A Lightweight Identification Method for Complex Power Industry Tasks Based on Knowledge Distillation and Network Pruning

Lightweight service identification models are very important for resource-constrained distribution grid systems. To address the increasingly larger deep learning models, we provide a method lightweight of complex power services based on knowledge distillation and network pruning. Specifically, pruning Taylor expansion is first used to rank importance parameters small-scale delete some parameters, compressing model reducing amount operation complexity. Then, migrate from large-scale ResNet50 so that can fit soft-label information output neural through loss function complete migration network. Experimental results show this compress size small improve recognition accuracy. Compared with original network, accuracy improved by 2.24 percentage points 97.24%. The number compressed 81.9% floating-point operations 92.1%, making it more suitable deployment in devices.