Digital-Twin Attention-Augmented Adaptive Model Predictive Control for Degradation-Aware Residential Battery Energy Management

Residential battery energy management systems increasingly rely on model predictive control (MPC), yet most schemes assume a fixed horizon, a noise-free state and an uncalibrated forecaster. This paper proposes DT-AAMPC, an adaptive MPC that combines a Kalman-filtered battery digital twin with an attention-BiLSTM forecaster (TAB-Net) whose attention entropy yields a confidence signal that adapts the controller’s horizon, uncertainty penalty, and state-of-charge back-off online. We evaluate it using corrected code executed on the real dataset at a reduced but realistic validation scale (three seeds, ten training epochs; full protocol: five seeds, forty epochs), against a PID/droop controller, a naive-forecast MPC, and a fixed-horizon ablation (Fixed-DT-MPC) across six household stress-test scenarios. Averaged over eighteen matched blocks, DT-AAMPC lowered net cost to 104.83 currency units (against 149.29, 170.79 and 106.49 for PID, Classical MPC and Fixed-DT-MPC) and cut state-of-health loss to 0.69% (against 1.70%, 2.36% and 0.73%), significant against both classical baselines after Holm–Bonferroni correction (p<0.001); DT-AAMPC was not, however, significantly different from Fixed-DT-MPC on any headline metric (p>0.11), so online adaptation itself is not yet shown to drive the benefit, and self-sufficiency/comfort violation were not significant against either classical baseline. A previously unflagged bug, where ambient temperature was in Fahrenheit, silently treated as Celsius, was found and fixed here; the thermal interlock then never activates across 72 runs, and equivalent-cycle counts fall to a physically unremarkable 0.8, −4.1 per six-day window (against 45–90 previously). Calibration analysis shows the confidence signal is uncorrelated with realised forecast error (ρ=0.047, p=0.568), moderating the confidence-adaptive claims.