Energy disaggregation (or Non-Intrusive Load Monitoring – NILM) is the task of estimating the electricity consumption of each appliance in a household from the total electricity consumption. Disaggregated consumption gives information on each appliance and helps to find ways to reduce a household’s energy consumption. Recent progress in deep neural networks for computer vision and natural language processing gives inspiration to train general architectures on time series data in order to improve the state of the art on NILM, but lack of supervised data is one of the main problems stalling the improvement of disaggregation algorithms. In this paper, we introduce a new multi-agent based simulator that enables to generate synthetic data according to real time-use surveys. This synthetic dataset is used as a training set in the NILM learning process: we show that this data augmentation improves the accuracy of the disaggregation. In addition, we present four neural network architectures to estimate appliances consumption and establish a baseline architecture on the data coming from the synthetic generator.