Adaptive Digital Twin synchronization under bandwidth constraints: A quantile forecasting and Kullback–Leibler divergence approach

Maintaining real-time synchronization between physical assets and their Digital Twins over bandwidth-limited wireless links is challenging. Existing approaches either transmit all sensed data, resulting in unnecessary network overhead, or rely solely on forecasting, which cannot guarantee fidelity when operating conditions change. In response, we propose an uncertainty-aware, event-driven synchronization framework that preserves twin fidelity while reducing network traffic. A Gated Recurrent Unit (GRU) trained via quantile regression forecasts key twin states and produces 10th-90th percentile prediction intervals; an online quantile-coverage test tracks predictive reliability, and a Kullback--Leibler (KL) divergence monitor detects distributional drift. During stable periods, the twin runs in predictive mode: only a subset of real states is fetched from the physical asset, while the remainder is inferred locally. If either monitor signals deviation, the system switches to direct synchronization, updates all states, and retrains the predictor when prediction reliability falls below the required threshold. Evaluations on public LTE and 5G datasets achieve R^2 > 0.80 in most cases, with a peak of 0.99. A multi-output variant forecasting two parameters matches single-output accuracy, highlighting the utility of multi-output regression for adaptive synchronization. On-device tests with a commercial smartphone reduce network traffic by 5.41%-7.01% compared to direct synchronization and lower mean absolute error (MAE) by about 64% on average relative to simpler predictive baselines. Estimated operational cost savings range from $3.84-$38.40 per device annually, indicating substantial cost reductions at scale. Overall, the framework combines calibrated forecasting with drift-aware control to enable high-fidelity digital twinning under strict bandwidth constraints.