In this paper, we present an open-source synthetic industrial energy decomposition dataset (SIDED) generated using digital twin simulation to address the data shortage problem in the field of industrial non-intrusive load monitoring (NILM) and the complex variability of industrial energy consumption patterns. SIDED captures diverse device behaviors, weather conditions, and load profiles across three types of industrial facilities and three geographical locations. In addition, we propose a computationally efficient device modulation data augmentation (AMDA) technique that intelligently adjusts device power contributions based on the relative impact of devices to improve the generalization performance of NILM models. Experimental results show that NILM models trained with AMDA-augmented data significantly improve the energy consumption decomposition performance of complex industrial devices such as cogeneration systems. In particular, in the out-of-sample scenario, the model using AMDA achieves a normalized decomposition error of 0.093, which outperforms the model trained without data augmentation (0.451) and the model using random data augmentation (0.290). Data distribution analysis confirms that AMDA effectively aligns the training and test data distributions to improve the generalization performance of the model.
