In smart grids, renewable energies play a predominant role, but they produce more and more data, which are volatile by nature. As a result, predicting electrical behaviours has become a real challenge and requires solutions that involve more all microgrid entities in learning processes. This research proposes the design of a coordination model that integrates two decentralised approaches to distributed learning applied to a microgrid: the gossip federated learning approach, which consists of exchanging learning models between neighbouring nodes, and the gossip ensemble learning approach, which consists of exchanging prediction results between neighbouring nodes. The experimentations, based on real data collected in a living laboratory, show that the combination of a coordination model and intelligent digital twins makes it possible to implement and operate these two purely decentralised learning approaches. The results obtained on the predictions confirm that these two implemented approaches can improve the efficiency of learning on the scale of a microgrid, while reducing the congestion caused by data exchanges. In addition, the generic gossip mechanism offers the flexibility to easily define different variants of an aggregation operator, which can help to maximise the performance obtained.