Non-intrusive load monitoring (NILM) decomposes aggregate load measurements into individual appliance profiles using only single-point meter data. Power sequences in NILM exhibit superimposed characteristics across multiple time scales

containing both short-term transients and periodic patterns. However

existing methods predominantly rely on single-time-scale and single-time-domain architectures

making it difficult to capture cross-scale features effectively or to leverage prior information from the frequency domain. To mitigate the aforementioned limitations

this study presents TF-SeqFormer

a sequence-to-sequence Transformer framework based on joint time-frequency modelling. Specifically

the proposed model establishes a unified time-frequency joint perception framework that employs parallel branches to achieve aligned modelling and synergistic representation of both feature types

thereby mitigating the limitations of single-domain representation. Within the time-domain branch

a multi-scale convolution cross attention (MSCA) module is designed to capture the coupled characteristics of local load transients and global temporal trends. Simultaneously

the frequency-domain branch incorporates a frequency linear modulation (FLM) module that retains and modulates selected low-frequency RFFT coefficients using learnable complex channel-by-bin weights

while an adaptive gating mechanism is employed to fuse the resulting frequency representation with the time-domain features. Experimental results on the UK-DALE dataset show that TF-SeqFormer ranks first in four of the five averaged evaluation metrics and second in the remaining metric

reflecting a strong overall balance among appliance-state identification

power reconstruction and aggregate-energy estimation.