基于改进BERT的多头自注意力非侵入式负荷分解方法OA

To address the challenges of inadequate load feature extraction and insufficient decomposition accuracy in non-intrusive load monitoring(NILM)methods,this paper proposes a multi-head self-attention-based approach named frequency and temporal attention-BERT(FAT-BERT),which leverages an enhanced bidirectional encoder representations from transformers(BERT)architecture.First,the time-domain data are converted into frequency-domain representations via Fourier transform,while multi-scale convolutional layers are adopted to comprehensively extract temporal and spectral features of load signals,thereby strengthening the model's capability to characterize diverse load signatures.Second,a frequency-enhanced attention mechanism is integrated into the multi-head self-attention framework to amplify the model's awareness of frequency components in sequential data,which effectively refines the representation of complex load patterns.Concurrently,localized self-attention is incorporated into the modified BERT architecture to eliminate redundant global computations and accelerate operational efficiency.Furthermore,residual connections combined with regularization techniques are implemented to stabilize the training process and enhance overfitting resistance.Extensive experimental evaluations conducted on the REDD and UK-DALE benchmark datasets demonstrate the superior performance of the proposed method.The results quantitatively confirm significant improvements in decomposition accuracy and computational efficiency compared to state-of-the-art baselines,validating the practical effectiveness of FAT-BERT in NILM applications.