Nonintrusive load monitoring（NILM）is an important application area of big data and artificial intelligence

which can significantly improve the intelligence level and energy-saving effect of the power grid. For a long time in NILM

the advantage when using steady-state characteristics for load decomposition is that it can identify loads with similar power

but it cannot handle multi-state loads. To this end

a sliding time window is used as an event detection algorithm

and a NILM model based on dynamic time warping（DTW）with multi-state features is proposed. Firstly

the model extracts the characteristics of the multi-state load

and establishes a steady-state waveform template library of the multistate characteristics. Then

it uses the sliding time window algorithm to extract the steady- state waveform characteristics of the load to be decomposed

and applies the DTW algorithm to the steady-state waveforms. The load characteristics in the state waveform template library calculate the minimum distance for identification. The proposed load decomposition model can significantly improve the decomposition ability of multi-state loads under steadystate characteristics. Lastly

the test based on the REDD dataset verifies the effectiveness of the method.