In this article, the smart grid is modeled as a decentralized and hierarchical network, made of three categories of agents: producers, providers and microgrids. To optimize their decisions concerning the energy prices and the traded quantities of energy, the agents need to forecast the energy productions and the demand of the microgrids. The biases resulting from the decentralized learning heavily affect the smart grid performance. Indeed, at their individual level, providers and producers incur additional costs in case where they cannot meet their consumer demand. We prove that there exist prices that provide to the producers a guarantee to avoid penalties, reporting all the risk on the providers. Furthermore, the providers can minimize their average regret by sharing information and coordinating forecasts. The results on a toy network confirm the theory: information sharing between the intermediary agents minimizes the average regret. As output of the simulations, we observe that the convergence rates of the collaborative learning strategy, where providers share their information through a grand coalition, are clearly superior to rates resulting from distributed learning, using external and internal regret minimization.