Towards cost-optimal joint electricity-computation management: A novel predict-then-optimize framework

The escalating computing demand due to the flourishing of artificial intelligence is catalyzing more comprehensive and intricate interactions between modern power systems and data centers (DCs), necessitating joint electricity-computation management towards cost-optimal operation. The power system operator (SO) dispatches the generators, and the DC operator (DCO) optimizes the server dispatch strategies, where coupled information interactions exist. In practical, SO and DCO would encounter uncertainties arising from power outputs of renewable energy sources (RES) and computing workload requests submitted by end-users, respectively. Conventional accuracy-oriented predict-then-optimize (PTO) framework may lead to sub-optimal solutions due to the asymmetric relationship between prediction error and decision error. To achieve cost-optimal dispatch strategies, developing a cost-oriented PTO decision-making framework for the joint management is essential. Specially, the prediction models are trained by minimizing the decision regret. In addition, a privacy-preserving dual-boundary feedback-embedded adaptive iterative algorithm is specially proposed to solve the joint dispatch problem, realizing guaranteed and faster convergence. Simulation results on a modified IEEE-30 bus system over extensive scenarios demonstrate that the cost-oriented PTO framework saves about 1.4% of the total operational cost compared to conventional accuracy-oriented decision framework on average. Moreover, the proposed iterative algorithm averagely reduces 20% of iteration times than the existing binary search method.