摘要
为提高烟气轮机状态趋势预测的精度,提出一种改进Elman神经网络的趋势预测方法。首先,引入以四分位数和四分位距为基础的箱线图方法判别异常值,并对缺失的数据进行插补,为状态趋势预测提供可靠的全数据序列;其次,根据设备运行状态数据序列具有的时间依存性,计算数据序列不同时延的相关程度,以相关系数最大值点对应的时延为最优预测步长;最后构建三层最优预测步长Elman神经网络对烟气轮机运行状态全数序列进行趋势预测实例分析。研究结果表明,箱线图法能够简捷快速、直观明了地判别异常值;邻近点中位数插补方法更贴近原始数据分布规律,为最优插补方法;相较其他预测步长的Elman神经网络预测方法,最优预测步长的预测误差最小、预测精度最高;同时,Elman神经网络最优预测步长方法的预测误差较BP、RBF神经网络更小、预测精度更高。改进Elman神经网络趋势预测方法能够为烟气轮机的状态趋势预测提供一种有效的预测方法,该方法还可应用于其他关键设备的趋势预测中。
In order to improve the accuracy to predict state trend of flue gas turbine,a trend prediction strategy based on improved Elman neural network was proposed. Firstly,the boxplot method based on the quartile and the interquartile range was adopted to determine the outliers,and the missing data was interpolated,which provided a reliable full data sequence for state trend prediction. Secondly,the correlation degree of different delays of the data sequence was calculated according to the time dependence of the flue gas turbine operating state data sequence,and the delaycorresponding to the maximum value of the correlation coefficient was set as optimal prediction step size. Finally, a three-layer Elman neural network with the optimal prediction step size was constructed to predict the state trend of the full data sequence for the flue gas turbine. The case analysis shows that the boxplot method can distinguish the outliers in the data, quickly and intuitively. The interpolation method of neighboring point median is more close to the distribution law of the original data,which is the optimal interpolation method. The Elman neural network prediction method with the optimal prediction step size has the smallest prediction error and the highest prediction accuracy compared with other prediction step size methods. And the prediction error of the Elman neural network with the optimal prediction step is smaller than that of the BP and RBF neural networks,and the prediction accuracy is higher. The improved Elman neural network trend prediction method can provide an effective prediction method for the state trend prediction of flue gas turbines. This method can also be applied to trend prediction of other key equipment.
作者
陈涛
王立勇
徐小力
王少红
CHEN Tao;WANG Li-yong;XU Xiao-li;WANG Shao-hong(Key Laboratory of Modern Measurement & Control Technology Ministry of Education,Beijing Information Science and Technology University,Beijing 100192,China)
出处
《广西大学学报（自然科学版）》
CAS
北大核心
2019年第2期367-375,共9页
Journal of Guangxi University（Natural Science Edition）
基金
国家自然科学基金资助项目50975020
北京市教委科技计划项目(KZ201611232032
KZ201611232004)
关键词
异常值判别
缺失数据插补
最优预测步长
ELMAN神经网络
状态趋势预测
outliers distinguish
missing data interpolation
optimal prediction step length
Elman neural network
state trend prediction
