ABSTRACT Rapid AI advancements have primarily focused on enhancing digital twins (DTs) using artificial intelligence (AI), as seen in the author's earlier research addressing scaling challenges such as computing resource efficiency and manual insight generation. Conversely, existing research on using DTs to enhance AI remains nascent, mostly limited to DTs serving as data factories for synthetic training and providing on‐demand validation environments to accelerate AI model testing and training. This work proposes that DTs have the potential to enhance AI through other diverse applications and explores a mutualistic symbiosis where DTs can improve AI via context, memory, factual grounding and bridging the ‘trust gap’ through better explainability, demonstrating that AI and DTs act as essential drivers for each other's scalability and development. We demonstrate how DTs facilitate the transition of AI from stateless chatbots to persistent adaptive cognitive agents. Ultimately, the integration of these fields leads to the emergence of cognitive digital twins (CDTs), where this work identifies a potential opportunity where DTs begin to enter the realm of AI, potentially contributing to the advancement of emerging AI concepts such as world models (WMs), obscuring the traditional distinctions between the AI and DTs.
ABSTRACT High‐quality multiobjective engineering optimisation requires large candidate populations to approximate dense Pareto fronts, yet CPU‐bound algorithms such as NSGA‐II become impractical for large populations. Although GPU hardware can relax this constraint, only algorithms whose internal structure permits data‐parallel execution can fully exploit it. We present a PyTorch‐native GPU MOPSO framework that extends the Coello Coello external‐archive design with linearly decaying inertia, boundary‐safe iterative archive pruning and periodic particle mutation, and we empirically characterise how swarm size, GPU throughput, Pareto‐front quality and run‐to‐run reproducibility interact across a range of engineering optimisation problems. The principal empirical finding is that GPU hardware makes it practical to run swarms of – particles and that these larger swarms produce substantially better engineering outcomes: GPU PSO‐5000 achieves 40% lower generational distance than CPU NSGA‐II‐100 on combustion engine calibration (, Wilcoxon signed‐rank test, 20 independent seeds), whereas GPU PSO‐2000 achieves the best hypervolume on a 30‐dimensional neural‐network surrogate in 2.5 s, which is 55% faster than CPU NSGA‐II and has lower run‐to‐run variance. A scalability study from to identifies a GPU throughput sweet spot at (2.5 μs per particle per iteration), where variance collapses relative to , enabling reproducible single‐run deployment without ensemble averaging. These gains arise from two compounding effects: GPU batch inference reduces per‐particle surrogate evaluation cost by over CPU, whereas PSO’s parallel update rule avoids the nondominated sorting bottleneck that prevents NSGA‐II from scaling beyond . A controlled GPU‐vs‐GPU comparison isolates this second effect, revealing a timing gap at attributable solely to algorithmic structure. Results generalise across a real‐world UCI Gas Turbine dataset, a 6D/4‐objective engine calibration problem and a fixed 60‐s wall‐clock budget protocol under which GPU PSO’s 5‐s convergence enables 12 independent seeds within a single NSGA‐II run budget.
ABSTRACT For large‐aircraft surface inspection, nonuniform viewpoint distribution and wide task scope make Euclidean‐distance clustering difficult to jointly achieve path efficiency and workload balance. To address this, we propose a density‐adaptive clustering distance , allowing density information to influence cluster‐boundary adjustment throughout the iterative assignment process. Based on a C919 digital‐twin model, a nonuniform viewpoint set satisfying coverage and viewing‐angle constraints is generated; then, a Gaussian‐kernel local density on the viewpoint set (KDE‐style) and an S‐type mapping are used to construct an adaptive weight, and k‐means++ algorithm with is used to complete multi‐unmanned aerial vehicles (UAVs) task allocation under total‐distance minimisation, coefficient of variation (CV)‐controlled workload balance and single‐UAV energy constraints. For fair comparison, all methods share a unified ‘Nearest‐Neighbour + 2‐opt’ backend to evaluate each UAV closed‐loop length. Simulations on the C919 digital geometric model with 582 viewpoints and 3 UAVs show that DAF–CPP achieves the shortest mean total path (679.1 ± 2.6 m), the lowest mean maximum path (239.5 ± 13.4 m) and the lowest mean CV (0.1025 ± 0.0166) among eight allocation baselines under a unified 20‐seed protocol. Extended fairness metrics and Holm‐corrected tests further support the statistical significance of the allocation improvement.
ABSTRACT With the growing penetration of renewable energy, traditional power flow methods struggle to capture the multidimensional uncertainties and abrupt meteorological shifts in active distribution networks. To address this, the paper proposes a digital twin‐driven probabilistic power flow framework based on the semi‐invariant method and Gram–Charlier series expansion. In the modelling stage, real‐time data continuously calibrates Weibull, Beta and normal distribution parameters, transitioning grid evaluation to agile online tracking. A highly efficient convolution‐free engine utilises linear matrix mapping to directly derive nodal voltage semi‐invariants, whereas the Gram–Charlier expansion dynamically reconstructs continuous probability distributions. Simulation results on the IEEE 33‐bus system demonstrate that the proposed framework effectively quantifies the spatial accumulation of network uncertainties and the temporal evolution of operational risks, providing grid operators with dynamic probability boundaries of nodal voltage.
ABSTRACT Battery‐electric bus (BEB) dispatch reliability in hot climates is limited by day‐to‐day variability in energy demand and by uncertainty in auxiliary loads and charging access. We develop a telemetry‐anchored digital twin workflow that converts a calibrated MATLAB/Simulink longitudinal energy model into reliability‐based dispatch guidance using day‐aware splits (DayID) and day‐weighted evaluation. Using telematics from two BEBs operating in Belize (131 bus‐days; 1341 exported segments; 1187 QC‐passing segments and 674 segments used for fitting and scenario simulation), the calibrated model achieves median absolute day‐level noncharging energy errors of 11.3% (train), 12.4% (validation) and 13.5% (test), with median filtered‐power RMSE of 7.6–7.8 kW across splits. We define shortfall risk via p ‐quantile ‘safe distance’ thresholds computed from simulated range‐to‐reserve and estimate conservative envelopes using day‐bootstrap lower confidence bounds. Out‐of‐sample quantile‐transfer diagnostics show pooled p10 thresholds transfer close to nominal shortfall frequencies, whereas the long‐duty regime requires conservative treatment. Scenario sweeps over temperature, reserve SoC (0.10–0.25), HVAC efficiency and usable capacity indicate that battery‐only feasibility tightens sharply under heat stress; temperatures above the observed telemetry window (≈ 23.8°C–31.5°C) are treated as scenario‐planning stress tests (e.g., pooled reserve limit declines from 0.17 at 28°C to 0.11 at 32°C and becomes infeasible at 35°C within the tested reserve grid). Crediting observed opportunity charging (median 22–64 kWh/day depending on day type) expands feasible reserve envelopes and converts residual gaps into minutes‐scale charging requirements under typical charger powers; this ‘with‐charge’ case is an upper bound unless comparable charging access can be ensured operationally. The resulting dispatch charts and temperature‐binned reserve rulebook provide an operational interface from calibrated energy modelling to reliability‐based planning, with explicit flags where reserve policy alone is insufficient.
ABSTRACT The intractable nonlinear relationship between energy flow and carbon emission flow (CEF) challenges the low‐carbon scheduling of integrated energy systems (IES). To address this issue, this paper proposes a digital twin‐driven low‐carbon economic scheduling approach for electricity–heating–gas IES based on energy–carbon coupling constraints learning. Under optimisation with the constraint learning (OCL) framework, a neural network (NN) is trained in the digital twin environment to approximate the mapping from source‐side power injections to load‐side carbon emissions, employing a sparse training strategy to reduce NN model parameters. To improve computational efficiency, an improved ReLU linearisation method based on ideal mixed‐integer programming (MIP) formulation is developed, where strengthening inequalities are separated as needed to obtain tighter relaxations. The learnt CEF constraints are incorporated into the scheduling model, together with a stepped carbon pricing mechanism for low‐carbon demand response. Case studies show that the proposed approach reduces system‐wide carbon emissions, accurately captures energy–carbon coupling and improves the tractability of scheduling optimisation with embedded NN constraints.
ABSTRACT Predictive maintenance (PdM) plays a critical role in enhancing safety, operational efficiency and cost‐effectiveness in the aviation industry by enabling condition‐based maintenance strategies instead of traditional schedule‐driven approaches. This paper presents a systematic scoping review of the core technologies underpinning data‐driven PdM in aviation, with a particular focus on digital twin (DT) systems, engineering data management and artificial intelligence (AI) algorithms, and their integration across the PdM pipeline. The reviewed studies are systematically categorised according to three primary data types used in aviation PdM—time‐series sensor data, graphical data and natural language data—together with their associated methods for data collection, preprocessing, storage and analysis. In addition, the review analyses AI‐based approaches for remaining useful life estimation and fault detection, highlighting commonly adopted models and benchmark datasets such as C‐MAPSS. Key challenges identified in the literature include data heterogeneity, real‐time processing constraints, scalability and cybersecurity risks. Emerging solutions, including multimodel database architectures and fog–cloud hybrid computing frameworks, are discussed as enablers of robust and scalable PdM systems. By providing an integrated and aviation‐specific perspective, this review offers researchers and practitioners a structured foundation for the design, development and deployment of DT–enabled PdM systems in safety‐critical aviation environments.
ABSTRACT Digital twin‐enabled additive manufacturing (DT‐AM) represents a critical advancement in industrial intelligence, yet its current fragmented implementations necessitate a comprehensive systematic analysis. This review rigorously examines DT‐AM architectures, integration frameworks, and pathways towards achieving autonomous and scalable production environments. Employing a structured multidimensional taxonomy, DT‐AM systems are categorised based on functional scope (component, asset, system, process twins), integration depth (digital model, digital shadow, digital twin) and operational sophistication (ranging from descriptive to fully autonomous twins). The synthesis highlights a pronounced emphasis in existing literature on simulation and control functionalities, with notable gaps identified in validation frameworks and intelligence‐driven decision‐making mechanisms. Major challenges, including data heterogeneity, computational scalability and inadequate validation strategies, currently hinder seamless interoperability and broader industrial adoption. The analysis further identifies promising technological solutions such as agentic artificial intelligence, secure digital thread infrastructures, hybrid cloud‐edge computing and human‐centric augmented and virtual reality interfaces. Crucially, the paper underscores the imperative of software standardisation as foundational to the progression of DT‐AM systems. Five strategic research trajectories are proposed to systematically bridge existing technological and operational gaps, fostering resilient, interoperable and human‐centric cyber‐physical manufacturing ecosystems. Ultimately, this comprehensive review establishes a structured foundation for advancing DT‐AM, guiding future scholarly and industrial efforts towards cohesive, intelligent and scalable production systems.
ABSTRACT Gait analysis offers significant potential for personalised health assessment and enhances the understanding of musculoskeletal dynamics in biomechanical studies. To meet the growing need for intelligent gait analysis, a digital twin (DT)‐driven gait analysis framework integrating wearable inertial sensors, multi‐task deep learning, and OpenSim‐based musculoskeletal biomechanical simulation was investigated and achieved closed‐loop analysis by collecting, processing, and mapping gait data. The DT system architecture followed a layered design methodology, combining interaction between physical, data, DT and service layers to create a comprehensive framework. The inertial measurement unit (IMU) sensors captured the changes in limb angles, acceleration and angular velocity. Data were stored in data layer. A multi‐task deep learning model with three convolutional layers was designed to combine with the Zero Velocity Update (ZUPT) algorithm with a short‐distance step length calculation strategy to suppress integral drift, enabling accurate annotation of key gait events, such as heel strike, toe‐off and related parameters. Gait parameters were extracted through the closed‐loop adaptive model. A step length estimation accuracy of mean error ( μ ) −0.023 m with standard deviation ( σ ) 0.025 m and excellent temporal parameter regression performance were achieved. To reveal the biomechanical mechanism of abnormal gaits, the lower limb joints were simulated through model scaling, inverse kinematic analysis and inverse dynamic calculation based on OpenSim platform. The virtual mapping model in data‐driven digital layer synchronised with the real motion and iteratively optimised. Gait classification was achieved by DT based convolutional neural network (CNN) with an overall accuracy of nearly 98%. Experimental validation confirmed that the system can conduct analysis and visualisation of the posture and mechanical data of the lower limbs. The DT‐driven gait analysis system showed significant potential for providing performance optimisation guidance feedback to assist precision gait analysis.
ABSTRACT Conventional bioreactor models and many existing digital twin formulations commonly assume constant transport and thermal properties, which limits their ability to represent the gradual degradation that occurs during long‐term operation. In practical bioprocess systems, ageing mechanisms such as biofilm formation and fouling progressively modify oxygen and heat transfer behaviour, influencing overall system performance over time. To address this limitation, this study presents a physics‐inspired, ageing‐aware digital twin framework in which degradation mechanisms are incorporated directly as time‐evolving internal state variables within the governing transport formulation. Unlike conventional approaches that treat transport coefficients as fixed or periodically adjusted parameters, the proposed framework enables continuous mechanistic representation of degradation‐driven transport evolution. The model integrates mechanistic mass and energy balances with ageing‐dependent transport relationships to describe the coupled evolution of oxygen transfer and thermal behaviour during prolonged operation. Simulation results demonstrate that ageing‐induced transport degradation leads to gradual reductions in oxygen transfer capacity, productivity decline, and measurable thermal drift, even under nominal operating conditions. The framework further provides mechanistic insight into how biofilm accumulation influences long‐term transport behaviour and system dynamics in ways not captured by time‐invariant formulations. By explicitly coupling ageing evolution with transport degradation, the proposed digital twin establishes a physically grounded and interpretable framework for degradation‐aware bioreactor modelling. The present study is intended as a mechanistic foundation that can support future experimental calibration, parameter estimation and integration with monitoring systems for long‐term sustainable bioprocess operation.
ABSTRACT Addressing the issues of limited endurance and load capacity in unmanned aerial vehicle (UAV) inspection of distribution networks, the lack of dynamic adaptability in charging networks and insufficient multi‐objective collaborative optimisation, this paper proposes a four‐dimensional collaborative system architecture consisting of ‘digital twin‐UAV‐charging facility‐communication network’. It clarifies the essential differences between digital twins and traditional simulation/data‐driven platforms and constructs a full‐element, high‐fidelity, closed‐loop iterative digital twin operation mechanism. This architecture achieves precise replication and trend prediction of inspection scenarios, equipment status and grid operation through full‐element modelling of digital twins, real‐time synchronisation between virtual and real worlds, and dynamic deduction. On this basis, a multi‐objective optimisation model is constructed with the goals of minimising the total life cycle cost of the charging network, minimising the waiting time for UAV charging, and minimising the operational fluctuations in the distribution network. A dynamic weight coefficient is introduced to coordinate the objectives and achieve balance. An innovative approach is proposed to integrate digital twin deduction with an improved nondominated sorting genetic algorithm (DT‐NSGA‐II). This method enhances solution efficiency and convergence accuracy through heuristic population initialisation driven by twin data, adaptive genetic operations involving virtual–real interaction, and population optimisation strategies based on closed‐loop feedback. Additionally, the computational complexity and convergence properties of the algorithm are analysed. A digital twin simulation verification is conducted using a 120‐km 2 complex terrain distribution network in East China as a case study, and cross‐validation is performed with actual distribution network operation and maintenance data. The results show that the robustness and adaptability of the proposed method are improved by 40% (quantified based on multi‐objective optimisation standard evaluation metrics) compared to traditional methods in complex scenarios. This provides an engineering paradigm for the precise configuration and dynamic optimisation of UAV intelligent inspection and charging networks in distribution networks.
Natural gas scheduling on offshore platforms is essential for ensuring safe and economically efficient production. Traditional control methods, such as PID, often exhibit limited robustness and slow response under nonlinear dynamics, disturbances and fluctuating demand. This paper proposes a hierarchical scheduling optimisation framework that integrates an improved genetic algorithm (GA) with model predictive control (MPC), aiming to achieve long‐term global optimisation together with real‐time dynamic control. The improved GA incorporates adaptive crossover and mutation rates, simulated binary crossover (SBX), polynomial mutation and an elitism strategy to enhance search efficiency and avoid premature convergence. A simulation‐driven digital‐twin prototype is constructed using representative operating conditions—normal, high‐demand, low‐demand and fault—to evaluate flow and pressure regulation performance. The optimisation objective minimises total operating cost subject to physical feasibility, safety limits and equipment performance constraints, with decision variables including valve openings and compressor speeds. The GA generates hourly reference trajectories for global scheduling, whereas MPC ensures minute‐level tracking under disturbances and operational variability. Comparative results demonstrate that the hierarchical GA + MPC approach outperforms standalone MPC, reducing operating cost by 9.4%, accelerating convergence by 34.2% and lowering steady‐state tracking error by 18.6%. Relative to PID control, convergence speed improves by 43.5% and steady‐state error decreases by 35.1%. In addition, the improved GA achieves a 65.6% reduction in convergence generations compared with the traditional GA, confirming its superior efficiency and robustness. These results, validated within the simulation‐driven digital‐twin prototype, highlight the hierarchical architecture's ability to combine global optimisation with real‐time dynamic control, demonstrating its practicality, robustness and potential for broader application in intelligent offshore energy systems.
Digital twin technology, continuously updated, high‐fidelity virtual replicas of physical assets, has matured rapidly in aerospace and is now poised to transform construction in extreme environments on Earth and beyond. This Opinion Article synthesises recent research, global case studies, and policy developments to argue that digital twins are pivotal for planning, autonomous execution, and predictive infrastructure maintenance on the Moon, Mars, the deep ocean, polar regions, and remote deserts. By coupling real‐time sensor data with physics‐based and AI‐driven simulation, twins enable risk‐free prototyping, swarm‐robot coordination, and lifecycle resilience while reducing environmental footprints. We highlight strategic implications, data‐governance frameworks, workforce upskilling, and federated twin ecosystems and forecast emerging trends such as AI‐enhanced autonomy, XR interfaces, and smart‐material feedback loops. The article aims to inform multidisciplinary researchers, engineers, and policymakers by advocating the early adoption of digital twin standards to accelerate sustainable resilient construction in the most challenging frontiers.
Despite the increasing affordability of data processing and storage and the enhancement of artificial intelligence (AI) and digital technologies in recent years, scalability and adoption continue to be a challenge when it comes to digital twins (DTs). Common challenges that are often cited include the effort of designing and building DTs, high customisation, the cost to operate and maintain DTs, interoperability between DT components and DTs, and the extensive analysis and effort required to turn DT outputs into useful insights. AI has seen significant advancements and growth lately, driven by the release of popular AI products such as ChatGPT, Google Gemini and DeepSeek's R1. Many of the recent developments have the potential to address the challenges of scaling and adopting DTs. This paper examines the intersection of AI and DTs and explores how AI can be used to address some of the challenges of scaling and adopting DTs. It concludes with a set of principles that aim to apply to most DT applications, regardless of use case or industry, and proposes AI methods and techniques that can potentially be used for each principle. These principles are (1) reduce effort, cost and/or time; (2) optimise resource and system efficiency; (3) improve interaction and outcome and (4) improve interoperability, reusability and maintainability.
Leveraging associated gas for power generation is a critical pathway to enhance comprehensive resource utilization during the development of offshore oilfields. However, this process currently suffers from significant gas wastage and inefficient pipeline distribution. Meanwhile, compressor pressure and valve settings remain heavily reliant on manual experience, especially under the complex multi-platform architectures, which severely impedes the intelligent transition of this industry. To address these challenges, this paper proposes a digital twin-based predictive control strategy for multi-platform natural gas distribution. The strategy integrates four core modules: (1) a Digital Twin Module, which employs a hybrid mechanism-data modeling approach to construct high- fidelity digital twin models for simulating and monitoring gas distribution systems; (2) an Operating Condition Prediction Module, which establishes a benchmark library of operating conditions based on unit equipment models. Combining with power generation demands and actual platform gas flow, the library enables rapid, accurate prediction of stop valve openings, platform pressures and the selected control valve to be adjusted; (3) an Intelligent Distribution Module, integrating the steady-state system model with a PID parameter self-tuning algorithm to autonomously generate a control valve control scheme; (4) a Scheme Verification Module, which validates correctness of the control scheme with the dynamic-state model. A case study applying this control strategy to gas distribution in an offshore oilfield in China demonstrated that, according to the desired power station load, predictive control schemes are generated within one minute. The discrepancy between the gas flow of the dynamic system model and the desired one was less than 5%, verifying the engineering applicability of this strategy. Against the accelerating trend towards unmanned development of offshore oilfields, the proposed strategy provides a reliable solution for the intelligent allocation of natural gas resources from multiple platforms.
To enhance the techno-economic performance and robustness of multi-microgrids (MMG) systems, this paper proposes a two-stage bi-level collaborative optimisation strategy integrating energy sharing and price incentives. In the day-ahead stage, the shared energy storage operator (SESO) at the upper level employs conditional Wasserstein generative adversarial network (CWGAN) and conditional value-at-risk (CVaR) to quantify renewable uncertainty risks, formulating day-ahead transaction prices to maximise profit while deriving internal clearing prices based on supply-demand ratios. The digital twin operator (DTO) is further utilised to execute the internal energy-sharing clearing process, incorporating grid-constrained pre-dispatch to ensure the feasibility of the derived schedule. Simultaneously, the lower-level MMG system minimises operational costs by optimising internal energy-sharing and resource scheduling schemes. During the real-time stage, the SESO adjusts real-time prices based on a coupling mechanism with day-ahead shared energy volumes, whereas MGs execute rolling optimisation via model predictive control (MPC). The bi-level problem is solved using an improved particle swarm optimisation algorithm. Case studies demonstrate that compared to traditional P2P bidding models, the proposed method significantly reduces system operational costs and boosts SESO revenue. Specifically, compared to the traditional P2P bidding model, the proposed strategy reduces the aggregate dispatch cost of the MG cluster by 8.29% while simultaneously ensuring the operational profitability of the SESO. Furthermore, the standard deviation of grid exchange power decreases by 40.5%, indicating effective smoothing of power fluctuations and mitigating impact on the main grid.
Digital twins, DT, are being developed to emulate behaviour, control, performance and operation of electrical assets. For the purpose of simulation, testing, monitoring and maintenance of asset components, work shall be done to integrate in a DT ageing modelling, reliability, design procedures, diagnostics, health condition evaluation and maintenance of the weakest component of an electrical or electronic asset component, that is, electrical insulation. Life and reliability models, potentially able to account for any type of stress, even time varying, are developed and their integration in a DT is discussed. Similarly, design procedures, accounting also extrinsic ageing, that can be scaled, with voltage and power, and algorithms for the dynamic assessment of insulation health condition through diagnostic testing or monitoring, are developed. The takeaway is that a DT approach should cover any feature regarding electrical insulation systems from design to condition monitoring.
As Industry 4.0 advances, the application of digital technologies is increasingly being popularised in the logistics and warehousing sector. For digital twins, the accurate and efficient collection of real-time data from the physical world significantly impacts the accuracy of modelling. With the enhancement of computer processing power, the application of computer vision as an information acquisition module is becoming increasingly widespread. This paper proposes a novel method in which YOLOv8 object detection is applied to locate and track packing boxes on the production line. At the same time, the digital twin model is updated in real time to synchronise the positions of the boxes. In the system, high-definition industrial cameras are used to capture the production line, and the YOLOv8 keypoint estimation model is utilised to calculate the pose of each packing box in every frame. The results are then mapped to the coordinate system of the digital twin model through perspective transformation and coordinate conversion algorithms. Data exchange between the vision algorithm and the digital twin model is performed in real time via a Redis database. The feasibility of the system was verified in the diverging and converging areas of a small experimental production line. Compared with the traditional approach of using photoelectric sensors to detect the positions of packing boxes, the proposed system overcomes two major limitations: the inability to distinguish between different types of boxes and the failure to continuously track and locate them. The results demonstrate that the proposed method outperforms conventional techniques. It shows notable superiority in accurately simulating the motion posture and trajectory of packing boxes within a digital twin system. Such capabilities hold great significance for enhancing the monitoring and management of production logistics.
This paper presents a communication‐free frequency support strategy for point‐to‐point voltage source converter high‐voltage direct‐current (PtP‐VSC‐HVDC) links. The proposed strategy is conceived as a harmonic‐amplitude modulation (HAM), where the frequency deviation experienced by one power system is described as an instantaneous bi‐dimensional trajectory problem, which is analysed using Green's theorem. The trajectory analysis allows to express the level of frequency deviation experienced by a disturbed AC network as a geometrical problem. This is used to generate the necessary active power adjustment command for the PtP‐VSC‐HVDC link depending on the severity of the power imbalance in the affected AC network. The novel HAM strategy introduces a fixed‐frequency AC voltage waveform into the reference of the DC voltage to facilitate the active power adjustment through the PtP‐VSC‐HVDC. Numerical simulations show that the HAM strategy effectively helps to bind the maximum frequency deviations during extreme emergency conditions, such as a network splitting (AC network separation) events in synchronously coupled multi‐area power systems.
AI‐driven control, particularly, Reinforcement Learning (RL), is becoming popular in modern power systems with increasing number of DC components. However, training RL‐driven controllers on actual systems may cause interruptions and/or create safety concerns due to random actions that may occur during training. This study evaluates the effectiveness of digital twins as a platform for training such control schemes. We develop an RL‐driven controller for a small‐scale DC microgrid using its digital twin and test it on an actual testbed. The digital twin enables safe and uninterrupted training, demonstrating improved learning efficiency and reduced risk compared to on‐system training. We compare the RL controller's performance with a PI benchmark to assess its advantages and limitations. Results show both benefits and constraints of this approach, contributing to the broader application of AI and digital twins in power system control. This study also highlighting the potential of digital twins in facilitating safe training and testing for developing advanced control schemes.
