Non-intrusive Load Decomposition Based on Progressive Model Structure and Temporal Information Embedding
Existing non-intrusive load monitoring algorithms have no regard for multiple-time-scale electricity consumption patterns,and have a high rate of misjudgment of electric appliance conditions and a big power prediction error.This paper optimizes existing models in terms of learning framework,information embedding,and loss functions,and proposes a load decomposition method based on progressive model structure and temporal information embedding.The model consists of a pre-decomposition module and a power prediction module,thereby progressively completing the determination of the on/off states of the electric appliances and the estimated power value.The two modules use Transformer-based network structures,and are optimized by different composite loss functions.Furthermore,the multi-scale time information encoding and embedding methods are proposed to enhance the feature extraction ability of power consumption behavior.The effectiveness of the proposed method is verified on REDD and UKDALE datasets.
non-intrusive load decompositiondeep learningself-attention mechanismTransformer model
万方数据
维普
