基于粗糙集和决策树的自适应神经网络短期负荷预测方法

Short-term load forecasting using adaptive ANN based on rough set and decision tree

利用数据挖掘中的聚类技术将历史负荷数据进行聚类,根据聚类后的分类信息对定性属性利用粗糙集进行属性约简,由约简结果进一步生成决策规则树供短期预测使用。根据聚类的结果对每类进行BP神经网络的训练,神经网络的隐含层单元采取逐步试用的方法根据训练误差最小化进行选择。在实际预测中,首先对待预测的记录利用决策规则树进行归类,然后选取相应类别的神经网络予以预测。通过实例证明,该方法的平均相对误差为2.16%,而同结构BP神经网络预测的平均相对误差为2.67%,ARMA预测的平均相对误差为3.81%,证明所提方法有效。

The clustering technology of data mining is used to classify the historical toad data and the rough set is then used to reduce the related qualitative attributes of the clustered information,based on which,a decision tree is constructed to provide the rules for short-term load forecasting. The historical data of every class are used to train the BP-ANN network,the number of hidden layer nerve cells increases step by step in an experiential range until the minimal train error is achieved. The class of data for forecasting is decided by its qualitative factors according to the rules and the BP-ANN for forecasting is accordingly chosen. Example shows that, the mean relative error of and ARMA are 2.67 % and 3.81% respectively. proposed method is 2.16 %,while those of BP-ANN