User Level Ultra-short-term Load Forecasting Based on Multi-scale Component Feature Learning
Considering the high volatility of user-level load and low operation efficiency or limited accuracy improvement caused by increased data dimension after one-step decomposition,a novel multi-scale component feature learning framework is proposed for forecasting user-level load.An adaptive quadratic modal decomposition framework based on complete ensemble empirical mode decomposition with adaptive noise(CEEMDAN),permutation entropy(PE),and variational mode decomposition(VMD)is constructed to capture temporal features such as periodicity and reduce their non-stationary characteristics.The multi-dimensional feature fusion method explores the coupling relationship among intrinsic mode functions and enriches the feature information.The improved multiscale spatial attention(MSA)module extracts spatiotemporal features and multi-component coupling relations along multiple scales such as time,space,and channel,thus facilitating the convolutional neural network(CNN)to learn multi-component features.Based on the actual load data of 12 users in the real estate industry,education industry,and business service industry in Nanjing,Jiangsu Province,the average absolute percentage error is 5.82％,4.54％,and 8.78％,respectively,which is reduced by 10.46％,6％and 7.48％compared with the best comparison model.It is verified that the proposed model has high prediction accuracy and good generalization performance.
万方数据
维普
