Non-Intrusive Load Monitoring (NILM), or energy disaggregation, involves predicting the use of individual appliances based solely on the total energy consumption of a household. This device-level data, obtained without installing multiple meters, allows users to monitor their consumption, detect faulty components and make better decisions based on the new insights. Deep learning methods have shown their superiority compared to other methods in recent literature. However, they often lack the ability to capture consumer behavior to increase their robustness. One way to achieve this is to use a generic data augmentation technique based on predefined thresholds. In this article, we propose a statistical approach capable of dynamically generating synthetic data and improving the performance of a given model. The results obtained on the datasets used show that this method achieves higher performance by generating synthetic data outperforming current state-of-the-art data augmentation techniques.