System on Chip Battery State Estimator: E-Bike Case Study

Transcription

1 System on Chip Battery State Estimator: E-Bike Case Study Rocco Moreo, Roberto Di Rienzo, Federico Baronti, Roberto Roncea and Roberto Saetti Dip. di Ingegneria de Informazione, Università di Pisa, Pisa, Itay Abstract This paper discusses the hardware impementation and experimenta vaidation of a mode-based battery state estimator. The mode parameters are identified onine using the moving window east squares method. The estimator is impemented in a fied programmabe gate array device as a hardware bock, which interacts with the embedded processor to form a system on a chip battery management system (BMS). As a case study, the BMS is appied to the battery pack of an e-bike. Road tests show that the impemented estimator may provide very good performance in terms of maximum and rms estimation errors. This work aso proposes a new methodoogy to assess the performance of a battery state estimator. I. INTRODUCTION Lithium-ion batteries aow the impementation of ight onboard energy storage systems (ESS), thanks to their high power and energy densities. The ESS must be equipped with a battery management system (BMS) in order to ensure a reiabe and safe operation of the battery. The BMS monitors and contros the charge and discharge phases by estimating the battery state. The most important state variabes are the battery state of charge (SOC) and state of heath (SOH). SOC is reated to the residua charge stored in the battery and SOH quantifies its degradation in terms of capacity fading and the increase of the interna resistance [1], [2]. The BMS executes agorithms that use quantities directy observabe on the ces, i.e., their votage, current and temperature. Many agorithms have been proposed for SOC estimation, starting from the Couomb Counting (CC), which is based on the integration over time of the battery current. The integration resut is normaized with respect to the ce capacity. The CC must be initiaized with the correct vaue of the SOC and the current must be measured by a high accuracy sensor, in order to avoid the intrinsic divergence in the integration. Another simpe method is based on the open circuit votage (V OC ) measurement. In fact, V OC mainy depends on SOC and sighty on other factors, such as temperature, ageing and current rate [3]. However, the ce votage reaxes to V OC a ong time after the ce current is interrupted. Thus, this technique cannot be used in highy dynamic systems [4]. High accuracy and reiabiity are required in automotive appications, in order to guarantee the safety standard and a good estimation of the vehice autonomy. Furthermore, the agorithms must be executed in rea-time in an embedded system, ike the BMS. Mode-based agorithms, such as the Extended Kaman Fiter (EKF) [5], [6], the Mix Agorithm [7] and the Adaptive Mix Agorithm (AMA) [4] are adopted. A mode-based technique is a cosed-oop method in which the mode is used to correct the estimated SOC, by comparing the votage predicted by the mode with the measured one. Various kinds of mode can be used [8], but an Eectrica Circuit Mode (ECM) has some benefits when used in a rea-time embedded system [9]. The accuracy of the SOC estimate depends on the ECM capabiity to predict the ce votage, in a the operating conditions and for the entire battery ife. A good approach is to track the variation of the operating condition onine, which is the soution adopted in the EKF and AMA techniques. The AMA uses an enhanced version of the Mix Agorithm, in which SOC estimation is corrected by using an ECM where the parameters are updated onine by using the Moving Window Least Squares (MWLS) method [1], [11], instead of taking them constant, ike in the origina agorithm. The performance of this agorithm has been evauated by comparing it to the EKF method [4], [12], which is the most used in this fied. The comparison, carried out with current and votage profies obtained from the Urban Dynamometer Driving Schedue (UDDS) driving cyce, defined by the U.S. Environmenta Protection Agency [13], shows that both methods provide good estimation of both parameters and SOC. Both EKF and AMA are suitabe for an embedded system impementation. An attractive soution for industria and automotive appications is the use of a ow-cost fied-programmabe gate array (FPGA). The estimators can be impemented on the FPGA as hardware acceerators beside the processor, which is fuy dedicated to the other BMS functionaities. A first hardware impementation of the non adaptive version of the Mix agorithm is reported in [14]. A hardware-in-the-loop (HiL) simuation framework used to test and compare the AMA and EKF, impemented in hardware on an FPGA, is presented in [15]. In this framework, a muti-ce battery is simuated using an accurate mode. An experiment not common in iterature that can provide fina resuts is to test the estimators in a rea appication. This aows an effective measurement of the estimator performance, as we as the anaysis of the issues arising from its integration in a rea system. The aim of this paper is to present the resuts obtained using the AMA in a rea appication, i.e., on an e- bike. A simpe BMS provided with the AMA battery state estimator has been impemented as a system on chip on an Atera Cycone V. The agorithm detais and the hardware/software partitioning of the system are discussed in Section II. The experimenta testbed description is reported In Section III, whie Section IV describes the methodoogy appied to assess

2 > d Q n + il > W / Z d Z d Z d d Z d d d Y D s K >hd s K ^K Fig. 1. Adaptive Mix Agorithm bock diagram. 1V 1V SOC - Z d EZ Z V OC + - v n Fig. 2. Eectric circuit mode. ^K the estimator performance. The resuts obtained from road tests are discussed in Section V and finay concusions are drawn in Section VI. II. STATE ESTIMATOR A. The Adaptive Mix Agorithm The AMA is based on the combination of the Mix Agorithm and the MWLS method used with the AutoRegressive exogenous (ARX) structure of the ECM. A basic bock diagram of AMA is shown in Fig. 1. The Mix Agorithm is a simpe mode-based SOC estimation method (bue square in Fig. 1), which uses a constant parameter ECM in order to correct the estimated SOC variabe. A genera representation of the ECM is shown in Fig. 2 and a singe RC branch version is used in this work. The votages v n on the RC branches mode the reaxation effects. A singe RC group reduces compexity maintaining good accuracy, particuary when fast transients are dominant. The ce termina votage v M is generated as a sum of the votage v 1, the open-circuit votage V OC and the votage across the interna resistance R i L. The V OC -SOC reationship is modeed as a controed generator V OC = f(soc), impemented with a Look-up Tabe (LUT). The eft-hand side of the ECM modes the ce capacity and SOC. The atter is expressed as Q/Q n, where Q is the residua ce charge and Q n is the normaization vaue corresponding to the maximum charge that can be stored in the ce, expressed in Couomb. i L + v M - The mode ce votage v M is subtracted to the measured ce votage v T yieding an error signa, which is firsty ampified by the observer gain L and then used to correct the measured ce current i L. Finay, SOC is the resut of the integration over time of the corrected current signa. The observer gain L is chosen to strongy reduce the effects due to bad SOC initiaization and current measurement uncertainty [14]. The estimation accuracy depends on the capabiity of the ECM to reproduce the ce behaviour. For this reason, the ECM parameters are updated onine by the Parameter Identification bock, which impements the MWLS agorithm. In this way, the agorithm takes into account the parameter variations due to the battery ageing, the operation conditions (as ce SOC and temperature) and the manufacturing process toerances. The ARX structure can be obtained starting from the ce time-domain state space mode: dsoc = i L dt Q n dv 1 dt = v 1 + i L (1) τ 1 C 1 v M = V OC R i L v 1 where τ 1 = R 1 C 1. The ce operating point sowy changes over time, so that the mode can be inearised around it, considering the parameters constant in the identification window. The V OC -SOC reationship is thus approximated by a piecewise inear curve V OC = α + α 1 SOC, where α and α 1 depend on the operating point. The discrete-time transfer function is obtained by the appication of the biinear transform to the transfer function from the current input to the ce votage output, as obtained from (1). Then, the discrete-time reationship between the input and output sampes becomes: y(k) = a 1 y(k 1) a 2 y(k 2) + α (1 + a 1 + a 2 ) + b u(k) + b 1 u(k 1) + b 2 u(k 2) which is a second order ARX mode of the ce. The coefficients of the ARX mode depend on the ECM parameters and we can simpy verify that 1 + a 1 + a 2 =, so (2) is reduced to: y(k) y(k 2) =a 1 (y(k 2) y(k 1)) + b u(k) + b 1 u(k 1) + b 2 u(k 2) Eq. (3) is used to buid an overdetermined system by using the current and votage sampes within a given time window. This window is shifted in time to track the parameter variations during the ce operation. In this way, a different system is obtained at every shift, which is soved by the Least Squares (LS) method. Finay, the resuting vector [a 1, b, b 1, b 2 ] is used to compute the ECM parameters [R, R 1, C 1 ]. B. Hardware impementation The AMA is fuy deveoped as a hardware modue and impemented on an Atera Cycone V SoC 5CSXFC6D6F31C6N device, by using the Atera SoCkit deveopment board. The hardware impementation enabes a high (2) (3)

3 TABLE I E STIMATOR R ESOURCE U SAGE AND P ERFORMANCE Logic utiization (in ALMs) Variabe-precision DSP Bock Memory bits Execution time 17244/4191 (41 %) 36/112 (32 %) 723 Kb/553 Kb (13 %) 61 µs (@5 MHz) throughput and ow atency estimation in an embedded system, in spite of the computationa compexity of the agorithm. The agorithm is described as a Simuink mode and then automaticay synthesized to a ow-eve hardware description, by using the Atera DSP Buider too. This deveopment environment optimizes the hardware modue performance and resource usage depending on the too parameter settings. The modue is aso provided with a Memory Mapped (MM) interface. The atter consists of input and output registers which are used by the companion processor to write the agorithm inputs and to read the computed resuts. The estimator FPGA resource usage and execution time are shown in Tabe I. The time required to estimate both SOC and the parameters, for a cock frequency of 5 MHz, is very short. Therefore, the BMS is capabe of estimating the state of a arge number of ces by using the estimator in Time Division Mutipexing (TDM) across mutipe ces. This feature is very usefu in high-power appications, where the batteries consist of hundreds of ces. (a) (b) III. C ASE STUDY Fig. 3. Photographs of the demonstrator setup V OC (V) The demonstration system is buit on an eectricay assisted bike, equipped with a 1 ce battery pack. This soution aows us to verify the estimator on more than one ce in rea road tests. Two photos of the demonstrator are shown in Fig. 3. The eectric bike is an Ataa E-Scape (see Fig. 3(a)), a city bike with auminum frame. The bike is equipped with a brushess DC motor with a nomina power of 25 W. The battery pack consists of 1 Lithium-ion series-connected ces, yieding a nomina votage of 36 V and a rated capacity 1 A h. The entire setup, incuding the battery pack, was arranged on the bike rack, as shown in Fig. 3(b). A 14-pin connector aows us to access the battery ces terminas. A Ha effect current sensor from Aegro Microsystems is used to measure the battery current. Its fu scae range is ±12.5 A, with an interna resistance of 1.2 mω and a sensitivity of 11 mv/a (for a suppy votage of 3.3 V). The ces and current sensor output votages are acquired by a DC1894B demo board. This Linear Technoogy board is equipped with an LTC684 battery stack monitor capabe of measuring up to 12 series connected ces with a tota measurement error of ess than 1.2 mv. The LTC684 is aso provided with two auxiiary anaog inputs, which are used to accuratey acquire the output and suppy votages of the current sensor. A the measurements take 11 ms to compete. The battery was preiminary characterized in the aboratory with a Pused Current Test (PCT) profie at 25 C [4]. The input current and the output votages during the test are used to determine the AMA initiaization parameters, i.e., SOC (%) Fig. 4. Open circuit votage VOC, as a function of SOC. the VOC -SOC reationship, the capacity of the ces and the initiaization vaues of the ECM parameters. The VOC -SOC curve is considered invariant for every ce [3] and it is shown in Fig. 4. The extracted ce capacities are very simiar to each other and are around the rated vaue of 1 A h, except for the second ce which has a 9.4 A h capacity. This resut, together with the average vaues of the ECM parameters (R = 24.5 mω, R1 = 13.3 mω and C1 = 6945 F), are used to initiaize the agorithm impemented on the FPGA. The Aegro current sensor, the DC1894B and SoCkit boards form a basic BMS, usefu to test the hardware estimator. Fig. 5 shows a bock diagram of it. The SoCkit Cycone V SoC FPGA contains a dua-core ARM Cortex -A9 Hardware

4 d dd Fig. 5. Basic BMS bock diagram. Processor System (HPS). When the board is turned on, the Linux Ångström distribution, based on a rea-time kerne, is booted from a partition on the mini SD card and the main appication is aunched. The appication manages the acquisition of the ce data, their ogging and the communication with the user interface. The current and votage sampes are read from the battery stack monitor connected to an SPI port. The sampes are then sent to the estimator, impemented on the FPGA, via the MM interface, and the computation is started. The operation is cycicay executed every.1 s, yieding an acquisition sampe rate of 1 Hz. When a the ce states are computed, the appication saves the sampes and the estimation resuts in a og fie, which can be downoaded directy from the SD card or by using the Secure Copy Protoco (SCP) through the Ethernet port. The user interacts with the system via the push buttons on the board. Other data are shown on the LCD dispay and LEDs on the deveopment board. The bike speed, usefu for a better anaysis of the data, is captured by a GPS unit. IV. VALIDATION METHODOLOGY As a vauabe contribution of this work, the impemented SOC estimator is vaidated in fied tests, i.e., riding the e-bike as in every day use. To evauate the SOC estimation error, the SOC estimate needs to be compared with a reference vaue [16]. The common approach is to use the CC technique to obtain the SOC reference vaue, as shown in (4). SOC ref (t) = SOC 1 t i L (τ)dτ (4) Q n Eq. (4) can provide a reiabe SOC reference SOC ref, assuming that the initia SOC vaue SOC, the normaization capacity Q n and the battery current i L are known precisey. The battery current shoud therefore be acquired by a highy accurate current sensor, different from the one used within the BMS, which feeds the SOC estimator. This is an important aspect, often negected in iterature, which makes it possibe to assess the robustness of the estimator against errors in the current measurement effectivey. In this work, i L in (4) is measured by a Nationa Instrument (NI) 9227 modue mounted on a NI cdaq-9178 chassis. The atter is aso equipped with three NI 9215 modues, which are used to acquire the votage of the 1 ces and to verify the LTC684 stack monitor behavior (see Fig. 3(b)). The cdaq-9178 is powered by a suppementary battery and connected to a aptop, carried by the cycist in a backpack, which runs a LabVIEW appication for data acquisition and ogging (see Fig. 3(a)). SOC is taken equa to 1 %, as tests are started after a fu charge of the battery. Q n is assumed to be equa to either the rated battery capacity or the capacity measured in a preiminary characterization test. These assumptions however may ead to inaccuracies in the computation of SOC ref. Inhomogeneities in the ces of the battery pack can cause not a the ces to be fuy charged to 1 % SOC. Moreover, the maximum charge that can be extracted from each ce during a vaidation test (we refer to this vaue as the rea capacity) may differ from that measured during a characterization test, because of different discharge rates and temperature. To overcome these probems, we propose to combine the CC and V OC estimation techniques for computing SOC ref. These two techniques are used to define two aternative metrics for assessing the SOC estimator errors [16], whereas here they are merged in a unique definition of SOC ref. The basic idea is to set SOC and Q n in (4), so that SOC computed by (4) coincides with the estimate obtained through V OC at the beginning and end of the vaidation test. This does not require any specific condition during the execution of the test, apart from the fact that the test starts with the battery in a rested state and ends in an SOC region where it is possibe to reiaby extract SOC from the V OC -SOC reationship, once the battery has reached the steady state after discharge. Let V OC = f(soc) express the V OC -SOC reationship shown in Fig. 4. For each ce, we can compute SOC before and after the vaidation test by inverting the V OC -SOC reationship, i.e., SOC = f 1 (VOC start) and SOC end = f 1 (VOC end start ), where VOC and VOC end are the ce votages measured, when the battery is in a steady state, before and after the vaidation test, respectivey. Given the above definition, we can compute the rea ce capacity reated to a vaidation test, which coincides with the desired normaization capacity to be used in (4), by the foowing equation [17]. tend t Q rea = i L (t)dt (5) SOC SOC end

5 P e (W) Speed (Km/h) SOC (%) Ce 1 Ce 2 Ce 3 Ce 4 Ce 5 Ce 6 Ce 7 Ce 8 Ce 9 Ce 1 Fig. 6. Eectric power and speed during the discussed vaidation test. Capacity (Ah) Ce Rated PCT Rea Fig. 7. Capacities of the ces extracted using the PCT and the discussed vaidation test. V. EXPERIMENTAL RESULTS Severa road tests have been carried out, proving both the functionaity of the basic BMS provided with the AMA estimator and the effectiveness of the designed experimenta set-up. As a representative exampe, we discuss the resuts obtained during a vaidation test, consisting of the repetitions of a cycing period foowed by a rest pause. The eectric power P e provided by the battery pack and the speed measured by the GPS data ogger during the test are shown in Fig. 6. First of a, we compute the initia SOC and the rea capacity of each ce of the battery pack from the data acquired by the NI cdaq-9178 chassis. Then, we evauate the corresponding SOC reference using (4). Fig. 7 shows the rea ce capacity, for the discussed test, and compares it to the vaue obtained in the PCT characterization test. Apart from the second ce, which presents a pronounced degradation of the ce capacity, there is ony a sight difference between the PCT, rated and rea ce capacities, as expected by the high Couombic efficiency provided by the Lithium-ion battery technoogy. The computed SOC ref (t) for each ce is reported in Fig. 8, where it is evident that the SOC of the second ce corresponds to the battery SOC, as the second ce is the first ce to reach the discharge cut-off votage. With the avaiabiity of the SOC reference, we can cacuate the estimation error introduced by the AMA. The AMA SOC estimate for the second ce is shown in Fig. 9 and is very cose to the reference vaue during a the test. On the contrary, the pure uncompensated CC drifts over time. This Fig. 8. SOC behavior of the 1 ces during the discussed vaidation test. SOC (%) Ref AMA CC Fig. 9. Comparison of the SOC estimation resuts for the second ce. ceary demonstrates that the AMA is capabe of correcting the uncertainties in the current measurement (especiay the offset of the current sensor) and in the rea capacity of the ce. Tabe II shows the maximum and rms errors of the AMA and CC SOC estimations. The AMA provides a more reiabe estimation than CC, being its rms error beow 2 % and the maximum absoute error around 3 %, for a the battery ces. This is a vauabe achievement, as it is obtained in road tests, when the battery is subjected to the rea appication oad. In fact, if we consider the SOC estimation errors reported in [6] for a simiar appication, AMA outperforms EKF, as the atter introduces a maximum SOC error we above 5 %. Better performance in terms of maximum error is obtained in [6] using an Adaptive EKF, at the expense of higher compexity. It is worth noting that the resuts reported in [6] refer to aboratory tests and not to road tests, as in this work. Finay, Fig. 1 shows the R and τ 1 ECM parameters identified by the AMA for the second ce. Whie the R identification is robust against the choice of the window ength, τ 1 identification seems to be sensitive to this choice. The resuts reported in Fig. 1 refer to a window ength of 2 min and an LS matrix with 3 rows. This impies that the ECM

6 R (mω) τ 1 (s) TABLE II SOC ESTIMATION ERRORS AMA CC Max (%) rms (%) Max (%) rms (%) Ce Ce Ce Ce Ce Ce Ce Ce Ce Ce Fig. 1. ECM parameters estimated onine by the AMA for the second ce. parameters are updated every 4 s. Even if the windows in which the battery current is amost constant are discarded, the identification of τ 1 is ess reiabe, when the window overaps a rest period. VI. CONCLUSIONS This work has demonstrated the impementation of a battery state estimator in an industria FPGA. The estimator is based on the AMA and is impemented as a hardware acceerator bock, which interacts to the processor embedded in the FPGA to form a system on a chip BMS. Thanks to its effective design, the estimator modue can be used to estimate the state of a the ces in the battery pack in a time mutipexing fashion. As a case study, we have integrated the BMS into the battery pack of an e-bike. This aowed us to vaidate the estimator in a rea appication scenario, which represents a significant achievement, as amost a the iterature on this topic reports on experimenta resuts obtained in aboratory tests and for a singe ce. The road tests carried out have shown the effectiveness of the designed BMS and experimenta set-up, and that the AMA estimator provides a very good estimation comparabe or even better than the we estabished EKF. ACKNOWLEDGMENT The authors woud ike to thank Aessio Tarabori and the System Soution Engineering group of Atera now part of Inte for the support and fruitfu discussion on the topic of this paper. This work has been funded by the ECSEL Joint Undertaking under grant agreement No (3CCar project). REFERENCES [1] H. Rahimi-Eichi, U. Ojha, F. Baronti, and M.-Y. Chow, Battery Management System: An Overview of Its Appication in the Smart Grid and Eectric Vehices, IEEE Ind. Eectron. Mag., vo. 7, no. 2, pp. 4 16, Jun [2] Dai Haifeng, Wei Xuezhe, and Sun Zechang, A new SOH prediction concept for the power ithium-ion battery used on HEVs, in 29 IEEE Veh. Power Propus. Conf. IEEE, Sep. 29, pp [3] B. Pattipati, B. Baasingam, G. Avvari, K. Pattipati, and Y. Bar-Shaom, Open circuit votage characterization of ithium-ion batteries, J. Power Sources, vo. 269, pp , Dec [4] R. Moreo, W. Zamboni, F. Baronti, R. D. Rienzo, R. Roncea, G. Spagnuoo, and R. Saetti, Comparison of State and Parameter Estimators for Eectric Vehice Batteries, in IECON st Annu. Conf. IEEE Ind. Eectron. Soc., 215, pp [5] G. L. Pett, Extended Kaman fitering for battery management systems of LiPB-based HEV battery packs: Part 3. State and parameter estimation, J. Power Sources, vo. 134, no. 2, pp , Aug. 24. [6] C. Taborei and S. Onori, State of charge estimation using extended Kaman fiters for battery management system, in 214 IEEE Int. Eectr. Veh. Conf. IEEE, Dec. 214, pp [7] F. Codeca, S. M. Savaresi, and G. Rizzoni, On battery State of Charge estimation: A new mixed agorithm, in 28 IEEE Int. Conf. Contro App. IEEE, Sep. 28, pp [8] V. Ramadesigan, P. W. C. Northrop, S. De, S. Santhanagopaan, R. D. Braatz, and V. R. Subramanian, Modeing and Simuation of Lithium-Ion Batteries from a Systems Engineering Perspective, J. Eectrochem. Soc., vo. 159, no. 3, pp. R31 R45, Jan [9] M. Chen and G. Rincon-Mora, Accurate Eectrica Battery Mode Capabe of Predicting Runtime and I-V Performance, IEEE Trans. Energy Convers., vo. 21, no. 2, pp , Jun. 26. [1] H. Rahimi-Eichi, F. Baronti, and M.-Y. Chow, Modeing and onine parameter identification of Li-Poymer battery ces for SOC estimation, in 212 IEEE Int. Symp. Ind. Eectron. IEEE, May 212, pp [11], Onine Adaptive Parameter Identification and State-of-Charge Coestimation for Lithium-Poymer Battery Ces, IEEE Trans. Ind. Eectron., vo. 61, no. 4, pp , Apr [12] F. Baronti, W. Zamboni, N. Femia, H. Rahimi-Eichi, R. Roncea, S. Rosi, R. Saetti, and M.-Y. Chow, Parameter identification of Li-Po batteries in eectric vehices: A comparative study, in 213 IEEE Int. Symp. Ind. Eectron. IEEE, May 213, pp [13] Urban Dynamometer Driving Schedue (UDDS). [Onine]. Avaiabe: [14] F. Baronti, R. Roncea, R. Saetti, and W. Zamboni, FPGA Impementation of the Mix Agorithm for State-of-Charge Estimation of Lithium- Ion Batteries, in IECON 214-4th Annu. Conf. IEEE Ind. Eectron. Soc., 214, pp [15] R. Moreo, F. Baronti, X. Tian, T. Chau, R. D. Rienzo, R. Roncea, B. Jeppesen, W. H. Lin, T. Ikushima, and R. Saetti, Hardware-in-the- Loop Simuation of FPGA-based State Estimators for Eectric Vehice Batteries, in to be Present. ISIE 216, 216. [16] B. Baasingam, G. Avvari, K. Pattipati, and Y. Bar-Shaom, Performance anaysis resuts of a battery fue gauge agorithm at mutipe temperatures, J. Power Sources, vo. 273, pp , Jan [17] M. Einhorn, F. V. Conte, C. Kra, and J. Feig, A Method for Onine Capacity Estimation of Lithium Ion Battery Ces Using the State of Charge and the Transferred Charge, IEEE Trans. Ind. App., vo. 48, no. 2, pp , Mar. 212.