In coming years, energy systems are likely to be organized as a heterogeneous system, where independent energy communities coexist with a main market. The price of electricity exchanges between communities and the main system should then reflect not only production costs, but also network constraints, potential distribution congestion and incentives towards local flexibilities. This price would necessarily be local, and setting it optimaly would require an all-knowing operator. The present contribution aims to investigate the potential of reinforcement learning to predict this exchange price. A minimalist case study is introduced to improve the interpretability and generalizability of the results obtained. In particular, learning speeds will be studied in order to discuss the volume of data required to guarantee a given level of performance. The transfer of trained algorithms from one case study to another will also be discussed.