In order to balance production and consumption on the power grid, efforts are mostly made on the production side. Demand response is a classical solution consisting in harnessing electricity consumption to allow the introduction of renewable energy sources and self-production in the energy mix. In this context, the main problem is to coordinate the network nodes, considered here as agents, in order for them to be able to anticipate their need and also to adjust their contribution to the collective load-shedding effort. To tackle this last point we present an original bottom-up approach of distributed load-shedding with a decentralized algorithm based on gossip protocols. These protocols offer both reliability and scalability to be used on a large scale power grid. On this ground, we built an efficient decentralized control mechanism using a self-evaluation process improving the system performances. In this paper, we present our model and evaluate it using realistic simulated data. We then discuss current limitations and further improvements as part of future work.