Abstract: Abstract: To address the shortcomings of existing deep learning–based Non-intrusive Load Monitoring (NILM) methods in capturing long-term dependencies and complex nonlinear dynamic characteristics, a NILM method combining BiLSTM-Transformer and Kolmogorov–Arnold Network (KAN) was proposed. First, the BiLSTM-Transformer module was designed by combining the advantages of the Bidirectional Long Short-Term Memory (BiLSTM) network in modeling bidirectional sequence dependencies with the capability of Transformer in capturing global context. A multi-head attention mechanism was employed to effectively capture long-term dependencies of power loads, thereby improving the accuracy of disaggregating long-cycle appliances. Then, the KAN module was introduced, which was based on the Kolmogorov–Arnold representation theorem. Through a hierarchical nonlinear mapping mechanism, complex nonlinear functions were decomposed into combinations of univariate functions, nonlinear dynamic features of load signals were more precisely represented, and the disaggregation accuracy for complex appliance patterns was improved. Experiments conducted on the REDD (Reference Energy Disaggregation Dataset) and UK-DALE (UK Domestic Appliance-Level Electricity) datasets show that, compared with the best baseline model TransUNet-NILM(NILM based on TransUNet), the proposed method achieves the maximum reduction of 7.3% in mean absolute error, the maximum improvement of 12.1% in F1-score, and an increase of 2 percentage points in accuracy. The proposed method captures long-term dependencies and nonlinear dynamic features of power load signals more accurately and improves the disaggregation of complex appliance operating modes.

Key words: Keywords: Non-intrusive Load Monitoring(NILM), Kolmogorov-Arnold Network(KAN), Transformer, Bidirectional Long Short-Term Memory (BiLSTM), network, multi-head attention mechanism

摘要： 摘 要: 针对现有基于深度学习的非侵入式负荷监测(NILM)方法在捕捉长期依赖性和复杂非线性动态特征方面存在不足的问题，提出一种融合BiLSTM-Transformer与Kolmogorov-Arnold网络(KAN)的非侵入式负荷监测方法。首先，BiLSTM-Transformer模块将双向长短时记忆(BiLSTM)网络在双向序列依赖建模上的优势与Transformer在全局上下文建模上的能力相结合，利用多头注意力机制，有效捕捉电力负荷的长期依赖特征，从而提高长周期电器负荷分解的准确度；其次，KAN模块基于Kolmogorov-Arnold表示定理，通过分层非线性映射机制能够更精确地捕捉电力负荷信号中的非线性动态特征，从而提高对复杂负载模式的分解精度。在REDD(Reference Energy Disaggregation Dataset)和UK-DALE(UK Domestic Appliance-Level Electricity)两个数据集上进行的实验结果表明，与表现最优的对比模型TransUNet-NILM(NILM based on TransUNet)的结果相比，平均绝对误差(MAE)最高降低了7.3%，F1分数最高提升了12.1%，准确率最高提升了2个百分点。可见，所提方法能更精确地捕捉电力负荷信号中的长期依赖与非线性动态特征，并显著提升复杂电器运行模式的分解效果。

关键词: 关键词: 非侵入式负荷监测, Kolmogorov-Arnold网络, Transformer, 双向长短时记忆网络, 多头注意力机制