Energy disaggregation, also known as Non-Intrusive Load Monitoring (NILM), aims to analyze the aggregate energy consumption of a user, such as a household, to infer the energy usage of individual appliances. Deep learning methods have become popular for this task, typically relying on supervised learning with labeled data. To reduce the need for extensive data collection, researchers have investigated various data augmentation techniques. However, it remains unclear to what extent labeled training data can be extended to facilitate energy disaggregation. To address this question, this study explores how training data diversity can be augmented from both intra-user and inter-user perspectives. Specifically, we synthesize new training samples using appliance activation data extracted from the same user (intra-user) as well as from other users (inter-user). Based on the public dataset REFIT, we show that by using the proposed intra-user and inter-user data augmentation methods, the disaggregation errors of the energy disaggregation models based on WaveNet and BERT are decreased by 6.6% and 26.8% respectively, in terms of MAE and SAE.