or H predictor is expected to outperform the prediction simply based on battery model. Secondly, advanced optimizers can obtain better changing sequence than standard GA. Third, the expected SoC trajectory can be reasonably designed based on the knowledge of battery properties and the controller can automatically produce best charging sequence to follow it. Finally, the tness function is very simple in this work. It can be further studied to evaluate charging sequence better. R EFERENCES
[1] R. Cope, Y. Podrazhansky, A. Technol, and G. Norcross, The art of battery charging, in The 14th Annual Battery Conference on Applications and Advances, 1999, pp. 233235. [2] S. Zhang, K. Xu, and T. Jow, Study of the charging process of a LiCoO2-based Li-ion battery, Journal of Power Sources, vol. 160, no. 2, pp. 13491354, 2006. [3] G. Hsieh, L. Chen, and K. Huang, Fuzzy-controlled Li-ion battery charge system with active state-of-charge controller, IEEE Transactions on Industrial Electronics, vol. 48, no. 3, pp. 585593, 2001. [4] Y. Wong, W. Hurley, and W. Wole, Charge regimes for valve-regulated lead-acid batteries: Performance overview inclusive of temperature compensation, Journal of Power Sources, vol. 183, no. 2, pp. 783791, 2008. [5] P. Notten, J. Veld, and J. Beek, Boostcharging Li-ion batteries: A challenging new charging concept, Journal of Power Sources, vol. 145, no. 1, pp. 8994, 2005. [6] S. Park, H. Miwa, B. Clark, D. Ditzler, G. Malone, N. Dsouza, and J. Lai, A universal battery charging algorithm for Ni-Cd, Ni-MH, SLA, and Li-Ion for wide range voltage in portable applications, in IEEE Power Electronics Specialists Conference, 2008, pp. 46894694. [7] T. Ikeya, N. Sawada, J. Murakami, K. Kobayashi, M. Hattori, N. Murotani, S. Ujiie, K. Kajiyama, H. Nasu, H. Narisoko et al., Multi-step constant-current charging method for an electric vehicle nickel/metal hydride battery with high-energy efciency and long cycle life, Journal of Power Sources, vol. 105, no. 1, pp. 612, 2002. [8] J. Diaz, J. Martin-Ramos, A. Pernia, F. Nuno, and F. Linera, Intelligent and universal fast charger for Ni-Cd and Ni-MH batteries in portable applications, IEEE Transactions on Industrial Electronics, vol. 51, no. 4, pp. 857863, 2004. [9] Z. Ullah, B. Burford, and S. Dillip, Fast intelligent battery charging: neural-fuzzy approach, IEEE Aerospace and Electronic Systems Magazine, vol. 11, no. 6, pp. 2634, 1996. [10] A. Khosla, S. Kumar, and K. Aggarwal, Fuzzy controller for rapid nickel-cadmium batteries charger through adaptive neuro-fuzzy inference system (ANFIS) architecture, in International Conference of the North American Fuzzy Information Processing Society, 2003, pp. 540 544. [11] R. Aliev, R. Aliev, B. Guirimov, and K. Uyar, Dynamic data mining technique for rules extraction in a process of battery charging, Applied Soft Computing, vol. 8, no. 3, pp. 12521258, 2008.

Fig. 7. Curves during CCCV and MPC charging processes using 3C prole.

sequence in a small range. If the xed search range can be accordingly reduced, i.e. from 0C to 3C, the result will be improved. As an example, the charging processes of CCCV and MPC are compared in the case of 3C, shown in Fig. 7. At the beginning of charge, the internal resistance is large when SoC is very low. In this process, the temperature rising dominates tness function. Optimized charging prole applies small current than CCCV. However, the increase of SoC tracking error gradually requires bigger current to speed up. Meanwhile, the internal resistance is reduced signicantly along with increase of SoC. Therefore, in the middle period, charging current of MPC is larger than CCCV. At the nal stage, the current in CV period is decreased to keep terminal voltage from overvoltage. Since the trajectory of MPC is modied from CCCV, the same trend is also remained in MPC, keeping terminal voltage under 4V.

[12] Y. Liu, J. Teng, and Y. Lin, Search for an optimal rapid charging pattern for lithium-ion batteries using ant colony system algorithm, IEEE Transactions on Industrial Electronics, vol. 52, no. 5, pp. 1328 1336, 2005. [13] H. Saberi and F. Salmasi, Genetic optimization of charging current for lead-acid batteries in hybrid electric vehicles, in International Conference on Electrical Machines and Systems, 2007, pp. 20282032. [14] L. Chen, R. Hsu, and C. Liu, A design of a grey-predicted Li-ion battery charge system, IEEE Transactions on Industrial Electronics, vol. 55, no. 10, pp. 36923701, 2008. [15] Z. Salameh, M. Casacca, and W. Lynch, A mathematical model for lead-acid batteries, IEEE Transaction on Energy Conversion, vol. 7, no. 1, pp. 9398, 1992. [16] H. Bergveld, W. Kruijt, P. Notten, and P. Notten, Battery management systems: design by modelling. Kluwer Academic Publishers, 2002. [17] V. Johnson, Battery performance models in ADVISOR, Journal of Power Sources, vol. 110, no. 2, pp. 321329, 2002. [18] G. Plett, Extended Kalman ltering for battery management systems of LiPB-based HEV battery packs:: Part 2. Modeling and identication, Journal of power sources, vol. 134, no. 2, pp. 262276, 2004. [19] A. Szumanowski and Y. Chang, Battery Management System Based on Battery Nonlinear Dynamics Modeling, IEEE transactions on vehicular technology, vol. 57, no. 3, pp. 14251432, 2008. [20] E. Camacho and C. Bordons, Model predictive control. Springer Verlag, 2004. [21] S. Qin and T. Badgwell, A survey of industrial model predictive control technology, Control engineering practice, vol. 11, no. 7, pp. 733764, 2003. [22] A. Bemporad and M. Morari, Robust model predictive control: A survey, Robustness in identication and control, pp. 207226, 1999. [23] J. Yan, Q. Ling, and W. Chen, Nonlinear model predictive control based on evolutionary algorithms: framework, theory and application, in Soft Computing: New Research, A. Giordano and G. Costa, Eds. [24] S. Piller, M. Perrin, and A. Jossen, Methods for state-of-charge determination and their applications, Journal of Power Sources, vol. 96, no. 1, pp. 113120, 2001. [25] T. Markel, A. Brooker, T. Hendricks, V. Johnson, K. Kelly, B. Kramer, M. OKeefe, S. Sprik, and K. Wipke, ADVISOR: a systems analysis tool for advanced vehicle modeling, Journal of Power Sources, vol. 110, no. 2, pp. 255266, 2002. [26] A. Pesaran, Battery thermal models for hybrid vehicle simulations, Journal of Power Sources, vol. 110, no. 2, pp. 377382, 2002. [27] G. Plett, Extended Kalman ltering for battery management systems of LiPB-based HEV battery packs-Part 3. State and parameter estimation, Journal of Power Sources, vol. 134, no. 2, pp. 277292, 2004. [28] G. Plett, Sigma-point Kalman ltering for battery management systems of LiPB-based HEV battery packs:: Part 1: Introduction and state estimation, Journal of Power Sources, vol. 161, no. 2, pp. 13561368, 2006. [29] G. Plett, Sigma-point Kalman ltering for battery management systems of LiPB-based HEV battery packs:: Part 2: Simultaneous state and parameter estimation, Journal of Power Sources, vol. 161, no. 2, pp. 13691384, 2006. [30] J. Yan, G. Xu, Y. Xu, and B. Xie, Battery state-of-charge estimation based on H lter for hybrid electric vehicle, in International Conference on Control, Automation, Robotics and Vision, 2008, pp. 464469. [31] K. Deb, Multi-objective optimization using evolutionary algorithms. Wiley, 2001. [32] B. Jiang and B. Wang, Parameter estimation of nonlinear system based on genetic algorithm, Control Theory and Applications, vol. 17, no. 1, pp. 150152, 2000. [33] D. Goldberg and K. Deb, A comparative analysis of selection schemes used in genetic algorithms, Foundations of genetic algorithms, vol. 1, pp. 6993, 1991.