Abstract
Renewable-rich DC microgrids require energy management that quantifies photovoltaic and load uncertainty rather than merely reacting to it. This paper presents a forecast-embedded Digital Twin (DT) framework in which a hybrid BiLSTM–GPR forecaster feeds a Model-Predictive-Control (MPC) battery sizing and dispatch pipeline, propagating the GPR's predictive standard deviation into a time-varying, chance-constrained state-of-charge (SOC) reserve applied identically across four Energy Management Strategies (Rule-Based, MPC, Q-Learning, and Fuzzy Logic), evaluated over a 13-day study with pairwise Wilcoxon signed-rank tests on 200 matched-pair Monte Carlo trials. A train/inference bin-edge inconsistency in the Q-Learning controller is corrected, reducing its Monte Carlo cost variance roughly fivefold. MPC remains the lowest-cost strategy ($8.5259 versus $11.0837, $10.4079, and $11.3823 for Rule-Based, Q-Learning, and Fuzzy Logic), with 100% cost-rank consistency and a 32.65% cost reduction relative to a no-Digital-Twin baseline, though it never achieves the lowest peak grid import across the 13-day set (0.0% peak win-rate), exposing a cost-reliability trade-off. The uncertainty-aware reserve is obtained at essentially no cost, reducing MPC's cost by 4.55% with no measurable change in peak grid import relative to fixed SOC bounds. All pairwise EMS cost distributions differ significantly (p ≤ 3.00×10⁻³¹). Results are reported alongside a 17-item architectural validation checklist.

This publication has 47 references indexed in Scilit: