蓄電池外文翻譯

1、<p><b>　　翻譯部分：</b></p><p><b>　　英文原文</b></p><p>　　Research upon the High-capacity Lead-Acid Battery Charge Model</p><p>　　Kun Yang and Guangyao OuYang<

2、;/p><p>　　Department of Power Engineering</p><p>　　Naval University of Engineering</p><p>　　Wuhan, Hubei Province, China</p><p>　　yangkundexiangzi@sina.com</p><

3、p>　　Ping Zhang and Jianping zhang</p><p>　　Department of Power Engineering</p><p>　　Naval University of Engineering</p><p>　　Wuhan, Hubei Province, China</p><p>　　ya

4、ngkundexiangzi@sina.com</p><p>　　Abstract - The simulation of the submarine battery charge process is an important part of the submarine battery modeling, which is especially a firm foundation for the furthe

5、r optimization usage of the modules of the whole submarine power system. In this paper, the modeling strategy of the submarine battery charge process is discussed in detail firstly, in succession an integrated algorithm

6、in which both the improved back propagation neural network method and the linear simulation technology are</p><p>　　Index Terms-：Neural network; Linear simulation; Battery; Charge model.</p><p>

7、　　INTRODUCTION</p><p>　　Battery is the only power source for the conventional submarine underwater propulsion, so the submarine has to move up to the snorkel state to charge for the battery when the energy i

8、s exhausted, however the adoption of the snorkel sailing state throw the submarine into a fatal danger, because the snorkel state make the submarine easier spied by its enemy for its diversified increscent physical field

9、s, such as the infrared radiation" navigational wake etc., thus the research upon the submarine b</p><p>　　On account of that the voltage of the battery can be affected by many factors, such as the SOC

10、(SOC: State Of Charge)" battery polarization delamination" former discharge rate etc., moreover the factors above are mutual influential; it is hard to decouple for them. The conventional electrochemistry mecha

11、nism model electrochemistry experiential formula model and the equivalent circuit model considers that there must be some formula relations among the main factors of the battery [1][2][3], actual</p><p>　　In

12、 view of the electrochemistry mechanism complexity of the battery, an integrated algorithm in which both the improved BP(BP: Back-Propagation) neural network method and the linear simulation technology are used is introd

13、uced in the paper, basing on the deep analysis on the battery charge test curve data. Finally it is proved that the algorithm has a very famous practicality in the battery charge process simulation.</p><p>　

14、　THE BUILDING OF THE BATTERY MODEL</p><p>　　The Analysis on the Battery Charge Process Characteristic</p><p>　　Considering the submarine battery's life limit and the high expenditure, it is

15、unlikely to do large numbers of tests to obtain platitudinous data for the simulation, so we have to think of ways to simulate for it only through the limited several battery charge test curves. Fig. 1 shows us several t

16、ypical practical charge process test curves under different former discharge rate in a certain charge period (for convenience, some of the data is normalized.) From Fig. 1, it is easy to know that the </p><p&g

17、t;　　The charge times of the middle stages are comparatively short, the data points are comparatively centralized and the curves ‘trend changes acuter. The neural network algorithm is once used to try to simulate the whol

18、e five stages charge process directly, however the net is so hard to be convergent, even along with the introduction of the improved algorithm making the net performance threshold met, the output curve cannot reflect the

19、 charge process correctly all the same, even worse the data is i</p><p>　　Fig. 1 Battery charge process test curve</p><p>　　Modeling for the Battery Five Stages Charge Processes </p><

20、p>　　Considering the highly nonlinear characteristic of the five stages charge process test curve that is showed in Fig. I, it is conceived to simulate each stage of the five separately.</p><p>　　I) The Si

21、mulation of the First Stage Charge Process It is well-known that both the nonlinear system and the uncertainty system can be described commendably by the ANN (ANN : Artificial Neural Network) algorithm because of its exc

22、ellent performance in parallel processing and self-learning, therefore the ANN algorithm provides us a feasible resolvent for the modeling of the first stage of the battery dynamic charge process[4]. Besides it has been

23、proved that the three-layer feed forward BP neural </p><p>　　In this paper, a BP neural network model is to be constructed to simulate the first stage of the battery charge process under either former discha

24、rge rate, firstly the Neural network principium sketch map is showed as Fig. 2 below:</p><p>　　Fig. 2 The neural network principium sketch map</p><p>　　When the Sigmoid function is chosed as the

25、 neuron transfer function, namely ，considering the weights updating formula:</p><p>　　Updating Increment= (Learning Rate) x (Teacher signal-Neuron Output) x (Sigmoid Function Differential Value) x (Neuron Ou

26、tput) (1)</p><p>　　Actually from the differential of the Sigmoid function as follows:</p><p><b>　　(2)</b></p><p>　　It is obvious that th

27、e output of the neuron always ranges from 0 to 1. when the approach 0 or I, the Updating increment becomes smaller, thus the net stability boosts up, however virtually the huge frequency happening of this situation ofte

28、n leads to an even slower learning speed. Aiming at suchlike problems, large numbers of improved algorithms are put forward by scholars recent years, and the improved algorithm by adding the item of the momentum is adopt

29、ed in this paper, which is widely used </p><p><b>　　(3)</b></p><p>　　There into is the error square sum of the output layer until the time of day n-I, 1] is the learning rate, and t

30、he is the momentum constant. If the current correction direction (the first item on the right of (3)) differ with the former correction direction (the second item on the right of (3)), namely the sign is contrary, it is

31、 considered that there is a certain extent instability, here the absolute value of the correction sum becomes smaller preventing the excess adjustment; by contraries i</p><p>　　Considering the relations betw

32、een the learning rate and the momentum constant , on the assumption that:</p><p><b>　　(4)</b></p><p>　　In company with the (3) had or would:</p><p><b>　　(5)</b

33、></p><p>　　The weights at the time of day t equals N:</p><p><b>　　(6)</b></p><p><b>　　Farther:</b></p><p>　　When , and on the assumption that

34、 the evaluation function is slick, it comes to a conclusion that the learning speed of progress is almostly pro rata to </p><p>　　2) The Simulation of the Later Stages Charge Process</p><p>　

35、　Referring to the charge test data that is showed in Fig. 1,only the start point and the terminal point data are recorded as the charge track record of the later stages, therefore it is very difficult to confirm the chan

36、ge rule of the voltage. Although the qualitative analysis indicates that the change rule of the second" the third and the fourth stage are the same as the first stage, in view of the charge times of the three stages

37、 are comparatively short, the linear simulation is introduced in t</p><p>　　THE SIMULATION REALIZATION OF THE FIVE STAGES CHARGE PROCESS</p><p>　　Before the implementation of the modeling, it is

38、 supposed that the performance state of the batteries is excellent and the working condition is perfect, this is described in detail as follows:</p><p>　　The working state of the electrolyte beater is normal

39、;</p><p>　　The charge current of each stage keeps constant;</p><p>　　Each battery of the pile are at the same state.</p><p>　　A. The building of the neural network model of the firs

40、t stage charge process</p><p>　　Because that the number of data points of the first stage charge process under different former discharge rate are dissimilar in Fig. 1, meanwhile the net also needs a mass of

41、 eximious data to be the input tutor signal vector, So the charge test data is extended through spline interpolation firstly, on the other hand the extension of the test data guarantees the consistency of the dimension o

42、f the input vector.</p><p>　　The Setup of the Net Parameters</p><p>　　In the beginning the net is hard to be convergent when the net initial weights are assigned via the program itself.</p>

43、;<p>　　Considering that the system is nonlinear, so the convergence may be mainly determined by the selection of the initial weights. After many failures of the net training are suffered, in order to make the net

44、to be convergent, the net initial weights are designated to the random numbers that ranges from 0 to 1, Which makes the weights of the neurons adjusted at the maximal position of the transfer function. The result of the

45、simulation experimentation indicates that the net comes to be convergent q</p><p>　　The program is a learning algorithm realization of a net that contains input layer, hidden layer and output layer. The tran

46、sfer function of the hidden layer is designated to be the tansig function, and the transfer function of the output layer is the purling function. The input vector of the net is X = (XI, X2, X3) = (Charge Periods, Former

47、Discharge Rate, and Charge Time), the output of the net is the voltage of the different moment under different former discharge rate in a certain charge peri</p><p>　　Net Training</p><p>　　There

48、 is definite difference among magnitude order of the components of the input vector, which leads to the net precision cannot be commendably satisfied, so the input and the output data is normalized in the paper firstly [

49、5].Besides when the neural network algorithm is brought forward for modeling, the selection of the number of the neurons of the hidden layer is a key factor. When the net is too small, the learning ability becomes weak a

50、nd the net is hard to be convergent. By contraries if t</p><p>　　Through large numbers of simulation experimentations [6], the number of the neurons is finally hosed to be 9. The contradistinctive image of t

51、he net simulation result and the test data is showed as Fig. 3(The asterisk in the figure denotes the test data), and the maximal data error is 0.4 percent which indicates that the effect of the simulation is extremely h

52、eart-stirring. Besides the net training error curve is showed in Fig. 4. Finally the trained net is saved as a mat data document in order </p><p>　　Fig. 3 The contradistinctive image of the net training data

53、 and the test data</p><p>　　Fig, 4 Net training error curve</p><p>　　Fig.5 shows the first stage charge process curve data under different former discharge rates that is obtained through the tra

54、ined net.</p><p>　　Fig.5 The first stage charge curve outputted through the trained net</p><p>　　The Building of the Models of the Later Stages</p><p>　　It is easy to come to a con

55、clusion that the charge times of the later stages are almost the same under the same former discharge rate through the charge test curve, however the charge times under different former discharge rates are inconsistent.

56、The charge times of the later stages under the 1 hour, 5 hour and the 50 hour are offered in the table I (The data is normalized.) The charge times under the rest former discharge rates can be obtained through spline int

57、erpolation.</p><p>　　Besides, the charge start voltages of the later stages are dissimilar under the same former discharge rates which are showed in table II. In the same way that the charge start voltages o

58、f the later stages under the rest former discharge rate can be obtained through interpolation.</p><p><b>　　TABLE I</b></p><p>　　CHARGE TIMES OF THE LATER STAGES UNDER DIFFERENTFORME

59、R DISCHARGE RATE</p><p><b>　　TABLE II</b></p><p>　　THE CHARGE START VOLTAGES OF THE LATER STAGES UNDERDIFFERENT FORMER DISCHARGE RATES</p><p>　　The Modeling of the Whole

60、 Five Stages Charge Process of the Battery</p><p>　　When the first stage charge process model is integrated with the later stages charge models , the whole five stages charge process of the battery is obtain

61、ed, the contradistinctive image of the simulation result and the test data is showed in Fig. 6(The asterisk in the figure denotes the test data) , and the simulation output of the charge process under 3 hour discharge ra

62、te is also exhibited in Fig. 7.</p><p>　　Fig. 6 The contradistinctive image of the simulation result and the test data</p><p>　　Fig. 7 Simulation output of the charge model under 3 hour discharg

63、e rate</p><p>　　IV. CONCLUSIONS</p><p>　　Basing on the particular research upon the characteristic of the submarine battery charge process test curve data, a new algorithm that is the integratio

64、n of the improved BP neural network algorithm by adding the item of the momentum and the linear simulation technology is brought forward in the paper, and the simulation experimentation indicates that comparatively high

65、precision is achieved, which shows that the algorithm has a very good practicability in the battery charge process modeling.</p><p>　　REF ERENCES</p><p>　　[1J Stefano Barsali, Massimo Ceraolo. D

66、ynamical models of lead-acid batteries. IEEE Transactions on energy conversion ' 2002 ' 17(1) :16-23.</p><p>　　[2J Li Bei, Wang Yanglin, Chen Zhibin. Reducing of mathematical-model and study in charg

67、ing about minor Lead-Acid battery [1]. Telecom Power Technology, 2008,25(4):27-28,34.</p><p>　　[3J XU Guoshun, Zhuang Jinwu, Yang Feng, Mao Haitao. Research on mathematical- model and parameter identificatio

68、n of high-capacity storage battery [J].Journal of Naval University of Engineering, 2007,19(3):35-38.</p><p>　　[4J Hecht-Nielsen R. Theory of the hack propagation neural network. IEEE International Joint Conf

69、erence on Neural Networks. Washington DC:</p><p>　　IEEE Computational Intelligence Society' 1989: 593-605.</p><p>　　[5J Ma Guangzhi, Hu Shaofeng. Improvement of the stability and convergence

70、 in BP algorithm [J].Huazhong Univ. of Sci. & Tech.</p><p>　　(Nature Science Edition), 2002, 30(12):21-22.</p><p>　　[6J Tang Yong, Chang Li, Zhou Jianzhong. The application of improved BP ne

71、ural network to temperature modeling and forecasting in shaft bushing of hydroelectric unit],l]. Huazhong Univ. of Sci. & Tech. (Nature Science Edition), 2002,30(4):78-80.</p><p>　　程。這可能是因為該對象的神經網絡運算法則的性

72、能誤差是全球性的。所以，甚至是全球性的誤差也會得到滿足。它可能無法顯示每個本地曲線的部分已經正確表達?？傊紤]到電池充電曲線的特征，即使是神經網絡運算法則對于電池的動態(tài)過程建模給我們提供了一個新的方法，另一些技術應該被補充到建模課程中。</p><p>　　圖1蓄電池電池充電過程測試曲線</p><p>　　電池充電過程的五個階段的建模</p><p>　　考慮到

73、在圖一顯示的充電過程的五個階段的測試曲線的高度非線性特征，它被構思為分別模擬這五個階段的每一個階段。</p><p>　　第一階段充電過程的模擬</p><p>　　眾所周知，因為人工神經網絡運算法則在并行處理和自學習方面有優(yōu)良的性能，所以非線性系統(tǒng)和不可信度制度能被其贊揚地描述。因此，人工神經網絡運算法則在第一階段的電池動態(tài)充電過程建模為我們提供了一個可行的解決方法[4]。此外，三層前饋

74、BP神經網絡已被證明了可以被訓練逼近任何多輸入多輸出的真正作用（用有限的不連續(xù)數目）。顯然，第一階段充電過程中可以很容易地通過人工神經網絡技術進行模擬。</p><p>　　在本文中，BP神經網絡模型是根據模擬非過去的放電率第一的神經網絡來搭建的。，電池的充電過程的第一階段原理示意圖如圖2所示：</p><p>　　圖2神經網絡原理示意圖</p><p>　　當選擇

75、Sigmoid函數作為神經元的傳遞函數，即，考慮權重更新公式：</p><p>　　增量更新=（學習速率）×（教師信號神經元的輸出）×（sigmoid函數微分價值）×（神經元的輸出） （1）</p><p>　　其實從Sigmoid函數微分如下：

76、</p><p><b>　　(2)</b></p><p>　　很明顯神經元的輸出總是范圍從0到1。當的接近0或1時，更新中的增量變小，從而增強穩(wěn)定性凈起來，但實際上巨大的頻率發(fā)生這種情況往往導致學習速度更慢。在諸如近年來許多學者針對大量此類問題提出了改進算法，本文加入了目前廣為使用的通過動量項的改進算法，即在增量更新的重量將在一天的時間計算等于n噸，增量更新的權重

77、相對應的時間是一天中n我也考慮。在具體的計算公式如公式（3）：</p><p><b>　　(3)</b></p><p>　　這里直到n - 1天時間的輸出層平方是錯誤的，η為學習率，α為常數的勢頭。如果目前的調整方向與前校正方向不同（在第（3）右側第一項）（上右第二個項目（3）），即符號相反，它被認為是有一定的程度的不穩(wěn)定，在這里，殘差絕對值變小防止過度調整;常相

78、反，如果校正方向是相同的，對殘差絕對值變大的權重調整加快的過程。</p><p>　　考慮之間的學習率及不變動量的關系，假設：</p><p><b>　　(4)</b></p><p>　　與此同時，公式（3）將給出：</p><p><b>　　(5)</b></p><p&

79、gt;　　在一天的時間t的權數等于N：</p><p><b>　　進而有：</b></p><p>　　當 ，并假設該評價函數是光滑的，它涉及到一個結論，即學習的速度按比例。</p><p><b>　　后期充電過程的模擬</b></p><p>　　談及充電測試，如圖1所示，只有起點和點數據作為

80、后期負責記錄記錄終端，因此它是非常難以確定的電壓變化規(guī)律。雖然定性分析表明，第二個“變化規(guī)律的第三和第四階段為第一階段相同的三個階段的充電時間線性仿真看來是比較短的。第五階段的電流非常短，所以也可以被認為是第五階段模擬線性電壓變化規(guī)律。</p><p>　　充電過程五個階段的仿真研究</p><p>　　在建模實現以前，它應該是電池的性能狀態(tài)非常好，工作條件是完美的，這是詳細描述如下：&l

81、t;/p><p>　　電解液的工作狀態(tài)正常；</p><p>　　每個階段電流保持不變；</p><p>　　每個電池都在相同的狀態(tài)工作；</p><p>　　對第一階段的充電過程神經網絡模型的建立</p><p>　　因為，第一階段的充電過程中的數據點在不同的放電率前數在圖1所示不一樣，同時，還需要大量數據質量為導師信號

82、的輸入向量，充電測試樣條插補延長至第一另一方面測試的數據擴展保證了維的一致性，數據，輸入向量。</p><p><b>　　網絡化的參數設置</b></p><p>　　在開始凈值是很難融合時，可通過網絡初始權指定程序本身。</p><p>　　考慮到該系統(tǒng)是非線性的，因此收斂可能主要由初始權值的選擇決定的。培訓后的凈是遭受多次失敗，為了使網是

83、融合，凈初始權被指定的隨機數字，范圍從0到1，這使得神經元在調整權重最大的位置傳遞函數。該仿真實驗結果表明，凈來收斂的改善時迅速采用。的凈相關參數設置如下（泰蕾茲參數實現的凈設置堅持的NNTool的默認值。）：</p><p>　　該方案是一個網，包含輸入層，隱層和輸出層學習算法實現。隱藏的層傳遞函數指定為功能，以及輸出層傳遞函數是潺潺的功能。的凈輸入向量為X=（X1，X2，X3的）=（收費期間，前放電速率，和充

84、電時間），凈輸出是不同的時刻不同放電率前者在一定電壓充電時間。很明顯，在輸入層和輸出層的神經元數目應該分別為3和1。</p><p><b>　　神經網絡</b></p><p>　　中間有幅度的輸入向量，組件的順序一定的差異，導致凈值精度不能值得稱贊的滿足，所以輸入和輸出數據的文件歸第一[5]。此外當神經網絡算法帶來了建模推進，對隱含層神經元數的選擇是一個關鍵因素。

85、當網太小，學習能力變弱，凈是很難收斂。常相反如果網太大將導致大的自由程度，多余的住宿會發(fā)生。通過大量的模擬實驗[6]，對神經元的數量始終趨緊于9。該網仿真結果和試驗數據對比的圖像顯示，如圖3（圖中的星號表示的測試數據），最大誤差為0.4個百分點的數據表明，該仿真效果非常逼真。除了網絡訓練誤差曲線圖4顯示。最后，培訓網是保存為一個墊子數據文檔，以便在將來調用。</p><p>　　圖3 凈值訓練數據和測試數據對比的

86、圖像</p><p>　　圖4網絡訓練誤差曲線</p><p>　　圖5顯示了在不同的前第一階段排放即通過培訓獲得的凈數據率的變化過程曲線。</p><p>　　圖5第一階段的充電曲線</p><p><b>　　后階段的的建立</b></p><p>　　人們很容易得出一個結論，即后來的階段充電

87、時間幾乎是在同一前放電率相同的電荷通過試驗曲線，但根據不同的放電率前的充電時間是不一致的。下階段，1小時后，5小時和50小時，充電時間均采用我表（數據恢復正常。）在休息前放電率的充電可以通過樣條內插法求得倍。</p><p>　　此外，充電后期開始在相同的電壓放電前者如表二顯示率不同。在同樣的方式，其余下的放電率前后期開始充電電壓可通過插值。</p><p><b>　　表一&l

88、t;/b></p><p>　　在后階段不同前任放電頻率下的充電時間</p><p><b>　　表二</b></p><p>　　后期不同前任啟動電壓下的電荷放電速率</p><p>　　全體五個階段的電池充電過程的建模</p><p>　　當第一階段充電過程模型與模型集成充電后期，整個五

89、個階段的電池充電過程中得到，仿真結果對比的圖片和測試數據，如圖6所示（圖中的星號表示的測試數據），以及在3個小時放電率充電過程的模擬輸出還展出如圖7所示。</p><p>　　圖6 模擬結果對比的圖片和測試數據</p><p>　　圖7 3小時放電率下電荷模型仿真輸出的</p><p><b>　　結論</b></p><