基于ANFIS和减法聚类的动力电池放电峰值功率预测  

Research on Discharge Peak Power Prediction of Battery Based on ANFIS and Subtraction Clustering

作　　者：孙丙香[1] 高科[1] 姜久春[1] 罗敏[2] 何婷婷[1] 郑方丹 郭宏榆 

机构地区：[1]北京交通大学国家能源主动配电网技术研发中心,北京100044 [2]广东电网公司电力科学研究院,广州510080 [3]惠州市亿能电子有限公司,惠州516006

出　　处：《电工技术学报》2015年第4期272-280,共9页Transactions of China Electrotechnical Society

基　　金：国家高技术研究发展计划(863计划)(2012AA050211)资助项目

摘　　要：动力电池的短时峰值功率预测对于实际使用来说至关重要。本文采用基于一阶Sugeno模糊推理系统的自适应神经模糊推理系统(ANFIS)模型估计放电峰值功率。选取温度、SOC和欧姆内阻为模型输入量,10s脉冲放电峰值功率为输出变量。基于实测和曲线拟合相结合的方法得到训练数据组,采用305组数据组模型进行训练,采用网格生成法和减法聚类法分别生成模糊集合,并采用单一BP神经网络方法和混合训练方法分别进行模型训练。发现采用减法聚类法生成模糊结构,能大幅减少模糊规则的数目,并提高收敛速度,在满足预测准确度的前提下降低了模型的复杂程度;采用混合训练方法进行网络学习能够加强模型的收敛能力并克服单一BP算法的局部最优问题,准确度更高。最后,采用125组数据组模型进行验证,预测误差在10%以内,基于ANFIS的模型能够很好地估计电池的脉冲峰值功率。Short-term peak power prediction of power battery is essential for practical use. In this paper, the adaptive neuro-fuzzy inference system(ANFIS) model is used to estimate the discharge peak power based on first order Sugeno fuzzy inference system. Temperature, SOC and ohmic resistance are selected as the inputs and 10-seconds pulse discharge peak power as the output. The training data pairs are obtained by the combination of measurements and curve fitting. Mesh generation and subtraction clustering methods are used to generate fuzzy sets. BP neural network and hybrid training methods are used to train the model based on 305 pairs of data. It found that the number of fuzzy rules is significantly reduced by using fuzzy clustering method to generate subtraction structure, the convergence rate is improved and the complexity of the model is reduced on the premise of meet the precision. Network learning by hybrid training method can strengthen the convergence ability and overcome the problem of local optimum by BP algorithm. Finally, 305 pairs of data is used to validate the model, the prediction is within 10%, ANFIS model can be well estimated the pulse peak power of the battery.

关 键 词：动力电池 峰值功率 ANFIS 减法聚类 混合训练 

分 类 号：TM912[电气工程—电力电子与电力传动]