Kai Guan, Zhiqiang Wei, and Bo Yin
Abstract The prediction of a battery’s state of charge (SOC) is one of the key tasks of battery management. Lithium battery internal chemical reactions are complex and have many factors; its SOC prediction has strong nonlinear characteristics. This paper discussed a SOC prediction model which is based on hybrid genetic algo- rithm and BP neural network. Set BP neural network’s training error as genetic algorithm fitness value, and then iterate to find the optimal individual as the neural network initialization thresholds and weights. Simulation results show that this method can accurately predict the new kind of a lithium battery’s SOC and have higher accuracy compared with BP neural network.
Keywords State of charge • Genetic algorithm • BP neural networks • Prediction method
17.1
Introduction
A battery’s state of charge is an important parameter characterizing the state of the battery. It cannot be directly measured during the charging and discharging of the battery. SOC prediction has a strong nonlinear characteristic and is difficult to describe accurately and mathematically [1]. Artificial neural networks can imitate multiple input and output functions and have some good characteristics, such as high nonlinearity, fault tolerance, and robustness. It can give the corresponding output for external excitation and is very suitable for a battery’s SOC prediction [2]. BP neural network’s optimization is based on the steepest descent method which has slow convergence speed and is easy to fall into local minimum point [3]. This paper is focused on hybrid genetic algorithm and BP neural network to predict lithium battery SOC. Using genetic algorithm to optimize the initial param- eters of BP neural network can enhance the generalization ability of BP neural network [4]. Some simulations and experiments have been done.
K. Guan (*) • Z. Wei • B. Yin
College of Information Science and Engineering, Ocean University of China, 266100 Qingdao, Shandong, China
e-mail:[email protected]
©Springer International Publishing Switzerland 2015
W.E. Wong (ed.),Proceedings of the 4th International Conference on Computer Engineering and Networks, DOI 10.1007/978-3-319-11104-9_17
17.2
Related Work
The widely recognized definition of SOC is defined by the US Advanced Battery Association (USABC); SOC is the acronym for state of charge; it refers to a certain discharge rate, the ratio of the remaining charge and the rated capacity [5]. SOC is considered 100 % when battery energy reaches saturation at a certain temperature and 0 % when battery energy can no longer be released [6]. SOC is calculated as
SOC¼ Remaining charge
Maximum discharge capacity¼
Q Ið Þ m Q Ið Þn Q Ið Þm ð17:1Þ Q Ið Þ ¼n t ð Indt ð17:2Þ
whereQmrefers to the maximum discharge capacity of battery when discharged with a constant current I and Q(In) refers to the released battery power when discharged with a constant currentI.
The discharge test method is the most reliable battery state-of-charge estimation method. It takes a long time and is not on line [7].
AH Metrology: The current flowing into the battery is integrated by time to calculate the inflow and outflow of the total battery charge [8]; it is calculated as
SOC¼SOC0 1 CE ðt 0 idt ð17:3Þ where SOC0is the initial SOC value andCEis the battery rated capacity. Open-Circuit Voltage Method: When the battery is unused for a long time, the voltage and the battery SOC form a relatively stable linear function. The Kalman filtering method uses linear minimum mean square error criterion, treats the internal state of the battery SOC as a management system variable, and uses estimates of the last moment plus the value of real-time measurements to recursively approach real- time estimation [9].
Neural Network Method: Neural networks have nonlinear characteristics and have a good ability to learn. Using the neural network for SOC prediction, the model is generally constituted by an input layer, an intermediate layer, and an output layer. The accuracy of the training data samples, sample size, and training methods will have a significant impact on the network’s accuracy [10].
17.3
SOC Prediction Method Based on GA-BP
Neural Network
The neural network is a function approximation method; the most widely used neural network is a BP (error back propagation) neural network; it is a three-layer feed forward network; the output of the previous layer is the input of next layer; each input node has a weight value. A three-layer network model is shown in Fig.17.1. According to the Kolmogorov theorem, three-layer neural networks can approximate a continuous function with arbitrary precision [11]. A three-layer BP neural network is used to predict the battery SOC. The model is shown in Fig.17.2. Genetic algorithm is a randomized iterative probabilistic search algorithm. This algorithm is a kind of imitation of the law of biological evolution. Genetic algo- rithms do not include the form of the problem; it changes the gene configuration to achieve overall optimization of the problem, and it has inherent parallelism and better global optimization capability.
Use genetic algorithm to optimize the BP neural network. The length of an individual is determined by the structure of the BP neural network. Each individual in the population consists of a set of network weights and thresholds. Genetic
LA LB LC Wir Wrj input output m n Fig. 17.1 Three-layer BP neural network U I Y Current SOC Voltage Fig. 17.2 BP neural network SOC prediction model
algorithm uses selection, crossover, and mutation operations to find the best fitness value and the best individual. The BP neural network takes the best individual as the initialization weights and thresholds, and then the network was trained to predict the function output. The process step pseudo-code is shown in Fig.17.3.
The network structure model has 15 weights and 6 thresholds, so the genetic algorithm individual code length is 21. Each individual is a real number string.
Individual fitness function is like
F¼k X n i¼1 absðyioiÞ ! ð17:4Þ wherenis the network output nodes,yiis the expected output value of theinode, oiis the predicted output value of theinode, andkis the coefficient.
The selection probability of each individualPican be calculated as (17.5) pi¼ k Fi XN j¼1 fj ð17:5Þ
Fiis the fitness value of individualiandkis the coefficient.Nis the number of individuals of the population.
define BP neural network structure load input data
for i 1 to sizepop //sizepop is the population size
individuals.Code(i)=Code(lenchrom) //code randomly
individuals.fitness(i)=BP neural network training error //calculate fitness value end
bestchrom = min(individuals.fitness) //find the best chromosome
for i 1 to mangen //maxgen is the maximum evolution generation
Select() //do select operation
Cross() //do cross operation
Mutation() //do mutation operation
for j 1 to sizepop
individuals.fitness(j)=BP neural network training error //calculate fitness value end
bestchrom = min(individuals.fitness) //update best chrom
trace the bestfitness value and best individuals //trace best individual in gen j end
Initialize network weight and threshold = best individual x net = new GA-BP neural network
set train.epochs and train.goal
net = train(net,inputn,outputn) //GA-BP neural network training return simulation results
Fig. 17.3 Genetic algorithm to optimize BP neural network
Cross operation is calculated as
akj¼akjð1bÞ þaljb ð17:6Þ
alj¼aljð1bÞ þakjb ð17:7Þ
Mutation operation is calculated as
aij¼ aijþ aijamax r2ð1g=GmaxÞ2 r>0:5 aijþ aminaij r2ð1g=GmaxÞ2 r0:5 ( ð17:8Þ
amax is the upper bound of gene aij,aminis the lower bound of gene, r2is a random number,g is the current number of iterations, andGmaxis the maximum number of evolution.
17.4
Application Instance
Lithium-ion battery data was collected from a new lithium-ion battery of soft carbon negative materials. The carbon anode material is micron and spherical. And it is made up of multiparticles with secondary granulation technology. The charging and discharging properties of materials were improved by shortening the lithium migration path in anode material particles. The charging and discharging curve is shown in Fig.17.4.
Fig. 17.4 Battery charging and discharging curve
With the sample data, we carried out a median filter to remove noises and used the difference method to meet the shortage of samples. The results of data processing are shown in Fig.17.5.