基于DTW K-medoids与VMD-多分支神经网络的多用户短期负荷预测

Short-term Load Forecasting Based on DTW K-medoids and VMD Multi-branch Neural Network for Multiple Users

多用户电力负荷预测是指根据历史负荷数据对多个用户或区域的电力负荷进行预测,可使电网企业掌握不同用户或区域的电力需求,以便更好地开展规划和实施调度优化等。然而由于各用户呈现出复杂多样的用电行为,采用传统方法难以进行统一建模并实现快速准确预测。为此,构建了一种基于DTW Kmedoids与VMD-多分支神经网络的多用户短期负荷预测模型。首先,采用DTW K-medoids法进行用户负荷数据聚类,利用动态时间弯曲(dynamic time warping,DTW)计算数据间的距离,取代K-medoids算法中传统的欧氏距离度量方式,以改善多用户负荷聚类的效果;在此基础上,为充分表征负荷历史数据的长短期时序依赖特征,建立了一种基于变分模态分解(variational mode decomposition,VMD)-多分支神经网络模型的并行预测方法,用于多用户短期负荷预测;最后,使用某地区20个用户365天的负荷数据进行聚类、训练和测试实验,结果显示该模型结果的平均绝对误差和均方根误差等指标均较对比模型有较大幅度降低,表明该方法可有效表征多类用户的用电行为,提升多用户负荷预测效率和精度。

Multi-user power load forecasting refers to the power load forecasting of multiple users or regions based on historical loads data,which can make the grid companies understand the power demands of different users or regions,so as to better carry out the planning and scheduling optimization of the power system.However,different users have complex and diverse power consumption behaviors,so it is difficult to use traditional methods to universally model different power users'loads and achieve accurate prediction.Therefore,a new multi-user short-term load prediction model based on DTW K-medoids and VMD-multi-branch neural network is established.Firstly,in order to improve the clustering performance of traditional clustering methods,the DTW Kmedoids method is used to cluster users'load data,and the distance between loads data is calculated using the dynamic time warping(DTW)instead of the traditional Euclidean distance measurement method in K-medoids to improve the clustering effects of multiple users'load.Secondly,in order to fully characterize the long short-term time series-dependent characteristics of load history data,a parallel load forecasting method based on VMD-multi-branch neutral network model is established for multi-user short-term load forecasting.Finally,the 365-day load data of 20 users in a region is used for clustering,training and experiment,and the results show that the MAE and RMSE indexes of the proposed model significantly decrease compared with that of the comparative models,indicating that the proposed method can effectively characterize the power consumption behaviors of multiple users and improve the prediction efficiency and accuracy of multi-user loads.