DistilQ-NILM: A Hybrid Quantization-Knowledge Distillation LSTM Model for Edge-Based Load Monitoring

In addition to adding information to users when using load monitoring, non-intrusive load monitoring (NILM) based on long short-term memory (LSTM) can also be useful for determining strategic follow-up to electrical power information. However, applying LSTM to NILM can still increase the system's complexity, compromising real-time response. This study proposed using a compression model for responsive edge computing-based NILM. We used the NILM dataset from Kaggle, which has power data from three sources: lamps, air conditioners (AC), and plugs. Then, three different model compression methods were applied to LSTM: quantization, knowledge distillation, and hybrid quantization knowledge distillation (DistilQ). We measured the performance of the three models from the model size and accuracy loss. Then, we evaluated the performance of NILM based on its confusion matrix. The test results showed that the original LSTM has a training, validation, and testing accuracy of 0.99. The F1-score of LSTM-NILM for predicting AC, plug, and lights classes were 0.98, 0.99, and 0.99, respectively. Then, the DistilQ method gave the lowest model size and the highest compression rate (CR) among the four methods, with values of 12.9 kB and 11.8×, respectively. Finally, the quantization method gave the highest accuracy and lowest accuracy loss among the three compression models, with values of 0.987 and 0.000, respectively.