Pre-Trained Models and Generative Adversarial Networks for Non-Intrusive Load Monitoring

Non-intrusive Load Monitoring (NILM) is an essential method for obtaining fine-grained power consumption information, which can significantly improve the level of advanced power grid application services, such as demand response, load forecasting, and user behavior analysis. However, when conducting transfer learning research, the current NILM model usually relies on existing source domain data for pre-training and then fine-tunes the model with a small amount of target domain data to adapt to new scenarios. Although this method is effective, it often requires additional data collection and substantial computational resources, which limits the popularity and efficiency of NILM in practical applications. To solve these problems, an innovative method that combines transfer learning and generative adversarial networks (GANs) is proposed. This method first employs a ResNet model pre-trained on a large-scale dataset to extract high-dimensional features of load signals. It then utilizes a fine-tuning strategy on the target task to adjust the model, ensuring both the effectiveness of feature extraction and the adaptability to the new classification task. Secondly, the GAN is introduced to generate high-quality appliance Voltage-Current (V-I) trajectory features, expand the diversity of the training dataset, alleviate the problem of scarce target domain data, and further improve the generalization ability of the model. Experimental results on the PLAID dataset demonstrate that the method combining transfer learning and GAN significantly enhances the accuracy and stability of NILM models in the task.