Non-intrusive load monitoring (NILM) can infer load profiles for each individual appliance from aggregated power consumption signals without installing extra sub-meters. However

performance of traditional energy disaggregation methods deteriorates in complex environments

especially susceptible to the presence of other high power consumption appliances. Practicalities are also limited by diversity of household load patterns and measurement errors. In order to address these problems

a hybrid deep learning model consisting of two steps is proposed in this paper. First

an improved variational autoencoder (VAE) structure is introduced for preliminary energy disaggregation

where the encoder and decoder layers are long short-term networks (LSTM) to extract temporal characteristics of active power signals. Afterward

a post-processing method based on Siamese one-dimensional convolutional neural network (S-1D-CNN) is adopted to remove incorrectly predicted activation segments of target appliances. Experiments are conducted on two public datasets

and results show remarkable improvements on prediction accuracy over other deep learning methods. Both transferability and stability of the proposed model are verified under different working conditions.