EVS28 KINTEX, Korea, May 3-6, 2015 SOC estimation of LiFePO 4 Li-ion battery using BP Neural Network Liun Qian, Yuan Si, Lihong Qiu. School of Mechanical and Automotive Engineering, Hefei University of Technology, Hefei 230009, Anhui, China Abstract As a new generation of green high-energy rechargeable battery, Li-ion battery has multiple advantages such as high voltage, high energy density, brilliant cycle performance, low self-discharge and no memory effects. Therefore, the application of Li-ion battery in electric cars is getting much more widespread. SOC is an important parameter for battery management system (BMS). Accurately estimating SOC of battery could reduce permanent damage to the battery caused by over-charging and over-discharging which will in turn extend battery life, as well as improve vehicle performances, reduce requirements of the power battery and improve economy. Current SOC estimation methods mainly include electrochemical impedance spectroscopy (EIS), electro motive force (EMF) method, artificial neural network (ANN) model, fuzzy Logic, Kalman Filter, linear parameter varying (LPV) and sliding mode observer method. Compared with other methods, ANN model is highly accurate, adaptive, logically complex, and requires a lot of training data. For this paper, several experiments of rapid charging and discharging of Li-ion battery pack have been conducted under NEDC, obtaining a large number of experimental data. Parameter characteristics of Li-ion battery have been analyzed and performance parameter shave been found out, which affect the data of SOC. Feed-forward network model is built using BP algorithm. Then experimental data are imported, neural network is trained and network structure is constantly adusted for accuracy optimization. After that, Simulation Module based on BP algorithms is founded and its simulation figure is analyzed. It is shown that the average error of the simulation model after training is less than 0.45%. Simulation model based on this method can accurately measure real-time SOC value of battery. SOC variation in other conditions can be obtained by the same way, which is particularly practical in the battery management system. Keywords: LiFePO 4 Li-ion batteries; SOC estimation; BP neural network EVS28 International Electric Vehicle Symposium and Exhibition 1
1 Introduction As a new generation of green high-energy rechargeable battery, Li-ion battery has multiple advantages such as high voltage, high energy density, brilliant cycle performance, low self-discharge and no memory effects. Therefore, the application of Li-ion battery in electric cars is getting much more widespread. SOC is an important parameter for battery management system (BMS). Accurately estimating SOC of battery could reduce permanent damage to the battery caused by over-charging and over-discharging which will in turn extend battery life, as well as improve vehicle performances, reduce requirements of the power battery and improve economy. Current SOC estimation methods mainly include electrochemical impedance spectroscopy (EIS), electro motive force (EMF) method, artificial neural network (ANN) model, fuzzy Logic, Kalman Filter, linear parameter varying (LPV) and sliding mode observer method [1]. Neural network model, which is more accurate and sophisticated compared with other methods, requires a large amount of data for training. The accuracy of the model depends on the accuracy of the training data. Based on the large sum of data obtained from experiment, network model is chosen to predict the current SOC. 2 LiFePO 4 Li-ion battery test 2.1 Basic parameters of the battery Table 1.1 basic parameters of the battery framework and device connection is shown in figure1. Battery Test System Remote Console CAN bus Bettery Pack and BMS system control network System Electricity Cabinet DC bus Three-phase Distribution Network Figure 1 test framework and device connection 2.3 The NEDC cycle test The test is according to the auto industry standard that is <lithium ion battery of the electric car>, QC/T43-2006. Test procedure: (1) Determine the charging current intensify according to the current temperature of the battery measured by the sensor under the ambient temperature. When the cell voltage reaches 3.45V, charge with small current and reduce the current with certain step length till the end. (2) Set aside for half an hour and discharge it in working condition until the DOD reaches 10%. (3) Set aside for half and an hour after the discharge process complete and repeat step (1) and (2). The current of the NEDC working condition changes according to time, as shown in figure 2 Type capacity/ah energy/kwh Cluster placer LiFePO 4 60 19 5P95S 2.2 Battery test plant The battery test plant consists of CAN bus, electric box, controlling terminal and battery pack. Test EVS28 International Electric Vehicle Symposium and Exhibition 2
Figure 2 Current of NEDC condition A test cycle lasted 5.5 hours on average, including a charging process and a NEDC discharge process, gathering 75000 sample points in a discharging process. Data gathering includes the input and output capacity, energy and power of the battery, the input and output capacity, energy and power of the electric cabinet, each cell voltage, maximum and minimum cell voltage and the corresponding number, total battery voltage, current, temperature, and time. x and d. Calculate the components of Y and O with formula (1) and (2). T y = f ( V X), = 1,2,..., m (1) T o= fwy ( ), k= 1, 2,..., l (2) k 3. Calculate the error of network output: Assume that there are p pairs of training sample. Error E p differs corresponding to different sample. Either E max, the maximum value of the error or (3), root mean square of the error can be utilized. 3 Artificial neural network model The network this paper used is feed forward network based on BP algorithm. 1 P p E p P 1 Erms = (3) 4. Calculate error signal of each layer: calculate v δ k 3.1 Principle of BP algorithm BP algorithm contains forward transmission and reverse transmission. In forward transmission, sample is processed from the input layer and the hidden layers to the output layer, during which the output of each neutron layer only affect the states of the next layer. The algorithm enters the reverse transmission if the network output deviates from the expected output O d. In reverse transmission, error signal is opposite to the direction of the forward transmission and amend the weight coefficient of each neutron layer along the negative gradient direction of the error function so that the expected error function tends to the minimum. Therefore, BP algorithm is a search algorithm on the basis of gradient method, which fully demonstrated the features of parallel process of the neural network [4]. Program realization of BP algorithm: 1. Initialization: assign random value to weight matrix W, V. Set sample mode counter p and training frequency counter q to 1. Set error to 0 and acquisition efficiency to a decimal between 0 and 1 and accuracy of the network training E min to a positive decimal. 2. Input training sample and calculate the output of each layer. Assign current sample x p and d p to vector and δ y with formula (4). ( )( ) o δ = d o 1 o o, k = 1, 2,..., l (4) k k k k k ( ) δ y = o ( ) 1, 1, 2,..., δ w y y = m (5) k k 5. Adust the weight of each layer: calculate the components of W, V with formula (6) and (7). w w + ηδ y (6) o k k k v v + ηδ x (7) y i i i 6. Check for the train completion of all the samples in rotation: if p < P,Counter p, q add 1,return to step (2), otherwise, turn to step (7). In the current applications, there are two methods for Weight adustment. As can be seen from the above steps, each input sample should return errors and adust the weights in the standard BP algorithm, this weight adustment method of rotating for each sample also known as Single-sample training. Because single-sample training is short-sighted to follow selfish departmentalism principle, only adusting the errors for each sample he makes, so that increasing the numbers of the training, and leading to the convergence speed too slow. Another method is EVS28 International Electric Vehicle Symposium and Exhibition 3
calculating the errors of the network after all sample inputting. 1 E = d o 2 P l p p 2 ( ) (8) k k p= 1 k= 1 Then calculating layers errors based on the total errors, and adust the weight, this batch mode of cumulative errors can be called Batch training or Epoch training. Due to the Batch training followed the Collectivism principle of the goal of reducing the global errors, which guaranteed the total errors changing to the decreasing direction. When large number of samples, batch training is faster than Single-sample training over the convergence speed [4]. 3.2 Input selection of neural network Neural network's inputs are selected from the factors of related the current SOC. From the perspective of electrochemistry, the integral of charge and discharge current with respect to time is one of the most basic calculating methods of SOC because electronic migration leads to charge. Besides, SOC is influenced by varying degrees polarization phenomena affected by charge or discharge current. There are close ties between voltage and SOC due to the open circuit voltage method which is an interpolating estimation using the relationship between voltage and remaining capacity [8]. Temperature influences SOC through affecting the change of cell resistance. The test subect is the battery pack instead of individual cells, so cell consistency, which affects battery performance, should be considered. Depth of discharge (DOD) and amount of looping also change real battery capacity to a certain extent, which will change the calculated value of SOC [2]. To simplify the neural network model, don't choose DOD and amount of looping which were constant in a certain loop as input. The influence of cell consistency effects are shown on maximum and minimum cell voltage and temperature. From the above, the neural network has eight inputs including total battery voltage, current, the highest cell temperature, the lowest cell temperature, cell voltage maximum, minimum, and cell voltage time and discharge power. We classified the data of tests, checking the relationship between various factors and battery SOC, respectively and the result is shown in figure 3-6. Figure 3 the relationship between total battery voltage and SOC Figure 4 the relationship between time, current and SOC EVS28 International Electric Vehicle Symposium and Exhibition 4
Figure 5 the relationship between temperature and SOC model with 29 nodes in hidden layer has a good network convergence performance and can be used. Figure 6 the relationship between power and SOC 4 Neural network training and result After creating artificial neural network, the number of nodes in hidden layer is uncertain, and needs a lot of adustment. We compare the training effect of 8-50 nodes and select 29 nodes with less error and better simulation effect [7]. The training results as follows. Open the performance figure, the change of the neural network training error can be observed in real time. Error reaches 0.0094 after 582 steps training, which is less than the limit error value. Performance is shown in figure 3. Figure 8 Error Histogram Figure 8 shows the training error, validation error and testing error distribution focused around 0.004245, the error of absolute value is within 0.15, and the training results can achieve better prediction accuracy. Figure 9 is regression performance,in which the training data, validation data and testing data expected response and simulation output show strong linear correlation and regression constant is over 0.99999. Figure 7 Performance It can be seen from above that network converges quickly after 10 steps training, which proves the Figure 9 Regression EVS28 International Electric Vehicle Symposium and Exhibition 5
5 Simulation analysis Artificial neural network prediction model is one of the black box model, the main part of the module is a black box mainly composed of neural network. Input the 8 input values of battery actual works and then get current SOC under the neural network prediction. Comparing with simulation value and measured value and the result is shown in figure 10.Caculating relative error is no more than 0.018, average relative error is0.0020. (a) 6 Conclusion and reflection 1. After being imported with a large amount of data for training and test, multi-layer feed forward network model, based on the BP algorithm, begin to mature with average error less than 0.45% and regression coefficient more than 0.999. Then we built the simulation module of the model with simulink. The data obtained from this simulation module coincide to the experimental data, which further demonstrated the accuracy of the model. The simulation module obtained by this method is capable of accurate measurement of SOC value in NEDC condition. 2. The same method can be used to detect the SOC changes when the vehicle is operating in other working conditions through training. The simulation module greatly improves the real time estimation of the SOC which plays a very practical role in the battery management system. 3. The consistency of the cells in a battery has a huge impact on the lifespan of a battery, it still requires our effort to figure out a reasonable description method for cell consistency and to what extent the cell consistency affects the battery s performance. References [1] Seyed Ehsan Samadani, Roydon A. Fraser,Michael Fowler University of Waterloo. A Review Study of Methods for Lithium-ion Battery Health Monitoring and Remaining Life Estimation in Hybrid Electric Vehicles.SAE international,2012. (b) Figure 10 comparisons with simulation value and measured value (a) completed picture (b) local picture [2] Wang Zhengpo, Sun Yingchun. Electric vehicle power battery systems and application technology [M]. Beiing: Mechanical industry press, 2012 [3] Yin Andong, Zhang Wanxing, Zhao Han, Jiang Hao.SOC prediction research of LiFePO4 Li-ion batteries based on neural network[j].ournal of electronic measurement and instrument,2011,25(5)433-437 EVS28 International Electric Vehicle Symposium and Exhibition 6
[4] Shi Yan, Han Liqun, Lian Xiaoqin. Neural network design and case analysis [M]. Beiing : Beiing university of posts and telecommunications publishing house,2009 [5] W. X. Shen, Student Member, IEEE, C. C. Chan, Fellow, IEEE,E.W.C.Lo, Member, IEEE, and K. T. Chau, Member, IEEE. Adaptive Neuro-Fuzzy Modeling of Battery Residual Capacity for Electric Vehicles. IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, 2002, 49(3):677-684 [6] Ye Weiqiong, Li Bei. Dynamic measurement and estimation of power battery residual capacity. Advanced Materials Research, 724-725 (2013) pp 829-83 [7] Zhu Kai, Wang Zhenglin. Mastering MATLAB neural network [M]. Beiing: Publishing House of Electronics Industry, 2010. [8] Shi Wei, Jiang Jiuchun, Li Suoyu, Jia Rongda. Research on SOC estimation for LiFePO4 Li-ion battery [J]. Journal of electronic measurement and instrument, 2010, 24(8):769-774 [9] Charkhgard, M. and Farrokhi, M., State-of-Charge Estimation for Lithium-Ion Batteries Using Neural Networks and EKF, Industrial Electronics, IEEE Transactions on, vol.57, pp. 4178-4187, 2010 Authors Liun Qian received his Ph.D. degree in School of Mechanical and Automotive Engineering, Hefei University of Technology, Hefei, China, in 2004. Since 2004, he has been a professor in Hefei University of Technology. His research interests include vehicle modern design theory and method, vehicle safety technology, and electrical vehicle technology. Lihong Qiu received his B.E. degree from Hefei University of Technology, Hefei, China, in 2013. He is now pursuing his Ph.D. degree in School of Mechanical and Automotive Engineering, Hefei University of Technology. His research interests include the energy management strategy for plug-in 4WD hybrid electric vehicle and the dynamic control for the HEV. Yuan Si received her B.E. degree from Hefei University of Technology, Hefei, China, in 2014. She is now pursuing her master degree in School of Mechanical and Automotive Engineering, Hefei University of Technology. Her research interests include battery management, control of electric vehicle. EVS28 International Electric Vehicle Symposium and Exhibition 7