能源互联网关联数据融合的互信息方法

A Mutual Information Method for Associated Data Fusion in Energy Internet

信息物理融合框架下的能源互联网需要处理的数据是海量的,要从中提取知识或分析数据之间的关联特征难度很大。在此背景下,基于互信息(mutual information,MI)理论,将信息融合理论中的"数据—特征—决策"三层结构应用到能源互联网的海量监测数据中,构建了一种基于多层模式的数据融合方案。互信息方法能够度量条件属性与决策属性间的相关性、消除冗余特征,从而提取规则、形成知识。首先,采用互信息方法发现海量监测数据间的关联度,并在数据预处理过程中筛选出关联特征。接着,采用多层前馈神经网络(multiple-layer feedforward neural network,MLFNN)对海量数据进行决策融合。之后,将该方法与在大规模数据集并行计算领域中发展起来的著名的MapReduce模型相结合,构造能够处理海量数据融合的"MutualInformation-Multiple-layer Feedforward Neural Network-MapReduce"(3M)方法框架。最后,以风电场功率预测问题为例来说明所提出的方法。计算结果表明,与传统的变量筛选方法相比,所提出的方法在预测精度和计算效率方面都有明显改善。

In the framework of a cyber-physical system, the amount of data in an energy internet needs to process is massive, and it is very difficult to extract knowledge and analyze the associated characteristics among data. Based on the mutual information （MI） theory, an information fusion structure with a＇ data-characteristics-decision＇ three-layer framework is applied to the massive monitoring data of energy internet, and a multi-layer mode data fusion scheme is then presented. The MI method can measure the correlation between Condition attributes and decision attributes and eliminate the redundant features, and then to extract rules and form knowledge. First, the MI method is used to determine the correlation degrees among massive monitored data, and extract the associated features in the procedure of data preprocessing. Then, the multi- layer feedforward neural network （MLFNN） is used for fusion of decision-making with massive data. The proposed method is then combined with the well-known MapReduce （MP） model in the field of parallel computing for large-scale data sets so as to figure out a ＂ Mutual Information-Multiple-layer Feedforward Neural Network-MapReduce＂ （3M） methodological framework for the fusion of large amount of data. Finally, the output power forecasting of a wind farm is served for demonstrating the presented method, whose calculation results show that the proposed method is of better forecasting accuracy and computational efficiency, compared with the traditional variable selection method.