Decision focused online learning for real time energy aware scheduling of interconnected data centers with photovoltaic generation and battery storage

Data centers require intraday scheduling that balances energy costs with renewable consumption. However, existing methods struggle with multi-source uncertainties in forecasting electricity prices, rooftop PV generation, and task arrivals. We propose a decision-focused online learning optimal scheduling method, which embeds the downstream optimization into forecasting model training, using decision loss rather than prediction error as the objective so forecasts directly serve scheduling quality. The method integrates forecasts of electricity prices, PV generation, and task arrivals: day-ahead training on historical samples captures long-term trends, while intraday rolling execution with delayed parameter updates—executing only the first action per step and updating after true values are observed—fully exploits real-time information. A regularized surrogate loss handles the non-differentiability of storage charging–discharging complementarity constraints during backpropagation. Experiments on real operational data from a data center cluster (June–December 2025) show monthly energy cost reductions of \(16.5\) – \(32.4\%\) over conventional predict-then-decide methods, and cumulative PV consumption improving from \(64.1\%\) to \(78.3\%\) over seven months, confirming that decision-focused online learning effectively leverages real-time measurements to continuously improve both scheduling economy and renewable energy utilization.