With the large-scale utilization of renewable energy sources

power quality disturbances (PQDs) in power systems are becoming increasingly complex and diverse. Traditional methods struggle to simultaneously achieve type identification of multiple composite disturbances and precise timing estimation of these disturbances. This paper proposes a two-stage multi-channel network based on differential encoding and embedding to address this issue. In the first stage

a differential multi-head self-attention (DMHSA) mechanism is proposed to detect signal mutation points

allowing for the encoding of disturbance differential features. In the second stage

the original signal is combined with the encoded differential signal to create multi-channel features

and an improved temporal convolutional network

TCN-SENet

is designed for channel feature extraction and feature learning to achieve point classification of PQDs. The PQDs detection model based on these two modules can achieve efficient and accurate disturbance identification and timing estimation. In simulation experiments

the proposed model outperforms others in classification accuracy for disturbance dataset at a 30 dB signal-to-noise ratio

with an average timing error of less than 1.3 ms. In the hardware experiment

the proposed model shows the best generalization capability

significantly outperforming other models in disturbance identification accuracy and average timing error.