This atc has been accepted for publication in future ioe ofthis journal. bat has not been fll edived Content may change prior to final publication. Citation information: DOM 10. “Tama on fol note Active Balancing of Lithium-Ion Batteries Using Graph Theory and A-Star Search Algorithm Guangzhong Dong, Member, IEEE, Fangfang Yang, Kwok-Leung Tsui, and Changfu Zou, Member, IEEE Abstract—The heterogencity of cells in a battery pack is Inevitable but brings high risks of premature failure and even safety hazards. Accordingly, for safe and long-life operation, Js necessary to adjust the state of charge (SOC) of all in-pack cells 0 the same level. To address this problem, this paper first proposes a battery SOC observer and analyzes its stability and convergence analysis using the Lyapunoy direet method, Different to:most available estimators is thatthe proposed method does not require the information of cell capacities, Then, after model the equalization system as a directed graph, the equalization problem is cast as a path searching problem. Finally, an A- star algorithm subject to balancing constraints is proposed to find the shortest path in this graph, corresponding to the most elficient SOC equalization, Experimental results show that the steady-state error of the proposed observer Is less than 2%. It also demonstrates that the A.star algorithm cam decrease the balancing time and energy loss during the balancing process by 9.59% and 19.5%, respectively, relative to the mean-difference- average method. Index Terms-—State estimation, lithium-ion batters, cell bale ancing, parallel operation, power converter. 1. INTRODUCTION A. Motivation and Technical Challenges NERGY storage technology plays an imperative role for nereased penciration of electric vehicles (EVs) in trans- port and renewable energies integrated into clectricity systems, which are two important sectors for a sustainable society [13]. Lithium-ion batteries (LIBs) represent today’s leading nergy storage technology for high energy density, low se discharge rate and long lifetime [4], [5]. Among the transport and power grid applications, a LIB system is composed of at least hundreds of cells in series- and parallel-connection to ‘meet the high-voltage, power, and energy requirements [6] However, due to battery manufacturing variation and uneven ‘operating conditions, e.g., temperature distribution, battery cells ina pack easily become unbalanced [7]. Such imbalance seriously restricts battery performance at system level and may leven cause safety hazards [8]. Consequently, itis necessary to design an efficient battery balancing system. ‘The assessment and climination of battery inconsistency are challenging from several aspects. First, in addition to ‘manufacturing deviation, causes of battery inconsistency are “This work wax supped by the General Research Fund (Projet No 11200819). (Conesponding author: Fang Yang: Chant Z0) 'G. Dong, F. Yang and K-L Tal ar wih he Schoo! of Data Science, City University of Hong Kong, Hong Kong (cma: gedong@icec ag; fangtyang2- [email protected] hk: Kisuieetyucd Bk) ‘©. Zow ts ith the Deparment of lsc Enginering, Chalmers University “of Technoogy, Gasbesbury 11296," Sweden (el chang zou@chalmersse) 08) a TEE Penna speed. bt piano ei IEE prin. S hw ‘Ahonzedicensed use bed to Unwersty of Sigh: Downloaded on June 8 2000 13 various, such as spatially uneven temperature distribution and nonuniform degradation dynamies. Second, the inconsistency fof battery cells can be reflected by not only the state of charge (SOC), but also the capacity, internal impedance, and ‘degradation path, Furthermore, the degree of cell inconsisten« ‘changes with time and is likely to grow with battery ages, ‘ease the weaker cells often experience deeper charging and discharging. Last but not leat, battery equalization requires in situ state information of all eclls,e.g., SOC and state of health (SOH). Online estimation of all the states fora large number of | cells is a non-trivial task and can pose a serious computational burden for battery management systems (BMSs). “To address the above challenges, this paper applies graph theory to battery management and propose an A-star path searching algorithm for effective and efficient equalization, B. Literature Review Numerous balancing methods have been proposed, and a plethora of relevant estimation models/ techniques have been reported [9, [10]. Considering the internal states monitoring, ‘equalization circuits, and equalization algorithms, the related literature can be reviewed as the following three aspects 1) Internal State Monitoring: To maximize the energy storage potential of a battery pack, the state of charge (SOC) of inmpack cells must be accurately estimated and balanced. SOC is the available capacity of a cell as a percentage of its total capacity [11], Ampere-hour counting is straightforward method of estimating SOC by integrating the current over time. However, this method requires prior known cell capacity. Furthermore, its estimation accuracy ean suffer from initial value errors and accumulated errors from noisy/drifted mea- surements [12]. Another widely used approach is model-based state estimation by combining electrochemical or equivalent circuit models (ECM) with adaptive observers to get accurate SOC estimation [12], [13]. However, due to high complexity, itis difficult to implement so many estimators for thousands of individual cells. An assortment of techniques focused on developing multi-cell estimation methods has been proposed, such as the dual-seale extended Kalman filter (EKF) in [141 [15]. Although these approaches could effectively reduce the ‘computational complexity, their acceptable performance is under the assumption of prior knowledge of the capacity difference among in-pack cells. Therefore, a convenient and accurate algorithm with low computing complexity and with- cout the assumption of prior known capacity is desirable for ‘monitoring the internal SOC of a battery pack stn straying or mor SB UTC komiZEE plore. Restichons spp."This artic has been acepted for publication in a future ioe ofthis journal. bat has not been fll edited Content may changs prior to final publication. Citation information: DOM 10.1.00/TT.2020.2907828 TEER “Tama on laa 2) Equalization Circuit and Topology: The equalization cit cuits can be principally categorized into dissipation and non- dissipation methods [9], [16]. The dissipation methods remove the excess charge through shunt resistors or other energy- consuming components for the target of equalization [6] Although these methods are easy to be implemented, they are Timited by the energy loss, low balancing current, low-speed, and heat generation. On the contrary, non-dssipation methods employ enengy storage components to allocate energy between individual cells. The non-dissipation methods have much more advantages of faster speed and less dissipative energy than the dissipation methods, despite the increasing cost and control complexity [17]. Considering the encrgy flow topology, single- layer serial and multi-layer parallel architectures have been proposed. In [18], three single-layer architecures, including adjacent cell-to-cell, direct cell-to-cell, and cell-to-pack, have been analyzed and compared. In high voltage applications, whore there are a large number of series-connected battery modules, the equalization speed of the single-layer serial topology is limited by the capacity of a single equalizer on the equalization path [19]. To further improve equalization speed and efficiency, several parallel architectures have been proposed and analyzed in [16], [8-120]. All the layers ean equalize corresponding battery modules and cells simultane- ously, and the balancing time can thus be reduced, Therefore, rult-layer parallel topology is chosen for equalizing SOC slobally (not only from cell to cell but also from module to module). 3) Equalization Algorithm: fer the equalization cireuit and topology are chosen, an optimization equalization algo- rithm is necessary to obtain a global equalization target. The ‘mean difference-average method (MDA) is simple and easy to implement, where the high-SOC cells are set up to discharge, while low-SOC cells are set up to charge [21], However, these iethods feature with low equalization efficiency. Some other advanced intelligent algorithms have also been employed to solve the optimization problems in equalization systems. For Instance, fuzzy logic [22], quasi-stiing move contol {7], aulapive backward control [23], decentralized strategies [8], 1241, and minimal charge transfer [25]. AIL these algorithms have achieved acceptable equalization performance. However, most of these balancing methods are designed (0 equalize the single layer serial topology, which is not convenient for describing and analyzing global equalization issues. On the contrary, the graph-theoretic is a good choice to represent and analyze equalization systems, because unnecessarily complex models of power converters can be ignored. Furthermore, the cqualization problems ean be deseribed as a path searching process in a dirceted graph. C. Key Contributions ‘The first contribution of this paper is an efficient and robust SOC observer, without prior knowledge of battery capacity This is conducted by first estimation of the open-circuit- voltage (OCV) using a recursive least squares (RLS) based algorithm. With the OCV estimate as a virtual measurement, 1a Luenberger observer is devised for battery SOC. By using (IEEE. Pras etd ut epiinioon ei IEEE perinon She Fig. 1 Diagram ofthe two-level parallel equalization architecture Lyapunov direct method, the estimation error is proved to have a guaranteed stability and convergence, ‘The second contribution is a graph-theoretic battery model developed for a parallel equalization circuit topology that transfers energy from cell to cell. The equalization problem is formulated as a path searching problem in a directed graph. ‘This makes it possible to analyze and solve equalization, problems without complex converter models. Finally, an A-star search algorithm is for the first time ‘employed (0 find the shortest battery balancing route in the ‘constructed graph. Tlustrative resulls demonstrate that the A-star algorithm is able to significantly decreases both the balancing time and energy loss. IL, EQUALIZATION TOPOLOGY CONFIGURATION Equalization is more complex in medium and large-scale systems, which are typically partitioned into modules [26] Fig. 1 shows the architecture of a modular battery pack. ‘The implementation of the BMS functions is distributed over fa master-slave architecture. Each module management unit (MMU) can monitor parameters and protect cells belonging to the module, However, equalization among cells in a module ‘cannot guarantee that the whole pack is equalized. Therefore, ‘two-level parallel architecture is designed to equalize battery SOC not only from cell to cell but also from module to ‘module. Compared to the conventional single-layer topology, the two-layer architecture show better performance in less ‘energy loss, faster balancing speed, and higher flexibility [18]. In this section, a two-layer equalization architecture is first given. Then, the converter models for different layers are introduced. A. The Two-level Parallel Architecture As shown in Fig. 1, the MMUs communicate with the ‘central management unit (CMU) via a serial controller area ‘network (CAN) bus. Each MMU can independently measure the voltage and temperature of the cells belonging to the ‘module. The equalization of these cells can also be effectively carried out through individual cell equalizers (CE) and the ‘MMU. Firstly, the MMU can estimate the SOC of every ‘Ahoraed iceneed uss ined to Uwraty of nghion Dewnlsied on func 8 2000 13 4659 LTC wom IeSE Mpa Reston apy an strapon or mor norman."This artic has been acepted for publication in a future ioe ofthis journal. bat has not been fll edited Content may changs prior to final publication. Citation information: DOM 10.1.00/TT.2020.2907828 TEER “Tama on laa individual cell according to the measured voltage and bus ccurrent measured by CMU. Then, the firs-layer driving signals for ICEs are generated based on the equalization algorithm, However, if the equalization is carried out by the MMUs independently, the equalization of the whole pack cannot bbe guaranteed, Therefore, the second layer equalization for transferring energy from module to module is designed. This process can be carried out through individual module equal- izers (IMEs) and the CMU. Firstly, the CMU can collect all the SOC values of the battery pack via CAN bus. Then, the socond-layer driving signals for IMEs are generated according to the equalization algorithm. The proposed two-level parallel equalization topology includes n modules, and each module contains p cells. The proposed equalization topology makes it possible to transfer energy between cells of different modules to perform global SOC equalization of the battery pack. B. Individual Cell Equalizers In this paper, the isolated bidirectional flyback transformer based buck-boost converters are chosen as the ICEs. The ‘eell-o-pack” and ‘pack-to-cel” are chosen as the equalization topology of the frst layer ICEs, because the cell-to-pack ean avoid over-charging the cell to be equalized, in case it has a higher voltage or SOC than the reference one, while the pack- to-cell can avoid over-discharging the eell to be equalized, in case it has a lower voltage or SOC than the reference one [9]. Fig. 2(a) illustrates a schematic diagram of three cells in a ‘module, connected by the selected equalizers. For the #* ICE, composed of two MOSFETs Cz and Gi, with 10m0 drain-source on-state resistance, two resistors ftp and Ry, and a transformer 7. The equalization process and the working principle of the equalization circuit are analyzed as Follow. 1) Cellto-pack: the Cy is firstly turned on, and the Gi. is tuned off. T; is charged by the #* cell. Then, the G.,. turned on, and the Gi.y is tumed off. T; charges the battery pack. 2) Pack-to-cell: the Gi» is firstly tumed off, and the Gi, is turned on. 7; is charged by the pack. Then, the Gi» is tumed on, and the Gi. is tumed off. 7; charges the cell, By repeating the above steps, the energy ean be transferred either from cell-to-pack or pack-to-cell It should be noted that the duty eyele of Gis p times more than the duty eyele of Gi,y, where p is the number of series-connected cells in the pack. The equalization currents are regarded as the average inductor currents in the following sections, according to the principle of inductor volt-sccond balanee [7] ‘A preliminary simulation is conducted first to determine the equalization eurrents and energy efficiency. The battery module is composed of three cells. The frequency of the PWM applied on the MOSFETS is 124 kl, and the duty cycles applied on the Giy and Gi, are set as 759% and 24%, respectively. The circuit parimeters. are selected as Ry = Ry = 2m, and T, = 10 pH. The voltages of three cells are set as Up, = 4.01V, Up, = 3.91V and Up, = 3.88V, respectively. The simulation is performed in MATLAB/Simscape, where B1 is set to discharge, B2 has no ns (200 TEE, Per ne ipeed tpaiinrbsin es EE min Sc pms nbn nda pbc fr mornin ‘Aahorzed Weeneed used nvesty of ghion Downloaded an une 082000 a 1 8% ‘foe Res Fig. (a) Diagram of the first layer ICE: (6) Equalization cumens of ‘cli pack and (2) pack el: () The isolated PM signals fr primary fu secondary MOSFET, | sac a 7 Bee Fig. 3. (@) Digram of the second layer IMs; (b) Equalization currents of Mto-M2 and (b) MZ-to-MI:(@) The solated pon signals for MOSFETs Q, and Qe action and B3 is set to charge. The transient equalizing current curves through the equalizer are ilasrated in Fig. 2(6) and (©. During celbto-pack operation, the average equalization currents of primary and. secondary sides are about 0.564 and OSIA, respectively. The energy efficiency is about 90% through the 1" converter, During pack-to-cell operation, the average equalization currents of primary and secondary sides au about 0.53A and 0.57A, respectively. The energy eficiency is about 91% through the 3" converter. The PWM signals are isolated from the switches on the baller side, as shows Fig. 2). Remark 1: In onder to protect in-module cells from being overcharged and overdischarged, only the cell with highest SOC is selected 10 execute cell-to-pack equalization during the rest or charging process, while only the cell with lowest SOC is selected t execute pack-to-cell equalization during the ‘ischarging process SB UTC komiZEE plore. Restichons spp."This artic has been acepted for publication in a future ioe ofthis journal. bat has not been fll edited Content may changs prior to final publication. Citation information: DOM 10.1.00/TT.2020.2907828 TEER “Tama on laa ©. Individual Module Equalizers Ti this paper, a batery pack consisting of six individual cells is used to demonstrate the proposed A-star-based equalization algorithm. Thus, the non-solated bidirectional Cuk convert- ers are chosen as the TMES, because they feature with the charactristis of satisfactory equalization speed, excellent im- plementation, reasonable size, and acceptable cost [23], [24], [27]. It should be noted that the isolated Cuk converters are {beter choice for high voltage applications when increasing the cell numbers in a module. Fig. 3(a) illustrates a schematic diagram of two modules in a pack, connected by the selected equalizer. A n modular battery pack requires m — 1 converters ‘as the IMEs. For the IME between i™ and (i-+1)" module, it is composed of two uncoupled inductors [and yan enerzy transferring capacitor C,, and two MOSFETs Q, and Qy1 with body diodes and 10m0 drain-source on-stae resistance The equalization process and the working principle of the equalization cieuit are analyzed as follow 1) Module i to Module i+ 1: the Q, is tured on, and the Quay is tuned off. 2) Module i + 1 to Module the Q, is tumed off. ‘The equalization currents depend on the PWM duty cycle applied on MOSFET Q, and Qu. According to [24], the average inductor currenis are expressed as: TU VoD? the Qc is turned on, and "FE e. Ua) ° TU VeDiyr thoes ~ FhigalUe, — Ua...) ° where TT is the switching period. Dz and Dey are the duty cycles of the PWM signals on @Q, and Qcya, respectively. Unt, Une.» and Ve, ate terminal voltage of the i, (5+ 1)® ‘modules and capacitor Cy. 11, and J,,, are average inductor currents, which can be regarded as constants for a particular duty cycle [27]. Considering that the terminal voltage of two battery modules Ugg, and Us,,, are close, and Uc, Untazy + Use, The equalization current can be approximated as constants once the duty cycle is selected. A preliminary simulation is conducted first (0 determine the equalization currents and enengy efficiency. The battery pack is composed of two modules. The frequency of the PWM applied on the MOSFETs is 7 kHz, and the duty eycles applied om the Q; and Qis1 are set as 0.45. The circuit parameters are selected as Li = Lier = TOQUH, and Cy = A70uF. ‘The voltages of two modules are set as Ung, — LLSV and Unt, = ULBV, respectively. The transient equalizing current curves through the equalizer ar illustrated in Fig, 3(b) and (c). During M1 to M2 operation, 11, ~ 0.50 and J), ~ 047A. ‘The energy elficiency is about 90% through the converter. During M2 to M1 operation, 11, ~ 043A and J) ~ 049A. The energy efficiency is about 91% through the converter. The PWM signals are isolated from the switches on the battery Side, as shown in Fig. 0), III. BATTERY MODELING AND SOC IMATOR ‘This paper aims at overall SOC equalization of a battery pack, A sccond-order Thevenin ECM is firstly introduced. 08) a TEE Penna speed. bt piano ei IEE prin. S hw ‘Auhonzedicensed use led io Lnwersty of Sigh. Downloaded on une 082000 213 Fig 4 Thevenia equivaletcitcit mode TABLEL [SPRCIFICATIONS OF TEST CELE. ope cra Taior Cy) Cavoll valage TW) LiNiMaCoO iC 18650 wo 4207250 ‘Then, the RLS-based OCV estimator and a robust SOC es- timator are designed for individual cclls. The SOC estimation ‘method for a battery pack is finally proposed. A. Thevenin Equivalent Circuit Modet ‘The second-order Thevenin ECM is shown in Fig. 4, in Which Use is the OCV, It is the ohmic internal resistance, Fs, Cid = 1, ate patallelly connected polarization resistance and capacitance, respectively. Ir. is the load current (defined to be negative for charge and positive for discharge). Us is the terminal vollage. The dynamic behaviour of this model ean be mathematically captured by [12]. Op =—Uifnt Hh /Cy t= 1,2, i) Up = We — Lay ~ Us ~ Us “ where 7 = FC are the time constants. Ui is the voltage of the RC networks. Since the battery pack is composed of a large number of individual cells, the inconsistency of the model parameters between different cells must be considered to monitor battery inner states, ‘To investigate the inconsistency of model parameters, such as capacity and impedance, Fig. 5(a) plots pulsed voltage re- sponse at different aging conditions. These data were collected fon the second generation, 18650-size LIBs from the Idaho National Laboratory [28], with specifications in Table 1. As in Fig. S(a), the capacity decreases and resistance increases along with the battery aging. Furthermore, the battery capacity and impedance change significantly along with temperature [29] ‘Therefore, to monitor the SOC of in-pack cells without prior known capacity and impedance parameters, one has to estimate battery SOC and identify model parameters simultancousl. For the viewpoint of the implementation of the SOC estimation algorithm, we made the following assumption to clarify the OCV-SOC relationship. Assumption I: The individual cells in the battery pack share 4 common OCV-SOC relationship, which only depends on cell's SOC. The aging, temperature, and inconsistency effects are negligibly small. SB UTC komiZEE plore. Restichons spp. an strapon or mor norman."This artic has been acepted for publication in a future ioe ofthis journal. bat has not been fll edited Content may changs prior to final publication. Citation information: DOM 10.1.00/TT.2020.2907828 TEER “Tama on laa (a) The pater psd vollage pies (b) the OCV.SOC relationship under diferent enpeatrs Justification 1 (For Assumption 1): The Assumption 1 will be justified by following three aspects: ‘Temperature effects: To investigate the temperature ef- fects, 5(b) plots OCV-SOC curves of a 18650 LiNiMnCoO./Graphite LIB under different temperature. The experimental dataset was collected by CALCE (The Center for Advanced Life Cycle Engineering the University of Mary- land) [30], with spec ‘Table I. It ean be seen thatthe low current OCV curves at three temperature points overlap for a wide SOC range. The difference among the three curves at different temperatures mainly appears at the beginning and the end of SOC ranges and is found to be negligible (30] ‘Aging effects: The aging effects on battery OCV-SOC relationship were investigated by measuring the OCV of fresh and aged cells at 25°C and a low rate (C25) by (311. The results suggested that only slight differences ean be identified between the OCV curves of the fresh and the aged cells [31]. Inconsistency effects: Inconsistency effects: In terms of the implementation of state estimation algorithms for a multi cell battery pack, it assumes that all in-pack cells share the common OCV-SOC relationship because it mainly depends on the concentration and type of active material in the positive and negative electrodes, [32] (33) In 34] and [35], OCV-SOC curves of nine individual cells and four individual cells were plotted, respectively, and the results experimentally support this assumption. ‘The discrete form of (3) can be expressed as a) Ree at=1,2 6) where a7 = exp(—T;/72), the sampling rate is denoted as T,, Combining (4) and (5) by eliminating Ua, we can obtain, Une = aig aH. Tat O. Thus, if Py is initialized asa postive definite symmetric matix, Py can be @ bounded postive definite matix. Define the Layapunov candidate function as Vk = OPLO}, where 0, — — is the parameter estimation cor. 0, can be deduced from (8) and Algorithm 1, i — (1 Perdudh_\g, Perdue os (1 Pi i) de V+ @IPi ade XoR ‘Thus, the one step time difference of the Layapunov candi- date function Vj, can be expressed as, a2) AVe = Vi ~ View a od, 2M | sry FO SOE soe |e OD) where ddd 7 %= [a—ae, Tyee] 20 y= Pacis "GIP hide From (11), we have Py! > "P54, then AV ean be deduced as, Av <-(1 sa AIP OE Oe + vbr ANGE Be a/t vb where AVi_ is negative outside the range Ok-alle < Jov's JU ~A)/X. The Lyapunov function Vi, mono: ally decreases until Oks» enters into this range, which can be determined by appropriately selecting w. Furthermore, the second-order battery model in (8) is @ widely verified in describing battery behaviors [15], (32), ie. ~ O. Thus the parameter estimation error will converge tothe vicinity of the origin based on (14). ‘After we got the model parameters Oy based on the Algo- rithm 1, according to (7, the estimated battery OCV can be recursively calculated by aay Goo bs 2Woe1 + ba 6 Uoese2 ~ Os4Uea1 — Gag yaa Ui Oaatene ~Osilte1~ O54za-2 (15) Convergence analysis of OCV estimator: the dynamics of OCY estimation error (Unc) can be expressed as, [Te] [ae al Coe where the eigenvalues of this dynamic are a and a, which are both within the unit circle, Therefore, as long as the identified model parameters converge, the OCV observer is stable regardless of the choice of the initial values, The OCV- SOC relationship in [36] is utilized in this paper. Coea-2 Doeh—1 | 6) ns (200 TEE, Per ne ipeed tpaiinrbsin es EE min Sc pms nbn nda pbc fr mornin ‘Aahorzed Weeneed used nvesty of ghion Downloaded an une 082000 a 1 8% ‘foe Res ©. Robust SOC Estimator Considering the definition of battery SOC, the SOC estima tion model can be represented as ae = et Bln tik Woes = 924) + Pie an as) where = © [0,1] is the abbreviation of battery SOC. B =T,/Cy, and Cry denotes the nominal capacity. wi, = ABI, + 1, denotes the model bias due to the incorrect capacity and unknown model noise. We assume B + AB roflects the accurate battery parameter and that wy is zero mean Gaussian noise. g() is the OCV-SOC function. ‘denotes the OCV estimation errors caused by Assumption 2 ‘A robust observer for (17) and (18) can be designed by oF) where 2 is the estimated SOC. Cex. is the estimated OCV by the RLS algorithm in Section TEB. Ke R™ fy the observer gain. Using the first-order Taylor expansion of 9(-), the estimation error dynamics can be expressed by Sy = 24 Blan + Koo as) (1 KO) + (E- KE )wy 20) 9/02|,-s,, B = [1,0] © RY, F = [0,1] © RE, wa = fit, On(Se) + tc)". and n(x) is the high- onder term of the Taylor expansion. In this model, [well < ‘h-< oe is norm bounded. The aim is thatthe ervor dynamics in 20) is stable and thatthe Following inequality is satisfied under 2er0 inital conditions fora given attenuation level y > 0 BT Well < vhwell en Convergence analysis of SOC estimator; Select a Lya- punov candidate as Ax = [Wx, W > 0, the one step time difference of the Lyapunov function can be expressed as, 2 (1 KOW(E~ KFY" (E~ KF)'W(E~ KP) san be guaranteed that the error dynamics (20) is asymptotically stable, From (21), the performance index ean be defined as [ (i Koew—W (1 KC)W(B- KF) ] oa wt; ‘To satisfy (21), for Vk > 0, J need to be negative. It yields 03) =3 Ws o?w}wy +n) “Sas, sPwlw, + AA,) <0 ey SB UTC komiZEE plore. Restichons spp."This artic has been acepted for publication in a future ioe ofthis journal. bat has not been fll edited Content may changs prior to final publication. Citation information: DOM 10.1.00/TT.2020.2907828 TEER “Tama on laa which is equal 10 Tao 26 [Fa |0 Wo W-@c WE-QF Wego 1-W Oe | <0 (WE-QF) Ox. 9? laxe whore Wand @ are the optimal variables. The observer K’ = W—~'Q. This optimization problem can be easily solved by MATLAB Linear Matrix Inequalities (LMI) toolbox (38). D. SOC Estimation of A Battery Pack In battery pack, each estimator works independent. The 4 cell's curent 7), can be formated as: 19, = daa + TY, 26) where J4 denotes the bus current, which is the same for all in-pack cells. 9. donotes the balance current. Afterall he in- pack cells’ OCV is estimated by Algorithm 1, every individual cell's SOC can be calculated using the robust observer TV. A-STAR ALGORITHM BASED In this section, the graph-theoretic-based equalizing system ‘model is first given out. The equalizing problem formulation is then described based on the system model. Finally, an A- based equalization algorithm is presented to solve the cell equalizing problem, BALANCING: A. Graph-theoretic-based Equalizing System Model ‘To illustrate the graph-theoretic representation of the pro- posed two-level parallel equalizing architecture, all the cells and energy storage devices, such as transformers and capae- itors are represented as “nodes”. The paths of energy flow between these nodes are represented as “edges”. ‘The digraph of the proposed equalizing architecture is diagramed in Fig. 6, where the energy flow paths are bi-directional. The cells in the same module are equalized based on the frst level ICEs, which employ cell to module oF module t cell energy flow. The cells in the different modules are equalized based on the second level IMEs. The reachability matrix Ft, of the digraph 6 can be expressed as Cia Cha Cra Cas Coa Coa Ga (o 2 Bon aon Gf 0 8 y a 0 Rn= Cs | 3 0 0 a on Calon on nn 0 BB Galn nn % Oo Gs \n nn 2 BO (IEEE. Pras etd ut epiinioon ei IEEE perinon She Cu Cia ‘Madde cn Cu cn Cu ‘Module Fig 6 Digiph the 2 evel pall achietre where 1 © 90% according to the pretiminary simulations Rey gives out energy cost during the equalization process. Tt cean be found that the directed graph in Fig. 6 is complete, because every cell in the pack is adjacent. The average energy efficiency can be computed by averaging all the nonzero terms in Ry, which is 18n/15 forthe designed topology. B. Cell Equalizing Problem Formulation For the i" in-pack cell in the battery pack, its ue SOC (6) can be expressed as © 010 Alena tk where Cy,s denotes the nominal capacity of the #* cell, and Cys = 22A-h. EAT is the equalization period. Duc to Cy = 3600 x 220s, GAT/Cy, can be regarded as a small constant ¢ for all impack cells Therefore, the it cell's SOC canbe further expressed as ae = 2) — 6 Ioa —¢-, 28) ‘The equalization process can be regarded as redistibuting energy stored in in-pack ells so thatthe cells’ SOC differences converge toa tolerant range. In ordcr to improve equalization efficiency and speed, we also need to minimize energy loss and balancing time during the equalization process. Thus, the objective of cell equalization can be expressed as: EAT/CK ss en lle — Selloo SVRAT > Ts, e9) where = (30,22), ,3()P denotes the estimated SOC Vector ofthe Baty pack af sampling ime and the mame i aN umber of serially connected cells. 2,) = 2,2 +f. uf? is the estimation error. i, denotes the average of the estimated SOC, and ¢ is a given positive constant representing tolerable SOC difference. Defining a row demand vector D which contains the bal- aancing energy demand of each cell, ie. the deviation from the ‘mean value in terms of SOC, the objective of the equalization process is to find the shortest path to transform the o0-norm of D into «. Remark 3: From (28), it ean be seen that the common ‘change trend of all in-pack cells’ SOC are the same in a short period, no matter the battery is charging or discharging. ‘Therefore, the balance demand energy of each cell can be regarded as constant during a short period, Remark 4: In order to avoid frequent operation of equal ization switches, itis assumed that the duration of the battery ‘Ahoraed iceneed uss ined to Uwraty of nghion Dewnlsied on func 8 2000 13 4659 LTC wom IeSE Mpa Reston apy an strapon or mor norman.This atc has been accepted for publication in future ioe ofthis journal. bat has not been fll edived Content may change prior to final publication. Citation information: DOM 10. “Tama on laa Fig. 7. Schematic diagram of equalization proces. stant € which is defined 3 demand vector D can be ion process. The diagram is shown in Fig. 7, where D is the destination point, and the meanings of the eight adjacent nodes Sy = 1+ ,8) of point $ are shown in Table I ‘The cost of S, represents the charge loss from $ 10 5, Which is approximated by a constant because the balancing curtent and equalization switching frequeney are assumed 10 be constant For j € {1,5}, only the IME equalization is executed, and the charge loss can be calculated as: Costs, ¢ do 4 Estimate the cost from ¢ to its neighbors: 5: for j =1,---,8 (eight neighbors of <), do ® GS) = Gl) + Cost(S): 1% GUS) = FO) F(S}) = G(S,) + HC & end for 5: Determine current Equalization order: S* = argmjn FS) where S* represents the solved Equalization_order. Ww Update D = s+ — 25e,and ¢ = S* us end while im else 1 retum to step 2. ue end if Fig. 8 Veication ofthe popsed observer under USO6G@OPC: (2) OCV,() SOC, under FUDS@-45°Ce) OCY, ) SOC: (e)etimaed baer resistance, (0 the obser gain vs. SOC. in Scetion V-A. ‘Thea, a simulation study is uscd to verify the SOC estimator under dynamic load profiles for a six-cell battery pack. The simulation results and discussion on the equalization algorithm are finally described in Section V-B. A. Verification of SOC Estimator for Individual Cell ‘The test data under loading profiles US06@0°C and FUDS@45°C (the Federal Urban Driving Schedule) in [30] are chosen to evaluate the performance of the proposed battery ‘model and observer. The parameters of the proposed method are chosen as \ = 0.985, w = 100, To evaluate the perfo ‘mance of the proposed method, the Coulomb-counting method is used as a benchmark (marked as “Ref.), whose initial value and battery capacity are prior-known. The experimental results are plotted in Fig. 8. Fig. 8(a) and (b) illustrate the estimated OCV under USO6@O"C and FUDS@45°C, respectively. Fig. a 200 TEE Peon is pet bot bonding eis IEE perminon Sew ‘Auhonzedicensed use led io Lnwersty of Sigh. Downloaded on une 082000 213 ‘ant [NUMERIC RESULTS OF THE PROPOSED ORSERVER. Conditions MAE CRY cn a FUDS@4s°C 062053 086 00 8(c) and (d) illustrate the SOC estimation results under two loading profiles, respectively. It can be seen that the proposed RLS-based OCV observer can accurately following the actual trajectory. From Fig. 8(c) and (d), it ean be concluded that the ‘proposed SOC observer can effectively reduce the estimation ‘errors caused by the parameter estimation error and model bias. Under USO6@0°C condition, the actual battery capacity is 1.82Ah, Although there exists about 10%Cy model bias, the proposed SOC observer can still track the referring trajectory. ‘The numeric results are listed in Table TM, in which the “RLS” denotes SOC estimated by table look-up method based on the RLS-based (jc, and the “Robust” denotes SOC estimated by the proposed observer. The Maximum Absolute Errors (MAE) and Root Square Mean Errors (RMSE) are calculated, respectively, as en G8) where =; denotes the reference SOC, and IV is the data length. From Table II, itis manifested that the proposed method ean estimate battery SOC accurately and robustly, and that the estimation value can converge to the reference SOC within an cero of 42%. Furthermore, from Fig. 8), it can been seen that the proposed RLS-based parameter estimation methods ‘ean also provide accurate battery resistance, in which the black dash lines are the offline identified ombic resistance reported in [30]. Therefore, the advantages of the proposed methods in on-line parameter identification and SOC observer are demonstrated. The relationship between the observer gain K and SOC calculated by the LMI toolbox are also presented in Fig. 80. B. Verification of SOC Estimator for Battery Pack A simulation study is performed to validate the performance of the proposed SOC estimator for a battery pack, which is ‘composed of six series-connected individual cells. The pack’s model parameters in the simulations are manually chosen as ‘Table IV. The FUDS loading profile is used in the simulation ‘The simulation results are plotted in Fig. 9. Fig. 9(a) and (b) give out the FUDS current profile and simulated data of cell 1# as an example, Fig. 9(¢) and (@) plot the estimated OCV and SOC errors of six in-pack cells. From these results, we can sce that the proposed observer for the battery pack works well under dynamic loading current, It should be noted that only the OCV-SOC relationship is used in the estimation, without prior- known battery capacity and other model parameters. Tt can be SB UTC komiZEE plore. Restichons spp. an strapon or mor norman."This artic has been acepted for publication in a future ioe ofthis journal. bat has not been fll edited Content may changs prior to final publication. Citation information: DOM 10.1.00/TT.2020.2907828 TEER “Tama on laa Tine 8 Te lf oy Fig. 9. Vesestion of propsed OCV estnatoe fr batterypack (a) FUDS amen pote, simulated daa of ell 1 @) OCW estimation eros (@) SOC estination ero TABLE IV MODE. PARAMETERS OF THE BATTERY PACK INTHE SIMULATION STUDY. seen from the Fig. 90) that the SOC estimation error within an error of 2.5%. The average RMSE of the six cells is 1.26% ‘These results indicate the accuracy of the proposed method. Furthermore, the OCV errors at low SOC range are a litle larger than other area because of the fast falling OCV value at low SOC range violates the aforementioned OCV assumption (Woe. ~ Usa * 0). However, the proposed observer can still converge to the actual SOC value, which demonstrate the robustness of the proposed observer. ©. Verification of the A-star-based Equalization Strategy In order to evaluate the performance of the proposed A-star based equalization algorithm, the widely used MDA is used as benchmark: Benchmark: The MDA follows two stages. First, the IME. equalization is executed to balance the average SOCS of modules. Namely, the energy is transferred from module with higher average SOC to module with lower average SOC. The criterion for stopping the IME equalization isthe average SOC © between two modhles is less ¢/2. Then, the ICES jon is exccuted to balance the in-module cells until the criterion of cell equalization (29) is met. Simulations are employed to validate the performance of the A-starcbased cell equalization strategy when the battery pack is in standby, charge, and discharge modes, respectively. © is set at 1.2%. The cells’ initial SOC are set as. 7% (60%, 72%, 76%, 65%, 68%, 61.5%]", which can be obtained based on constant current discharging from fully charged states. The cells’ SOCs during standby equalization process are shown in Fig. 10(a) and (b), respectively, while those during charge (0.54) and discharge (2A) equalization process are plotied in Fig. 10(¢-f, respectively. Fig. 10(a, (c), and (e) are a 200 TEE Peon is pet bot bonding eis IEE perminon Sew ‘Auhonzedicensed use led io Lnwersty of Sigh. Downloaded on une 082000 213 soc eee € E E soci weds Fig 10. The in-pack cell? SOC during equalization process sing benchmark (G2, (©) and (e) and Asta alsoridus (). (2) and (and () stay ‘mode, () and (2) charging ode, () and () discharsing made. ‘TABLE V COMPARISON BETWEEN THE PROPOSED METHOD AND THE BENCHMARK, Dac MDA AT MDA AT Tamia) 613 SAD 6RT SHAT 320 Lon (Amia) 773 Gal 788 640493 3.76 the results using the benchmark method. Fig. 10(b), (4), and (faze the results using the A-star algorithm, The oo-norms of the demand veetors D under the three scenarios are shown in Fig. 11(a). The total charge losses under the three scenarios are shown in Fig, 11(b). As shown in Fig. 11, all of the o0- norms of D, which denote the maximum difference between actual SOCS and the average SOC, can converge to 1.2%. However, the equalization speed and charge loss are different. ‘The numerical results are listed in Table V. Tt can be seen that the proposed method works better than the benchmark method under all the three scenarios, In order to further validate the robustness of the proposed method, we randomly generated 1000 test cases under standby niode. Each test ease consists of a set of initial SOCS for the six cells in the battery pack. The initial SOCs are randomly generated subject to the uniform distribution U/(0.6,00.75). The uniform distribution U/(—1.5%, 1.5%) is employed to simulate the SOC estimation error. Ia these 1000 test cases, the average balancing time and charge loss of the benchmark method are 31.8 min and 3.60 Amin, respectively, while the average balancing time and charge loss of the A-star algorithin are 25.7 min and 2.90 Amin, respectively. The histograms of the reduced balance time and charge loss of all the test cases, using the proposed method are plotted in Fig. 12(a) and (b), respectively. Tt can be seen that the proposed method can shorten the equalization time and reduce the charge loss in ‘most of the test cases, as compared to the benchmark method. ‘The average reduced equalization time and charge losses. SB UTC komiZEE plore. Restichons spp. an strapon or mor norman."This artic has been acepted for publication in a future ioe ofthis journal. bat has not been fll edited Content may changs prior to final publication. Citation information: DOM 10.1.00/TT.2020.2907828 TEER “Tama on laa oy @) [ilo (charging toss during equation, lhl 8 pedcss Tin 2 0 Mas @ enc! 2. Comparison between benchmark and Asta slgoridims wader 1000 Simulated cases (a) reduced equalization ime, (b) raced charge cs of the ‘Asa algorthn; maximum difereace aller equaliaton:() benchmark, (@) Nae are 6.10 min and 0.70 Amin, respectively. ‘The maximum difference between actual SOCs and the average SOC after equalization processes are plotted in 12(c) and (), in which ‘we can see that the estimated SOC difference can converge to &, while the true SOC differences are below 3%, which is higher than ¢ due to the existence of the estimation errors. ‘The negative influences will also increase with the inetease of estimation errors. However, how to quantify the negative influence caused by the SOC estimation errors and to reduce it during the equalization process are beyond the scope of this paper. From the experimental results, it can be concluded that the proposed topology and A-star based algorithm can balance cells’ SOC difference to a tolerant range. The proposed A- star algorithm is more efficient and robust than the widely used MDA method. The A-star algorithm can decrease the balancing time and the energy loss during the balancing process by 9.59% and 19.5%, respectively. It can be found in Fig. 10 that the proposed battery topology can improve battery capacity usage rate, balance cells’ depth of discharge, and realize global equalization, VI. Conctustons Departing from currently used battery equalization strate- gies, this work proposes new equalization control strategies that aim to achieve the global consistency and reduce balance costs. These new strategies offer the following distinct advan- tages: 1) The SOC estimation ofall in-pack cells can be accurately ‘obtained without prior-known cells’ capacity and resis- tanee, However, traditional SOC monitoring technologi a 200 TEE Peon is pet bot bonding eis IEE perminon Sew ‘Auhonzedicensed use led io Lnwersty of Sigh. Downloaded on une 082000 213 such asthe Kalman filtering technologies require accurate information of cach cell's eapacity in real-time. 2) The multi-layer parallel architecture is utilized to ensure the global equalization from both cell o cell and module to module, based on a master-slave BMS topology. 3) The equalization issue is formulated as a path searching problem in the directed graph and solved by the A-star algorithm, which is easy to implement, flexible and ean censure the leat equalization cost Extensive resulls verily the efficacy of the proposed RLS- based SOC observer and the proposed equalization strategy. Future research directions could be the consideration of (1) more accurate SOC estimation by treating the OCV as 4 multidimensional function of SOC, aging and hysteresis parameters, (2) scalable equalization topology and algorithm by inereasing the dimension of the path search problem, and (3) equalization strategies with a robustness guarantee against SOC estimation errors REFERENCI TQ Ouyang, W. Han, C. Zou, G. Xu, and Z. Wang, "Cell balancing ‘onto for ithium-on Paery packs: hierarchical optimal approach TEE Thun. Ind nf, vol. 16, 0. 8, pp. 5065-5075. 2018. J. Wei, G. Dong, and 2. Chen, "Remaining useful life pion and “ate o health diagnosis fr lithium-ion batries using partic ee an Sappon vecorreression” IEEE Transactions on Industrial Electonics, vol. 65. no. 7, pp. 5634-5643, 2018. G-Dong J. Wet, 7. Chen, H. Sun and X. Yo, "Remaining dischargeable tine etic for lurk beri using unsceted Kalman filter Je ower Sources, YO 368. pp. 316-827, O&t. 2017 G. Dong, I We, C. Zany and Z Chen, “Online state of charge ‘simation and open circuit voltage hysteresis modeling of LaF ePOx Thaery using invariant imbodding method” Appl Energy, vo. 162, pp 16-171, Jen 2016 [5] K Xa, HL, and Z Lia, “ISOMAPLbised spatiotemporal modeling For lithium-ion batery thecal process” ZEEE Trans. Id. If, ol. 1, 10.2, pp 509-577, 2018 [o) L Zhong, C. Zhang. ¥. He, and Z. Chen, “A method forthe esiation of the battery pac sate of charge based on in pack cells unifrity fnalsig” Appl Energy, vl. 113. pp. 558-564, Jn. 2014, (©. Ouyang. J. Chon. 3. Zheng, and Y. Hong. "SOC estimation based ‘qustsliding mode contol for osll balancing ia lihiwn fon baery focks” IEEE Tran. Ind. Elecro, vol 68, no 4, pp. 327-3436, Ape Soe D. Prod and D. Howey, “Completely decentralized active balancing Yar management system” IEEE Trans. Power Elec, vol. 33, fo Ty pp, 729-738, Jan, 201% 4 Galrdo-Lozana, E. Romero-Cadaval, M. 1. Milages Monro, and M. A. Guorero-Manincr. "Baler equalization active meth” J Poor Sources, vl. 246 no. 0, py 931 949, 2014, ¥. Shang, C. Zhang, N- Cui, an J. Gucrer, "A clone tery quale with zoo current switching and zero volage gap based ot (quahresonant Ie comerter apd boost coment” IEEE Tuas. Power ‘lec, ol. 30, 10.7, pp. 3731-8747, du. 2045 . Chazniolaou ad D. Rogers, "Cell SoC balancing using xscaded Fuld motive converte baler energy storage systems" EEE Tran. Th Elect. yl. 3, 00.9, pp. S394 S402, Sep. 2016, Dong. Wei and 2 Chen “Kala er fr onboard wate of charge «timation and peak power cpabiity analysis of lthiam-on Baltics” FE Romer Sources, Yo. 328, pp 615-626, Ot. 2016. SLL Wang, C: Lyu, and M. Pech, "Sit of chars estimation based ‘ona simpli’ electrochemical model fora single CoO, batery and Itry pack” Energy, vo. 133, pp 872-SK3, Aup. 2017 Sam and R. Xiong, “A nove see cel ia charge estimation approach for sere-connesed baery pack used inclevre vehicles” Power Soures, vl. 27%, pp. 582-854, Jan. 2015, H.'Dai, X. Wel, Z Sun. J. Wang. and W. Ga, “Online cell SOC ctimation of Lion batery packs using a dual timescale Kalman fering foc EV applications” Appl. Exergy, 257, Jal 2012, PI ff ul m ts v uo} va us us ust SB UTC komiZEE plore. Restichons spp. an strapon or mor norman.This atc bas been acepted for pobication in a future ime ofthis journal bat has not been fll edited Content may change prior to inal publication. Citation information: DO} 10.1104 20 “Tama on laa 16} ¥. Ma, P Duan, ¥, Sun, and H, Chen, “Equalization of lihium-ion user pack based on fuzzy logic contol in clecic vehicle” IEEE Trans. tnd Electron, sol. PP 0. 99. 1-1, 2018 Yeng, M. Ouyang LL J Li, Xan, and L. Xu, “On ine equal ization Tor lihimion batery packs based on charging eal voles: Part 1. equalization based on remaining chagig opacity estimation” J. Pome Sources, ol 247, pp 676-686, Feb. 2014 [18 2. Zhang, H Gu, D. Gu, ¥ Yang, and X. Rem, "A hierarchical ative alancing architecture for lihiumion batcries” IEEE Trans. Power Elona. vol. 32, no. 4 pp. 2757-2708, 2017 B. Dogg, ¥. Li, and ¥. Han, "Pall architec for battery charge ‘equalization IEEE Tran. Power Electron, vol 90, 9. 9, pp. 4908 913, Sep. 2015 Y¥. Chen, X. Liu, H. Futhy 3. Zou, and S. Yang. °A graph theoretic Framework for analyzing the speads and eliceacies of battery pack ‘cqalvation crews? Int. Elec Power Energy Syst, vol 38 Pp 15-99, June 2018 Y¥. Wang. C. Zhang. Z Chen, J. Xi, and X. Zhang. "A novel ative Squalaton meliod for liam on Dales fa eecc vehicles” Appl Energy, vl. 145, no 0, pp. 3-4, May 2015, Y.eng, M. Ouyang, LL J. Li X Hoa, and L. Xu, "On ine equa ization for lihiamion baer packs based on charging ell voles: Prt 2 fuzzy hgie equation” Power Sources, wl 247, 20.0: 460-466, Fob. 2014 (23) N-Neusen, 8. Organi, K. Na, and F Bin, “An adaptive backward ‘conta atery squiliztion system for serially connected lium ion Tattery poche” EEE Trans Wok. Technol, vl. 63, 0. 8, pp 3651 $600, Oct 2018 [241 Q. Ouyang, 1. Chen, J. Zheng, and Ht Fang, “Optimal cellxo-cel ‘ulancing topology design for serially connociod um ion butery packs” IEEE Trans. Susan. Energy, wo. 8,0. 1, pp 380-360, Jan. Son 2. Shea. H. Gui, and L. M, Tolbert “A battery cel balancing control Scheme with minimum charge vane" in 2076 TEEE Energy Conver Sion Congress and Exposition (ECCE), 2016, Conferene Proceedinss, pps [2} Fr Baroni, C. Beravesehi, L. Cassano, A. Domenic, R. Roncel and R. Sleti, "Dosign an saety verleation of 2 db charge ‘cualizer for module Lion bterics” IEEE Trnt. Ind. In v0.1, 0.2. pp 1008-1011, May 20 ¥ Les and M. Cheng, “nigent consol tatery equalization for series contac lim in bey sings.” IEEE Trans Ind. Electron, Yo. 82, no, pp. 1257-1307, Oct. 2008, B. Bole 5, Kulkarei and M, Dug, "Adaptation ofan eletrchemisty- ‘used Lion batery model to account for deterioration observed under randomized sein Proceedings of Anmual Conference of the Prgnas tie and Heath Management Soviety, Fort Worth, TX, USA, vol 29, 2014, pp. 1-9. X. Tit Z Chen, C. Zhang. and J. Wa, “A novel temperature: ‘compated model foe power Lion haters wilh dual prt iter State of charge estimation” ppl Energy, vol. 123, pp. 268-272, 2014 Zheng, ¥. ing J Shng, B.S, Rim, ad M, Poh “Inuence of liferct open cicit voltage ustson state of charge online estimation Fortius bares” App. Energy, vol 183, pp. 313-525, 2016 V1 Oveja and A. Cuadras, “State of chugs dependency of the ‘overage genre in commiesal Lion cell” J. ower Source, Sol a18, pp. 176-185, 2018 F Sun, R Xiong. and H. He, "A systematic state-of charge estimation Framework foe multrcll hatcry pack im clei vehicles using bias correction technique” Appl Ener, vol 162, pp. 399-1409, 2016. [33] 2 Zhang, X- Cheng, Z La, and D- Gu, “SOC estimation of iia jem batery pack considering balancing cument” IEEE Trunsa. Power Elecinom. Ol 33n0. 3p. 2216-2229, 2018 S. Lev. J. Kim, J. Lee, aad B. H, Cho, "Suc ofcharye and capacity ‘imation of liam om ballery using «new open ie voliage Nera Sate of charge” Power Sources, vl 188,802 pp. 1367-1373, 208 X.Chen, Hei R. Xiong, W. Shon. and Yang. A nove approach 0 "econsrut open cieut volage for state of charge estimation of Lin tom bates in electric vehicles.” Apt Energy, wal 25, p. 113738, au. 5, Dong, J. Wei, and Z, Chen, “Consrained Mayesan el iering for Ste of charge estimation ofTibium-ion bars” Iu. J. Elecie Power Energy Spt, vol 98, pp. 516-524, 2018 Zhang. G. Lit, LF, ad H. Wang, “Estination of batery tte (of chargé ith Hig observer: Applied 0 a rob or inspecting power Aranrsion hoes ZEEE Tran fd. Electrons, vo 9,92, pp OM tows, 2012 um 9) 0) en ea ps pn bs 9 bor ou a Ba bs et oa ss: (200 IEE Penal ip bt pant aon rei IBE eminion See ‘Auhonzedicensed use led io Lnwersty of Sigh. Downloaded on une 082000 213 [38] Q. Ouyang, J. Chen, and J. Zhong, “Se-tcharge observer design for atteics with online model pamcte ideation A. robust Sppmoseh” IEEE Trans. Power Electrons p. 1-1, 2019. 139] Vite and. Chon “ACV path pling hase on smoothing A* algorithm” Jat J. Ste: Eng. Appt, el 6, 0.5. pp 1-8 Sep 2018, Guangzhong Dong (S'17-M'17) eecived the [BEng dare in automation andthe PhD desree in contol science apd enginering from the Univenily ‘of Science and Techrology of China (USTO), Het, (China. 2013 and 301%, respective He is curealy a Postdoctoral Fellow withthe School of Data Science, City University of Ho Kong His curen eesearh incest incle dynamic stem modding and control, such &s modeling. es ‘mation, health prognosis of ene storage system, ‘optinal energy dispatch and coorimated contol of ined angfong Yang eccived the BEng. degree in ax ‘omation from University of Science and Technology fof China (USTC), Hefe, China, and the PAD. degree in Systems Engineering. and. Engincering Management fom Cay. Universty of Hong Kons. “Hong Kong, in 2013 apd 2017 spectively. ‘She i currently 4 Research Fellow atthe School ‘of Data Science, City Universiyof Hong Kong. Her ‘search interests include system health estimation, ‘emulning useful fe proditon, and machine lara Kook-Leung Tsui recsivad the BS. depree in chemistry andthe MLPh. ia mathematics frm Ch ‘nese University of Hong Kang, andthe PD. degree { statsics To Uaivery of Wisconsin He wan Profesor with the Schoo! of Industrial and Syems Engiccriag, Georgia Insitute of Techaology He was a rovpient of the Natal Sciace Foundation Young livestigaor Award. He is carrey the Chait Professor of industrial ngincring withthe School of Data Selene, Cy ‘University of Hong Kong. andthe Founder and he Disector of Center for Syatenm Inleematice Engecrng His cree eeearch intrest include data mining, surveillance ia healthcare snd public health prognosis and systems heh management, calibration and validation of ompuler model, process congol and moniworing. and robust design and ‘Taguchi methods He ia fellow ofthe American Statistical Associaton, the Amica ciety for Quality, and the Ineratonal Society of Engineering Asset Management fda US. representative to the Intemational Organisation for Sundar ain ‘Technical Comiittr on Saisie! Methods, He was he Chiro he Hite for Operations Research and the Mamagemeat Sciences (INFORMS) Section ‘08 Qual, Sats, and Reliability andthe Founding Chair ofthe INFORMS, Section ow Data Mining. (Changin Zou (M16) weesved his BE degree in ulomesive eaginceing Tom the Beijing lstiate Of Technology. Bejing. China and PRD. degree {in automation and congo! enginering. rom the University of Melbourne, Veta, Asai ia 2011 and 2017, respetiely He as a ising rscarcher with the Encry, CConuols Applications Lab, University of Cal fomia, Berkcley, USA, from 2015 to 2016, Since 2017, he has joined the Automatic Control group a Chalmers University of Technolo, Gothen, Sweden, where he was a pontctrl recachr and eure fan sian Professor His research focdses on intelgent management of energy stems td eletie vehicles, Dr. Zou recived the Swodish Research Council Suing Grant and the Marie Curie Individual Fellowship. He serves a an asacate editor of IEEE Transactions on Transportation Electrification, SB UTC komiZEE plore. Restichons spp. an strapon or mor norman.