This work proposes a corruption-aware predictive control strategy for bidirectional battery–converter systems for electric vehicles (EVs) using digital-twin-based simulation modeling. The proposed approach combines a nonlinear averaged model of the battery-converter coupling, a two-node thermal network, multi-state degradation dynamics, structured corruption in the measurement model, residual-based correction of state estimates, and finite-horizon health-aware predictive control into a single architecture. The electrical part of the model considers the interactions of inductor current (iL), Direct Current (DC)-link voltage, battery polarization voltage, and state of charge (SOC), while the thermal part models the junction temperature, capacitor temperature, the switching resistor variation, increased inductor resistance, increased capacitor Equivalent Series Resistance (ESR), reduced capacitance on the DC link, and battery capacity reduction. The residual-learning approach is exploited to incorporate model mismatch and improve estimates despite measurement corruption. The numerical results indicate the advantages of the residual learning twin over the physics-only approach in terms of Root Mean Square Error (RMSE) reduction, achieving RMSE reductions of 68.88%, 62.55%, 79.17%, and 60% for inductor current (iL), DC-link voltage (vdc), junction temperature, and switch on-state resistance drift, respectively. Under closed-loop control, the proposed scheme yields precise DC-link voltage (vdc) tracking with a steady-state error of 0.00683 V while also minimizing the maximum junction temperature to 54.1 °C and cumulative damage of 0.6372. Compared with Proportional–Integral (PI) control, the proposed approach reduces the peak junction temperature by 14.47%, cumulative damage by 56.95%, and steady-state voltage error by 74.63%, while increasing the final health index by 52.58%. The ablation study clearly showed that the complete framework has better electrothermal performance as well as simulated degradation and health performance when compared with the physics-only model, where total RMSE was improved by 80.48%.
Underground hydrogen storage is increasingly positioned as a long-duration flexibility option for renewable-dominated energy systems. Dependable deployment, however, is constrained by coupled uncertainties spanning multiphase flow, pore-scale interfacial physics, reactive transport, and potential microbiological effects. To improve the applicability and generalisability of physics-informed artificial intelligence for underground hydrogen storage, we conducted a PRISMA 2020 systematic literature review of studies published from 2020 to 2026. The Scopus and Google Scholar searches retrieved 1071 records. After deduplication, 872 unique records remained, and 676 full texts were assessed for eligibility. From these, 51 primary studies relevant to PI-AI for UHS were retained for synthesis after an evidence-quality appraisal gate, comprising 49 UHS-explicit studies and 2 methodologically adjacent studies, retained for transferable PI-AI insights. We examined the complexity of applying physics-informed learning across pore-to-field scales and how this complexity contributes to barriers to transferability, reproducibility, and decision readiness. Specifically, we synthesised evidence across five themes: pore-scale interfacial physics; reactive transport, sorption, and biogeochemistry; reservoir-scale surrogate architectures; uncertainty and data assimilation readiness; and multi-objective design and operational optimisation. Each study was also coded by task, model family, uncertainty practice, and validation strength. Only 9 of the 51 included studies (17.6%) reported predictive intervals or distributions consistent with quantified predictive uncertainty (UQ-A), and just one of these reported fully calibrated empirical-coverage checks, identifying uncertainty quantification as the most consequential evidence gap in the reviewed corpus.
