Incorporating Knowledge Distillation Into Non-intrusive Load Monitoring for Hardware Systems Deployment

As an information provision technology, non-intrusive load monitoring can extract detailed load consumption separately from the mains reading, which is conducive to promoting residential consumers to join the demand-side response. The current deep learning model has been proven to be effective in NILM, but the huge amount of trainable parameters limits its implementation in hardware systems. Therefore, the knowledge distillation method is introduced into NILM. In this article, bidirectional gated recurrent unit (BiGRU) based on the spatial attention mechanism is first constructed as a teacher model to guide student model learning. Then, a small-scale student model is constructed, and the training of the model is implemented under the guidance of the teacher model. The experiments confirm that the proposed teacher model achieves better results than other methods, and the student model can also obtain results close to the teacher model, which provides a research direction for the deployment of deep learning models in hardware systems.