Power Quality Disturbances Classification Based on CNN Combining Knowledge Distillation and Incremental Learning

With the integration of renewable energy sources and new power electronic devices, power grid complexity has increased, leading to frequent power quality disturbances. Existing classification models often require retraining when dealing with new disturbance types, resulting in long update periods and significant use of computing resources. This paper proposes the FKDIL-CNN, a convolutional neural network model that combines knowledge distillation and incremental learning, enabling the enabling rapid adaptation to new disturbances without requiring full retraining. The FKDIL-CNN transfers identification knowledge for the known old types from a high-accuracy teacher model to the student model. By introducing a dynamic weight mechanism to improve the traditional knowledge distillation loss function, the student model efficiently distills old knowledge from the teacher model and performs incremental learning on new knowledge. Compared with the existing methods, our proposed FKDIL-CNN significantly reduces the training time and computing resource consumption while maintaining high classification accuracy. Additionally, this model demonstrates strong noise immunity, making it a practical solution for PQDs monitoring in complex power grid.