Energy management optimization algorithm for data center microgrids based on solid-state transformers

The proliferation of hyperscale data centers, together with tightening carbon-neutrality targets, exposes the limits of the line-frequency transformer (LFT) that still dominates data-center power distribution and motivates a fresh look at the conversion and dispatch layer. This paper develops an energy management framework for a data-center microgrid whose grid interface is a three-stage solid-state transformer (SST), so that renewable, storage, and IT loads sit on independently regulated AC and DC ports rather than behind a passive iron core. A multi-objective mixed-integer nonlinear program (MOMINLP) is formulated to jointly minimise operating cost, carbon emission, a composite power-quality index, and average PUE, subject to bus power balance, SST port transfer, battery SOC, ramp, and demand-response constraints. Uncertainty in PV, wind, and IT workload is condensed through Latin hypercube sampling with Kantorovich-based scenario reduction, and tail risk is capped by a CVaR constraint at the 95% confidence level. Dispatch is decomposed across day-ahead, intra-day, and real-time MPC layers, and the nonconvex problem is solved by an improved whale optimization algorithm (IWOA) that combines Tent chaotic initialization, an adaptive convergence coefficient, differential-evolution mutation, and Lagrangian relaxation of the diesel and battery integer commitments. On a 10 MW-class pilot the proposed scheme raises renewable accommodation by 17.2% points, cuts annual operating cost by 11.7%, reduces $$\:{CO}_{2}$$ emissions by 12.9%, and lowers average PUE from 1.52 to 1.40 relative to a line-frequency-transformer baseline, while keeping bus voltage deviation below 1.7% on the DC bus and 2.4% on the AC bus. Against PSO, GA, standard WOA, and NSGA-II, IWOA halves the iterations needed to reach 1% of the final fitness (78 versus 134–187), tightens the relative gap to the best-known solution to 0.62% (versus 2.4–5.1%), and drops the coefficient of variation to 0.018 (versus 0.061–0.094), showing that the observed gains come from the coupling of hardware-aware modelling and algorithmic tailoring rather than from either alone.