Improvement of the Generalization Ability of NILM Algorithm Based on AFF

The non-intrusive load monitoring (NILM) is the basis and key to the sensing and measurement of the digital grid. However, the NILM algorithms, which rely on the training dataset, suffer from serious accuracy degradation problems when unknown appliances are present. For this reason, the appliance fictitious features (AFF) are introduced in this paper to improve the generalization capability of the NILM model. First, different construction methods of feature offset are designed for the appliance features with different characteristics, such as power and V-I trajectory. Second, the fictional feature map is generated by superimposing the fictional offset on the real appliance feature map. Finally, the resulting fictitious feature maps of appliances will be supplemented into the training data of the NILM algorithm. In this way, the problem of mutual confusion between the appliances covered and not covered in the algorithm training data during appliance identification is solved. The plug-load appliance identification dataset (PLAID) is used to evaluate the algorithm. The results indicate that after adding the AFF in the training dataset, the model’s generalization ability has been greatly improved. The recognition accuracy stabilized to 96.7% even in the presence of unknown appliances.