基于k-Medoids聚类和深度学习的分布式短期负荷预测

Distributed Short-term Load Forecasting Based on k-Medoids Clustering and Deep Learning

为了获得较高的预测精度,提出一种基于k-Medoids聚类和深度学习的分布式短期负荷预测。基于配电变压器的能耗分布,采用k-Medoids聚类将电力负荷数据集中的数据进行聚类,并构建基于深度神经网络(DNN)和长短期记忆网络(LSTM)的短期负荷预测模型。在拥有1000个变电站数据子集的武汉配电网络系统中进行验证,验证结果表明,所提的kMedoids聚类可以在减少44%训练时间的基础上拟合出单个变压器预测模型的平均参数,且DNN和LSTM预测模型分别以7.32%和11.15%的平均绝对百分比误差(MAPE)跟踪实际负荷。

In order to obtain higher prediction accuracy,a distributed short-term load forecasting based on k-Medoids clustering and deep learning is proposed.Based on the energy consumption distribution of distribution transformers,k-Medoids clustering is used to cluster the data of power load data set.A short-term load forecasting model based on deep neural network(DNN)and long short-term memory network(LSTM)is constructed.It is verified in Wuhan distribution network system with 1000 subsets of substation data.The verification results show that the proposed k-Medoids clustering can fit the average parameters of a single transformer forecasting model on the basis of reducing the training time by 44%.DNN and LSTM forecast models track the actual load with mean absolute percentage error(MAPE)of 7.32%and 11.15%,respectively.