基于mRMR的发电机DCS信号神经网络趋势预测方法

摘要

针对神经网络方法在发电机信号趋势预测过程中如何合理选择网络输入节点的问题,提出了一种基于最大相关和最小冗余(mRMR)算法的神经网络输入信号选取准则.该方法研究了发电机分布式控制系统(DCS)监测数据的特点,采用了mRMR算法从原始特征集合中选择了与被描述对象具有最大相关性,且特征集元素间冗余量最小的特征子集作为网络输入,进而有效地提高了网络模型对输入输出间非线性函数关系的拟合精度.研究结果表明,当对某电厂DCS信号进行分析时,与直接利用神经网络进行趋势预测的准确性相比,该方法预测准确性高、泛化能力好,具有良好的工程适用性.%Aiming at the problem that how to select the network input node reasonably in the process of the trend prediction of the generator signals,an input signal selection criteria method based on max-relevance & min-redundancy (mRMR) was proposed.The characteristic of generator distributed control system(DCS) monitoring data was researched,and the feature subset which had the maximum correlation with the described object and minimun redundancy between feature elements from original feature set was selected as the network input through mRMR,and then the fitting precision of the nonlinear function between input and output of network model could be improved effectively.The results indicate that compares with the accuracy of using neural network to trend prediction directly,the proposed method has high accuracy and good generalization ability,therefore has good engineering applicability.