Evaluating Preprocessing Techniques for Target Labels in Non-Intrusive Load Monitoring

This paper investigates the impact of various preprocessing techniques on target labels in Non-Intrusive Load Monitoring (NILM). NILM, a key strategy for real-time energy consumption feedback, relies on analyzing smart meter data to track household device power signatures. The data often exhibits a right-skewed distribution, which can lead to performance issues in deep learning models. This study evaluates seven preprocessing methods, including four scalers and three transformers, to determine their effects on label distribution, model performance, and training time. Our research reveals that scalers, especially the Robust scaler, outperform transformers by preserving the original data distribution more effectively, leading to enhanced performance. The findings emphasize that choosing the right preprocessing technique for target labels is critical for optimizing model performance and training efficiency in NILM. By identifying the most effective preprocessing methods, this study provides valuable insights for enhancing NILM research and implementation, with implications for energy conservation and real-time consumer feedback.