The field of Energy Management Systems for Smart Grids has been extensively explored in recent years, with many different approaches being described in the literature. In collaboration with our industrial partner Ubiant, which deploys smart homes solutions, we identified a need for a highly robust and scalable system that would exploit the flexibility of residential consumption to optimize energy use in the smart grid. At the same time we observed that the majority of existing works focused on the management of production and storage only, and that none of the proposed architectures are fully decentralized. Our objective was then to design a dynamic and adaptive mechanism to leverage every existing flexibility while ensuring the user’s comfort and a fair distribution of the load balancing effort ; but also to offer a modular and open platform with which a large variety of devices, constraints and even algorithms could be interfaced. In this thesis we realised (1) an evaluation of state of the art techniques in real-time individual load forecasting, whose results led us to follow (2) a bottom-up and decentralized approach to distributed residential load shedding system relying on a dynamic compensation mechanism to provide a stable curtailment. On this basis, we then built (3) a generic user-centered platform for energy management in smart grids allowing the easy integration of multiple devices, the quick adaptation to changing environment and constraints, and an efficient deployment.