EV Battery State of Charge: Neural network based estimation A. Affanni, A. Bellini, C. Concari, G. Franceschini, E. Lorenzani, C. Tassoni Dipartimento di Ingegneria dell’lnformazione University of Parma Parco Area delle Scienze, 181iA 43100 Parma, ITALY A b m a n Different Electric Vehicles (EV) types have been recently developed with the aim of solving pollution problems caused by the emission of gasolinepowered engines. Environmental Considerations promote the adoption 01EV for urban transportation. As it is well known one of the weakest points ofelectric vehicle is the battery system. Vehicle autonomy and therefore accurate detection of battery state of charge are among the main drawbacks that prevent he spread of electric vehicles in the consumer market. This paper deals with the analysis of battery state of charge: performances of B few sizes of batteries are analyzed and their state of charge is estimated with a Neural Network (NN)bared system. The obtained results have been used to design a ion. lithium battery pack suitable lor electric vehicles. The proposed System presents high capability of energy recovering in braking conditions, together with charge equalization, over and under voltage protection. Moreover a Neural Network based estimation of battery state of charge has been implemented in order to optimize autonomy instead of perfarmanas or viceversa depending on journey. simple and effective implementation. Experiments were made on different discharge profile modeled on the typical urban EV duty cycle. Then a battery pack has been designed and realized. It performs battery equalization and protection in order to performenergy recovering during vehicle braking operation. Moreover battery State of charge is continuously monitored allowing different control strategy depending on joumey or dnver requirements. These battery packs will he s e d to supply a high performance electric scooter [SI. ~ 1. NTRODUCTION Different Electric Vehicles (EV) types have been recently developed with the aim of solving polluCon problems caused by the emission of gasolinepowered engines [l]. As far as overland transports are concemed large range requirements lead to hybrid vehicles solution while environmental considerations promote the- adoption of EV for urban transportation [2]. As it is well known one of the weakest points of electric vehicle is the battery system. Vehicle autonomy and therefore accurate detection of battery State of Charge (SoC) are among the main drawbacks that prevent the spreadofelechic vehicles in the consumer market [3], [4]. The aim of the paper is to obtain a simple and reliable method to estimate-the battery state of charge using only the electrical quantities usually available. State of charge information allows tuning control strategy according with journey characteristics. This allows for example the optimization of battery lifespan instead of useless performances in case of urban journey. The proposed method includes a neural network based estimator in order to comply with the non linear behavior of the battery and to provide a 11. BAITERIES MODELIZATlON FOR EV The experiments are referred to Lithium batteries whose main features are briefly reviewed in the following. Lithium batteries can be divided into: LPB (Lithium polymer) batteries, and Ion Lithium batteries. The formee feature the best energy density values. Specifically they provide 170 W K g , however they are very expensive. The authors chose Ion lithium cells nevertheless the lower energy density and the stronger self discharge, because of their cost g and their weight. Specifically they provide 140 M (Cobalt), 120 Wh/Kg (Manganese). During the experiments the following sizes of Lithiumlon battery cells were used: 3.7 V, 1.5 Ah and 3.7 V, I O Ah. The battery voltages are constrained during discharging operation to a minimum level of 2V in order to prevent capacity loss. As already stated, batteries are one among the weakest part . of an electric vehicle. Limited autonomy and recharge facilities are some of the main drawbacks. Therefore exploiting the battery up to its end is a key element, and an accurate howledge of current State of Charge is needed. Accurate estimation of SoC is a complex task. To accomplish it a few approaches are possible: circuit nodels, empirical models, statistical or A.I. based models [61, [71, [SI,[9]. Circuit models require a very heavy job for each type of battery, and require noElinear elements. Classical methods -for SoC estimation include: measurements of the extracted charge, measurements of the battery intemal impedance or resistance, measurements of the battery no-load voltage. Though none of them provides acceptable results, a hybrid method, including measurements of intemal impedance or resistance &), extracted charge and no-load voltage a) 0-7803-7817-2/03/$17.W 6?W3 IEEE 684 Authorized licensed use limited to: Universiti Tunku Abdul Rahman. Downloaded on March 11,2021 at 02:42:09 UTC from IEEE Xplore. Restrictions apply. Vim), can achieve a better estimation of Soc. Fig. 1 refers to the 1.5Ah battery cell, it shows the R, profile as a function of time: the profile is almost constant during battery discharge increasing suddenly close to the end of discharge. Therefore this profile, alone, is not sufficient to evaluate the SOC. Fig. 2 refers to the lOAh battery cell, it shows a discharge profile as a function of time at constant discharge current can be computed using this equal to the battery capacity. profile howeverthis quantity is not sufficient to state SOC. The knowledge of the amount of energy that can be extracted from battery is a key element in order m optimize system performances. This information can be obtained starting from SoC only if battery discharge status is fully defined. A battery can be defined completely discharged when it cannot provide the power needed by the system it supplies. Therefore an accurate estimation of S o C should be made only fully discharging the battery. This procedure is impractical because of the large amount of time required. The authors’ assumption is that SoC decreases linearly with time ifdischarge occurs with a periodical current profile. This simple approximation was then verified a posteriori by laboratory tests. Starting from this assumption the Authors propose a NN based procedure based that allows the estimation SoC mapped as a functionf of fem, Qe audRi . a .. ., .. .. .. .. .. . ... . ... ... . ... . ..., ... . .. .. .. .. 0 50 I K b I M O 2b.l 2.503 30% 3504wo45W r a [SI Figure 2: lOAh battery cell discharge profile. As usual the neural network estimation is more accurate if the training set is large and fully representative of different operating conditions. Therefore the behavior o f f was analyzed in several different conditions, performing several different discharge cycles. Specifically the test set-up must allow discharging the battery cell with a given current to measure cell voltage and current during discharge. Fig. 3 shows a current profile chosen accordingly to the typical E V urban duty cycle for a single cell. The test set-up for the experiments was made by a Programmable current Generator HP6684A, IOOA 60V; a Data-logger HP 3 4 9 7 0 4 a host PC with IEEE-488 interface and suitable LabView software. Two GUls were designed in LabView which allow to specify the desired current profile during battery discharge, s e t u p the measurements, store collected data on the host PC. Starting from measured data (battery current and voltage during discharge) a MATLAB script is used to produce the training set. Qe is computed with the composite trapezoidal approximation from current sampling. Ri is computed measuring the voltage drop A V corresponding to a current variationdl: R( =& !.! Time [sec] AI Figure 1: 1.5Ah battery cell Rprofile. 111. QATE OF CHARGE ESTrmATlON The aim is to find a function f so that SoC=f(lem,aRi ). As stated before fcan be obtained from an analytical study of physical and chemical behavior of batteries, or it can be mapped from empirical data. The latter approach tums out to be preferable, since it is simpler and often more efficient. In this paper an empirical estimation o f f was made, using a neural network, which maps the input vector: /em, R, into the state of charge (SoC) [9], [IO]. a Finally feemwas estimated directly as the voltage with noload in time intervals where the current is zero. The SoC values suitable for NN training are obtained in accordance with the previous hypothesis on cell discharge behavior. N. EXPERIMENTAL RESULTS The collected training data were used to train a two-layer neural network with tansig activation function. The leaming procedure uses the BFGS algorithm (Broyden, Fletcher, GoldfarbeShanno)[lO],[ll]. 685 Authorized licensed use limited to: Universiti Tunku Abdul Rahman. Downloaded on March 11,2021 at 02:42:09 UTC from IEEE Xplore. Restrictions apply. 0.76 ,................................................................................... Amps i ,,&I 1 1 1 I 0.755 0.75 0.745 -.n.-* 0.74 Cycles 1111....................................................... 0 10 Ec( E? "^ ...................... m M. . Timslsecl :... CO 70 Figure 3: Typical discharge current profile as a function of time for a single cell The trained NN was then used as a real time estimator of battery state of charge. To validate the estimation results different tests were performed changing the discharging conditions. Specifically the current profile has been varied during every validation test in order to prove the NN performances in operationsclose to'the real ones. Each test was stopped after a consistent number of cycles, the banery was fully discharged and the measure of the extracted charge was compared with NN output. The two different IonLithium battery cells detailed in section I1 were used. As an example Fig. 4 shows the estimated normalized SoC compared with the current state of charge during the discharge of the smallest Ion-Lithium battery. The curves are almost completely overlapped, therefore in Fig. 5 an enlarged version of the same discharge process is reported. Figure 5 Neural network estimation (light grey) compared with current SOC. Several discharge cycles were repeated with the described test set-up, obtaining a good agreement between neural network estimation and measured results. Specifically tab. 1 reports the results in terms of Mean Square errors and maximum errors for different tests performed with different discharge current profiles. Test I 1 I 2 3 4 5 MeansquareError 2.9R71r-4 5.0878e-3 2.9377e-3 4.2381e-3 3.2647e-3 I . Maximumemr I on46 0.031 0.030 0.024 0.032. . . . . . . . . . . . . . . . . . . ................. .. . .. .... .... . ... . . . ................................ . . . . . . . .. .. ........... ............. . . .. . . . . .. ... ... . .................. . .. .. ... .. .. . . . OD 81 w 1 - Em a0 . . ....... .. . . ,on ran I Cycles , la0 Figure 4 Neural network estimation compared with current SOC 686 Authorized licensed use limited to: Universiti Tunku Abdul Rahman. Downloaded on March 11,2021 at 02:42:09 UTC from IEEE Xplore. Restrictions apply. among the ten cells to a DSP by means of its output impedance. The block diagram in fig. 6 shows the structure of the board. 2.4" 3.2v 3.6" Cdl "oh$* I y O h , ] 1" 4.4v 4.8V Figure 7 Optocouplers input voltage as a function of voltage across the cell. Figure 6 Block diagram of the battery pack control board. Normal operating voltage range of each cell is [3V, 4.15V1. When the voltage exceeds this range undervoltage and overvoltage protection circuits start to operate by sending the actual cell voltage level to the control by the optocoupler. Specifically in overvoltage operation range [4.15V, 4.25V], overvoltage circuits produce a linear voltage depending on the chargestate of the relevant cell. With a dual behavior in undervoltage operation range [2.7V, 3V], undervoltage circuits produce a linear voltage depending on the chargestate of the relevant cell. When two or more cells are operating in a non-optimal voltage, the control board acts as an e x a and sends the output of the sensing circuit which shows the worst case of chargestate. The temperature sensor guarantees that the battery pack temperature does not overcome 7O0C, in order to prevent the damage of the cells. The outputs of the above mentioned protection circuits are sent through a proper interface (implemented with an a m y of optocouplers) to the motion control DSP allowing tuning control strategy according with journey characteristics. The chargPequalirers allow the maximum energy storage in the battery pack shunting the current of the cell at the threshold of overvoltage during the charge. It is remarkable to observe that the circuits which sense the overvoltage and undervoltage state do not suffer from commommode voltage; so that each sensor observes the voltage across the relevant cell independently of the position of the cell inside the pack. Extensive simulations were performed to validate the behavior of the protection circuits. Fig. 7 shows the simulation of the input voltage of the optocouplers as a function of the voltage across a cell. Simulations were performed with a DC input sweep between 2.4 V and 4.8V superimposed by a ripple which models the disturbance caused by the current requested by the motor. 2.w The input of optocoupler is at high level if each cell is in the optimal voltage range [3V, 4.15VI and falls with linearity if the voltage across a.cell exceeds this range. The circuit does not suffer from the disturbance introduced in the simulation, as the output is function of the average input. The slopes of,overvoltage and undervoltage transitions are different, because of the different priorities. In fact when a cell is in overvoltage condition the DSP must stop the charge at constant current starting the charge at constant voltage. On the other side when a cell is in undervoltage condition the DSP must reduce the current requested by the user providing less acceleration in order to,guarantee the best autonomy of the vehicle allowing a larger margin ofbattery-safety. The choice of the components of the control circuit was driven by the needs of low-power consumption of the board (900mW). Because of the large number of devices, they have been placed onto two faces into a four layers PCB. The conformation and the size of the board are related to the size of the battery pack composed by ten cells, in fact it underlies the board in order to reduce the pack volume. The size of the board is 201mm x 54mm (7.91 x 2.12 inches) and a global view is reported on fig 8. Figure 8 Control board layout The pads which provide the voltage of each cell have been placed in order to reach the cells connectors by screwed bars to make easier the assembly. The structure of the battery pack is shown in fig 9. 687 Authorized licensed use limited to: Universiti Tunku Abdul Rahman. Downloaded on March 11,2021 at 02:42:09 UTC from IEEE Xplore. Restrictions apply. . I I-- The proposed control board was embedded in a prototype electric scwter in order to increase its reliability and autonomy. 201 mn Figure 9 Battery pack lay-out CONCLUSIONS This paper presents a simple approach for protection and estimation of State of Charge goC) of batteries tailored for electric vehicles applications. The proposed estimation is based only on the measurement of electrical quantities like /em, g, Ri usually available without introducing further invasive sensors. The data input set for training a NN is built relying on experimental measurements of the aforementioned electrical quantities. The NN is a two layer one with wellstated activation function and leaming algorithm. The real time measurements of battery charge has been performed in order to validate NN based estimation results. It can be noted that this approach does not require any physical or chemical knowledge of battery behavior. Experiments show that the method is reliable and accurate. Electronic circuits for charge equalization, undervoltage and overvoltage monitoring, temperature sensing were designed and implemented o n a dedicated PCB. Their outputs drive motor control strategy according with joumey characteristics. The implemented charge equalization strategy allows to obtain the maximum energy storage also during vehicle braking operation. EFERENCES Rajasheh K.: 'History ofElectric Vehicles in General Moms' IEEE Trans. On IAS ~01.30 n.4 IulyiAugmt 1994 pp.897-9M Miller, J.M., Gale, A.R.: 'Hybnd electric vehicle luefess will depend on low COSI, efficient pawer electronicssystem', Power Conversion and Intelligent Motion, November 1997. M. Reizenman, "The EV future", L E E Spun" novmkr 1998. P. B<uuerlein, R. Herr, M. Kloss, 1. Khpers, M. Maul, E. Meisner, "Wvanced Lithim ion cells with Lithium Manspiner,Jo-iopo-sourcer,~o. 81-82 1999,pp. 585.~88. P. casaSs0,A. Fratta, G . Firando, P. Guglieltni, M. Pastoreili, A. Vagati. -High-performance elecbie raote?', to appear on Pmeeedingrof PClM 2003, N u t " & Germany, May 2003 0.Caumont, P. Le Moigne, C. Rombaut, X. Muneret. P. Lenain, 'Energy Gauge for Lead Acid Batteries in Electric Vehicles", I€€€ lr-~lctiom on Enerw Conwrsion. VOL. 15. NO. 3. Senembre2000. T. Yana&~% A. Kawmum,''Raidual Capacity Estimation of Sealed Lead-AcidBatteries for Electtic Vehicles", F'CCNogaob '97,pp. 943-946, 1. Yammki, K. Sakurai, K. Muramoto, "Estimation of the Residual Capacity of Sealed Lead-Ad Baneties by N d - Nerwork".PESC'98pp.210-214. J. Peng, Y.Chen, R. Eberhat, "Battery Pack Smlate of Charge Edmator b i p UsingComputational Intelligence Approaches", Bonery Confwence on Applicnrionr ondA&nees, 2QM),pp.1 7 3 177. 1. Fukuda, "Theory and Applications of Neural Nenvorks fr Indusm.ai Canml Sysrrms': E€€rmrmchm on zdzwid Elenronier, VOL. IS, NO. 4, Derrmber 1992. The Mathwoh, "MATLAB Neural Networks Toolbox". A.Vagati, G.Francerehini, M.Panorelli, V.Omgoreanu: "Fluxobserver-based high performance contml of S~dnonOuF Reluctance Malm by including cmss-sahnatlon" LEEE lramocrions on S d u f r y Applicarionr, Vo1.35,N0.3 MaylJme 1999. 688 Authorized licensed use limited to: Universiti Tunku Abdul Rahman. Downloaded on March 11,2021 at 02:42:09 UTC from IEEE Xplore. Restrictions apply.