Abstract:

Aiming at the problem that the differences in signal characteristics between different electrical devices lead to the decrease in accuracy and insufficient generalization ability of traditional non-intrusive load recognition methods when facing unknown additional devices, a migration learning method based on the adversarial Discriminative domain adaptation (ADDA) is proposed. First, the multi-dimensional feature parameters of different devices are extracted by combining feature selection methods, and the source domain models are trained using the source device feature data. Then, an adversarial training strategy is employed, with a discriminator introduced to optimize the feature extractor of the target device. Finally, the model is fine-tuned and optimized to complete the load recognition task on the target device. Experimental results show that the optimized recognition model achieves an average accuracy of 98.90%, an improvement of 18.41% compared to the traditional migration learning methods.

Key words: non-intrusive load monitoring, unknown added devices, adversarial discriminative domain adaptation (ADDA), transfer learning