As　urbanization　accelerates,　substations　have　become　a　part　of　residential　and　downtown　areas.　The　low-frequency　noise　generated　by　transformers　has　increasingly　become　a　major　source　of　complaints　from　neighboring　residents.　Due　to　the　long　wavelength　of　low-frequency　noise,　traditional　passive　noise　reduction　technology　is　bulky,　expensive　and　limited　in　effectiveness.　Active　noise　control　(ANC)　has　gained　attention　due　to　its　compact　size,　light　weight,　low　cost,　and　excellent　ability　to　reduce　low-frequency　noise.　However,　the　precision　of　ANC　relies　on　the　control　algorithm,　and　conventional　least　mean　square-based　algorithms　still　face　challenges　such　as　sluggish　convergence,　susceptibility　to　dispersion,　and　poor　stability.　In　light　of　the　time-varying　characteristics　of　transformer　noise,　this　paper　proposes　a　machine　learning-based　algorithm　for　generating　inverse　noise　waveform　through　time-series　forecasting,　and　this　approach　aims　to　achieve　rapid　and　stable　noise　reduction.　The　experimental　result　shows　that　the　algorithm　achieves　noise　reduction　effects　of　49.67 dB,　and　the　time　used　for　training　and　prediction　is　1.78 s　and　1.64 s,　which　shows　a　promising　solution　for　transformer　noise　reduction.　©　The　Author(s),　under　exclusive　license　to　Springer　Nature　Singapore　Pte　Ltd.　2024.