用BP神经网络提高锂离子电池化成系统采样精度  

Improvement of sampling precision in Li-ion battery formation system by using BP neural network

作　　者：吴免利[1] 李劼[1] 肖昕[2] 邹忠[1] 

机构地区：[1]中南大学冶金科学与工程学院,湖南长沙410083 [2]中南大学信息科学与工程学院,湖南长沙410083

出　　处：《中南大学学报（自然科学版）》2010年第1期55-59,共5页Journal of Central South University:Science and Technology

基　　金：国家"十一五"科技支撑计划项目(2007BA12B01)

摘　　要：针对自行设计的YX-20A型锂离子电池化成柜采样精度不高的问题,分别采用动量梯度下降法和L-M优化法以三层BP神经网络为预测模型对采样电流数据进行校正;并用校正后的采样数据通过TL494芯片调节设定基准和充放电电流实测值的偏差。研究结果表明:L-M算法能快速收敛,效果优于动量梯度下降法,当隐含层节点数为9时,L-M算法效果最佳;校正后的电流最大相对误差由原来的5%降到1.1%左右,平均误差小于0.5%;校正后基准电流和实测值间的相对误差波动较平缓,其最大相对误差比校正前有明显下降。Aiming at the solution of low sampling precision problem of developed YX-20A Li-ion formation equipment, two improved algorithms of three layers back-propagation neural network, namely gradient descent with momentum and Levenberg-Marquardt optimization, were introduced as forecasting models to correct the sampling electric current data; then the corrected sampling data were used to adjust deviation between basic set-point values and measured ones through TL494 chip. The results show that Levenberg-Marquardt optimization with 9 nodes in its hidden layer has the advantages of faster learning rate and higher precision; the maximum relevant error between former electric current and corrected values declines from about 5% to 1.1%, and the average relevant error is less than 0.5%; the corrected relevant errors between measured values and basic set-point values fluctuate gently, and the maximum relevant error goes down evidently by amelioration.

关 键 词：L-MBP神经网络 锂离子电池 化成系统 

分 类 号：TM910.6[电气工程—电力电子与电力传动]