Propriétés algorithmiques des simulations de Smartgrids sur CPU-GPU

The growth of distributed energy resources implies a rise of constraints on the electricity market computation. Indeed, the increasing number of agents and their difficult forecasts may overload the system operator. The decentralization of computation may allow management to scale. Nevertheless, the simulation of this kind of management will face prohibitive computation time as the size increases. This article presents a hardware-software co-design approach to solve this issue on a single machine. Several implementations of different algorithms will be done, analyzed, optimized, and compared. The decentralization algorithms of the Alternating Direction Method of Multipliers (ADMM) and of Proximal Atomic Coordination (PAC) has been partitioned on Central and Graphic Processing Units (CPU-GPU). In the specific case of the Peer-to-Peer market, the use of an consensus ADMM converges faster (in time and number of iterations) than the PAC. Parallelization speeds up calculations, but remains slower than using ADMM.