TF-LGNet: A Time-Frequency Guided Local-Global Feature Network for Non-Intrusive Load Monitoring

Non-intrusive load monitoring (NILM) remains challenging due to the non-stationary and multi-scale characteristics of aggregated power signals. To address these issues, this paper proposes a time-frequency guided feature learning framework that integrates wavelet-based signal decomposition with multi-scale temporal modeling. A wavelet packet transform (WPT) is first applied to obtain a time-frequency representation of the input signal. An adaptive wavelet weighting (AWW) mechanism is then introduced to selectively emphasize informative frequency subbands. To further capture temporal patterns at different resolutions, a cross-scale convolutional pyramid (CSCP) is designed to fuse multi-scale local features through parallel dilated convolutions and attention-based weighting. The resulting features are finally aggregated using a global temporal modeling module for appliance-level power disaggregation. Experimental results on three public datasets show that the proposed method achieves competitive accuracy compared with state-of-the-art methods.