To solve the problems of complex feature extraction, low recognition accuracy and model weight reduction in the recognition of complex power quality disturbance (PQD), a method of PQD signal visualization using recurrence plot (RP) and a model training method based on knowledge distillation are proposed. This method first mines the hidden features of PQD signals based on RP and builds image data sets. It then uses a deep residual shrinkage network (DRSN) to extract deeper features of image data sets and complete classification. Finally, based on knowledge distillation (KD), it lets the trained DRSN guide the training of the lightweight network MobilenetV3, and realizes cross network transmission of knowledge through distillation. The simulation experiment and hardware experiment show that MobileNetV3 trained by knowledge distillation can achieve high-precision and lightweight composite disturbance recognition, and the accuracy can be improved by 1.06% in a 30 dB noise environment. It has good recognition effect on actual disturbance signals and good noise robustness.