Beyond Price Signals: Coordinated Wholesale Market Participation for Flexible Data Centers

Data-center (DC) flexibility is typically optimised against exogenous electricity prices, ignoring the fact that real wholesale-market participation requires sequential decisions across multiple timescales. The DC is a market participant and can affect prices, especially in marginal markets such as the balancing market. We introduce a three-level reinforcement-learning controller that coordinates a detailed digital twin of a DC with an endogenously cleared, full-asset electricity-market simulation. The controller carries workload, thermal state, and prior market commitments across each decision stage, allowing it to shift computation and cooling while managing procurement across the three markets. Across five independent training runs, the full controller reduces mean effective DC electricity cost by 16.1% on a frozen transfer scenario with 2024 weather and gas-price inputs, compared with 12.3% for the strongest rule-based controller. Layer ablations confirm that the full three-layer RL controller outperforms both the L1 and L1+L2 variants. The full controller's real-time actions are also directionally aligned with renewable forecast imbalances, producing a small but consistent reduction in absolute system imbalance. These results indicate that a DC can participate in three-level wholesale markets, delivering electricity cost reductions while also improving system balancing and influencing prices in the balancing market.