A predictable and reliable power supply is crucial for safe and efficient operation in the energy sector. Hence, digital twins with real-time monitoring, simulation, and optimization capabilities receive increasing attention. In particular, early detection of faults and other operational issues based on real-time processing of sensor data can greatly reduce downtime and maintenance costs. In this work, we compare the performance of conventional and state of-the-art unsupervised anomaly detection methods to detect tube leaks in the steam generator that recovers heat from a combined-cycle steam and gas turbine (CCGT). Since real CCGT operational data with known faults is not available, the comparison is performed using data from a high-fidelity dynamic model based on a CCGT installed on an offshore oil and gas platform. Specifically, we evaluate local outlier factor (LOF), one-class support vector machine (OC-SVM), principal component analysis (PCA), low-rank and sparse (LRS) decomposition, and a transformer autoencoder (TAE) using receiver operating characteristic (ROC)-area under the ROC curve (AUC), geometric mean (G-Mean), and true positive rate (TPR) at fixed false positive rate (FPR). For small-leak detection using the full sensor set, ROC-AUCs are PCA = 0.77, LOF = 0.87, OC-SVM = 0.95, LRS = 0.98, and TAE = 0.99, while all methods approach near-perfect AUC under large leaks. Under reduced data availability (e.g., no real-time makeup water flow data), LRS remains the only method that sustains high detection performance. In general, LRS delivers consistently strong accuracy and robustness across leak sizes and noise levels, offering a practical accuracy-complexity trade-off for deployment.

The strategic placement of bike-sharing infrastructure shapes urban accessibility and mobility outcomes. However, station-allocation approaches vary in their assumptions and decision logic. This study examines how alternative modelling paradigms prioritise urban space when applied to the same planning problem in Trondheim, Norway. We developed a unified analytical framework to compare three location-allocation approaches: weighted linear combination (WLC), maximal covering location problem (MCLP), and a data-driven suitability score based on exogenous spatial features (SSE). Each model designs a 68-station bike-sharing network from scratch using the same 24 spatial features and hierarchical weighting scheme. The resulting configurations are compared with the existing network, and consensus-based synthesis identifies 12 priority locations for expansion. The findings reveal systematic differences in spatial prioritisation across modelling approaches. WLC achieves the strongest coverage of population and transit demand, MCLP produces the widest spatial distribution prioritising geographic reach, and SSE balances demand intensity with accessibility. All model-derived configurations diverge from the existing network, highlighting the influence of historical and institutional factors on real-world deployment. Consensus synthesis identifies 12 expansion sites characterised by multimodal integration potential, underserved residential clusters, and high latent demand. This analysis demonstrates that methodological choices fundamentally shape spatial decision-support outcomes. By systematically evaluating classical optimisation and data-driven approaches under controlled conditions, the study provides evidence-based recommendations for bike-sharing network expansion and clarifies the strengths and limitations of alternative analytical frameworks for location-allocation planning.

Combined-cycle gas turbines (CCGTs) play a key role in modern power generation, offering both high efficiency and reduced environmental impact. However, their complex thermo-fluid and mechanical interactions complicate fault detection, particularly when labeled fault data are scarce. In this paper, we introduce the Kalman Prototypical Network (KPN), a metric-based few-shot learning (FSL) framework specifically tailored for CCGT fault diagnosis. We model the evolution of class prototypes as latent stochastic states in a dynamic system to reduce episodic variance and improve robustness in embedding representation. Synthetic data sets generated with a high-fidelity Modelica-based dynamic simulation of an offshore CCGT system were used, simulating both normal operation and progressive leak faults under transient conditions. Application of the proposed framework on simulated leak fault detection tasks demonstrate that KPN outperforms conventional FSL methods such as Matching Networks, Relation Networks, and MAML in both accuracy and stability under varying support and query configurations. The proposed framework significantly improves training convergence and generalization by stabilizing class representations, making it well-suited for real-world CCGT fault detection where labeled data is limited.

Accurate prediction of flow patterns in two-phase pipe flows is vital for the optimal design and flow assurance of production systems. Currently, mechanistic models are widely used to predict flow patterns in pipes at different operational, geometrical, and fluid properties conditions. However, if a two-phase flow system is highly complex in geometry and fluid properties, it is very challenging to develop a reliable physics-based mechanistic model. The predictions of such models often result in high uncertainty in flow patterns, especially in viscous two-phase liquid flows in vertical, horizontal, and deviated conduits. Alternatively, data-driven models, which learn and adapt to complex systems, are powerful tools to predict flow behavior and characteristics. This article aims to use machine learning (ML) to predict flow patterns in two-phase flow systems in horizontal, inclined, and vertical pipes across a wide range of liquid viscosities. An extensive database of two-phase flow patterns was compiled from the open literature, covering a wide range of inclination angles and liquid viscosities (2-2000 mPa.s). Several ML models are evaluated in this study: logistic regression, random forests, support vector machines, Knearest neighbors, and XGBoost. The ML predictions show that even the simplest machine-learning algorithm, i. e., logistic regression, gives more accurate predictions than the mechanistic model. Results revealed that trained machine learning models applied to test samples can predict the flow pattern with greater than 90 % accuracy, outperforming the mechanistic model, which achieves 60 % accuracy. The analysis of the relative importance of the input variables reveals that liquid and gas superficial velocities, and pipe inclination angle are the most important factors for correct classification.

Predictive maintenance (PdM) is recognized as a next-generation maintenance paradigm across industrial applications. However, implementing PdM in complex engineering systems, such as an offshore wind farm, remains highly challenging. These systems are exposed to continuous degradation and external shocks simultaneously and also feature recoverability. Meanwhile, they often exhibit failure interactions, such as common cause failures (CCFs) and dependent or competing failures. These characteristics hinder accurate condition assessment and complicate maintenance planning. To address these challenges, this study extends the stochastic hybrid system (SHS) framework to model the reliability of repairable systems with self-recovery capability under degradation and shock-induced CCFs. Furthermore, a PdM scheduling approach is proposed based on a Markov control policy that maximizes the expected asset value and determines optimal health levels under stochastic market conditions. A case study is conducted on a simulated offshore wind farm with 16 turbines exposed to lethal and non-lethal shocks. Partially observed data from simulations are used to estimate a transition matrix and predict system states for PdM decisions. The objective function minimizes the discrepancy between the estimated Markov control policy and turbine conditions through maintenance actions. The results show that the predicted optimal maintenance plan closely matches the theoretical optimal plan.

Neural operator learning has emerged as a powerful approach for approximating solution operators of partial differential equations (PDEs) in a data-driven manner. However, applying principal component analysis (PCA) to high-dimensional input and solution fields incurs significant computational overhead, particularly because the global singular value decomposition becomes expensive as the spatial resolution increases. To address this limitation, we propose a localized PCA-Net framework that decomposes the computational domain into smaller patch-wise coverings, applies PCA within each local patch, and trains a neural operator in the resulting reduced latent space. We investigate two patch-based formulations that balance computational efficiency, global coupling, and reconstruction accuracy: (1) local-to-global patch PCA, in which input fields are compressed locally while solution fields are represented globally, and (2) local-to-local patch PCA, in which both input and solution fields are compressed locally. To mitigate patch-interface artifacts in the local-to-local setting, we further study two refinement strategies: overlapping patch reconstruction with Hann-type weighted blending and a two-stage CNN-based RefinementNet. Experiments are conducted under a fixed-split, multi-seed protocol with relative error, structural similarity, interface-jump, PDE-residual, and runtime diagnostics. On the 2D Poisson benchmark, localized PCA reduces PCA fitting time by up to approximately 15 × relative to global PCA, while the fastest local-to-local configuration achieves a 1.7 × end-to-end speedup when PCA fitting, latent transformations, training, and inference are all included. The overlap-enhanced local-to-local model provides the best accuracy and physical consistency, reducing mean relative error from 7.52% for global PCA-Net to 2.31% while still achieving a 1.4 × end-to-end speedup. Additional studies on variable-coefficient Darcy flow, randomized SVD, patch-wise neural-network heads, and Gaussian-random-field roughness demonstrate the robustness, limitations, and extensibility of the proposed localized PCA-Net framework for scalable operator learning.

This study involves developing an unsupervised machine-learning (ML) framework (an artificial intelligence (AI) approach) for predicting micro-scale wind and turbulence. The localized wind and turbulence information from AI is a potential enabler for path planning of urban-scale drone operations and for use in smart-city through building-integrated renewable energy systems. In urban-city, the wind and turbulence is influenced by buildings and terrains and these complex local physics needs to be captured by the AI model. The trained ML in this work uses two input parameters: the meso-scale wind speed and the meso-scale wind direction in-order to predict the micro-scale building-induced wind and turbulence for regions around the Prague city centre. The performance of ML model has been compared with a traditional computationally-intensive computational Fluid Dynamics (CFD) solutions in terms of accuracy and computational speed-ups. After training, the ML performance has been compared for “unseen test datasets” in both the interpolation range and the extrapolation range of the input meso-scale parameters used in the training dataset. The results indicate that the machine learning model yields reasonably accurate solution for the velocity and turbulence fields for the test cases when the unseen input meso-scale parameters are “within the interpolation range” of training database and the ML infers the flow field in a matter of 5–10 s, i.e. a computational speed-up of over 1000 times as compared to the traditional CFD models. While for extrapolation (out-of-distribution) parameter test case, as expected, the ML model has scope of improvement in accuracy in flow-field. Overall, the ML-predictions due to its computationally efficiency can enable quick path planning and decision-making during drone operations, which is not possible with traditional computational fluid dynamics solution, thus highlighting the potential of machine learning for planning drone operations.

Large Language Models (LLMs) are increasingly used for decision-making and reasoning tasks, yet their potential as controllers for physical systems remains largely unexplored. This work investigates whether LLMs can function as interpretable controllers for a dynamic thermal environment, examining their ability to follow setpoints, interpret natural-language commands, reason about actuator effects, and incorporate prior model-based knowledge. Five LLMs of varying scales are evaluated under multiple scenarios, including settings with penalties on heater or fan usage and cases where the models have access to a physics-based prediction tool. The results show that control performance depends on model complexity: while low- and mid-scale models frequently misinterpret actuator dynamics or generate inconsistent reasoning, high-complexity models such as Qwen-3~14B and GPT-4o achieve accurate temperature tracking, stable actuator usage, and coherent explanations aligned with physical principles. Incorporating a physics-based model significantly improves control smoothness and energy efficiency by enabling anticipatory decision-making. A detailed reasoning taxonomy further reveals a clear progression from causal misinterpretation in smaller models to cohesive and temporally aware reasoning in larger ones. The findings demonstrate that LLMs can act as interpretable controllers when sufficiently capable and appropriately grounded in domain knowledge, highlighting promising opportunities for hybrid model-based and language-driven control strategies that can provide plausible explanations.

Urban bike-sharing systems require strategic station expansion to meet growing demand. Traditional allocation approaches rely on explicit demand modelling that may not capture the urban characteristics distinguishing successful stations. This study addresses the need to exploit patterns from existing stations to inform expansion decisions, particularly in data-constrained environments. We present a data-driven framework leveraging existing stations deemed desirable by operational metrics. A hybrid denoising autoencoder (HDAE) learns compressed latent representations from multi-source grid-level features (socio-demographic, built environment, and transport network), with a supervised classification head regularising the embedding space structure. Expansion candidates are selected via greedy allocation with spatial constraints based on latent-space similarity to existing stations. Evaluation on Trondheim's bike-sharing network demonstrates that HDAE embeddings yield more spatially coherent clusters and allocation patterns than raw features. Sensitivity analyses across similarity methods and distance metrics confirm robustness. A consensus-based procedure across multiple parametrisations distils 32 high-confidence extension zones where all parametrisations agree. The results demonstrate how representation learning captures complex patterns that raw features miss, enabling evidence-based expansion planning without explicit demand modelling. The consensus procedure strengthens recommendations by requiring agreement across parametrisations, while framework configurability allows planners to incorporate operational knowledge. The methodology generalises to any location-allocation problem where existing desirable instances inform the selection of new candidates.

Adaptive governance provides the institutional framework for iterative policy learning in mobility transitions, but often struggles to keep up with contemporary big data paradigms. This study develops a data-driven cybernetic framework as a high-frequency, feedback-oriented complement to adaptive governance. Using campus consolidation in Trondheim, Norway, we examine how it links policy intentions to commuting realities through integrated mobility analysis. We use a spatially-linked multi-source quantitative approach combining survey data (n=573) with public transit and crowd movement data. The framework integrates three analytical components to characterise commuting flows, modes, durations, distances, route concentration, and the spatiotemporal constraints shaping daily travel. The analysis reveals unevenly distributed mobility constraints, with 59.3% of respondents having children and private cars dominating in winter (49.4%). Although support for sustainable mobility goals is broad, 86.0% identify longer travel duration as the main difficulty. Baseline transit and crowd movement analyses highlight peak usage patterns and concentration risks as signals for recalibration. The study shows how multi-source quantitative analysis can establish pre-intervention baselines, identify system vulnerabilities under policy change, and support more responsive mobility governance. Interpreted through a cybernetic governance lens, the results inform context-sensitive interventions that align sustainability goals with lived realities.

Passenger count data from public transit systems reveals urban mobility patterns and is essential for planning, operation, and optimisation. However, non-linear spatiotemporal interdependencies across stops and lines make modelling and prediction challenging. Existing approaches often rely on fixed temporal, spatial, or stop-level formulations, limiting their ability to capture within-trip evolution and network context. This study proposes SMT-GraphFormer, a spatiotemporal multi-task graph transformer that frames trip-level transit prediction as sequence-to-sequence modelling. Given a line's stop sequence and trip-level context, the model predicts successive boarding and alighting counts, with delay and dwell time treated as encoder-side surrogate tasks. Key components include graph embeddings for multi-relational stop similarity, a context encoder for weather and temporal information, and a multi-gate mixture-of-experts module that produces task-specific decoder representations for boarding and alighting predictions. Evaluation on public bus transit data from Trondheim, Norway, shows that SMT-GraphFormer outperforms stop-level tabular benchmarks, with ablation studies examining each component's contribution. The sequential formulation yields substantial gains on alighting prediction (+0.24 in R^2) and consistent improvements on boarding, delay, and dwell, confirming the value of explicit trip-level sequential bias and inter-target dependencies. These findings demonstrate the potential of transformer-based sequence modelling for capturing complex spatiotemporal dynamics in public transit and underscore the value of architectures tailored to transit data rather than off-the-shelf tabular models. The proposed framework provides a horizon-agnostic basis for scenario analysis in digital twin environments, supporting informed decision-making by planners and transit operators.

Transformer-based models are very popular in time series forecasting because of their ability to capture contextual information in natural language processing tasks. However, their application to time series has yielded mixed results. Performance improvements of Transformers in this domain are usually based on metrics such as MSE or MAE and have marginal percentage gains. We include statistical tests to examine the significance of such performance differences. Moreover, comparisons rarely account for the uncertainty that training such complex models entails. We investigate this by looking at the robustness in terms of training stability of Transformer-based models compared to simpler long-term time series forecasting linear (LTSF-Linear) models. Instead of focusing on adversarial or anomaly robustness, we measure it through standard deviation to common perturbations in data-driven algorithms, including initialized parameters and data splits. By analyzing model performance and robustness, and statistical significance, across various benchmark datasets, we find that Transformer-based models perform generally worse in our experiments, as pointed out by other works, and exhibit significantly higher variability of results, hence lacking robustness. This study highlights the challenges of using Transformer architectures for time series forecasting and underscores the importance of considering model robustness and stability, and the significance of the results, when choosing algorithms for time series forecasting.

Anomaly detection is increasingly becoming crucial for maintaining the safety, reliability, and efficiency of industrial systems. Recently, with the advent of digital twins and data-driven decision-making, several statistical and machine-learning methods have been proposed. However, these methods face several challenges, such as dependence on only real sensor datasets, limited labeled data, high false alarm rates, and privacy concerns. To address these problems, we propose a suite of digital twin-integrated federated learning (DTFL) methods that enhance global model performance while preserving data privacy and communication efficiency. Specifically, we present five novel approaches: Digital Twin-Based Meta-Learning (DTML), Federated Parameter Fusion (FPF), Layer-wise Parameter Exchange (LPE), Cyclic Weight Adaptation (CWA), and Digital Twin Knowledge Distillation (DTKD). Each method introduces a unique mechanism to combine synthetic and real-world knowledge, balancing generalization with communication overhead. We conduct an extensive experiment using a publicly available cyber-physical anomaly detection dataset. For a target accuracy of 80

Anomaly detection is crucial in maintaining the safety, reliability, and optimal performance of complex systems across diverse domains, such as industrial manufacturing, cybersecurity, and autonomous systems. While conventional methods typically handle single data modalities, recently, there has been an increase in the application of multimodal detection in dynamic real-world environments. This paper presents a comprehensive review of recent research at the intersection of agentic artificial intelligence and large language-based multimodal anomaly detection. We systematically analyze and categorize existing studies based on the agent architecture, reasoning capabilities, tool integration, and modality scope. The main contribution of this work is a novel taxonomy that unifies agentic and multimodal anomaly detection methods, alongside benchmark datasets, evaluation methods, key challenges, and mitigation strategies. Furthermore, we identify major open issues, including data alignment, scalability, reliability, explainability, and evaluation standardization. Finally, we outline future research directions, with a particular emphasis on trustworthy autonomous agents, efficient multimodal fusion, human-in-the-loop systems, and real-world deployment in safety-critical applications.

Digital twins are evolving into self-learning, autonomous systems that link models, data and human interaction. Realizing their full potential depends on interoperability, standardization and the integration of artificial intelligence and advanced computational reasoning across sectors.

Effective control of indoor environments is crucial for maintaining occupant comfort and optimizing energy use. However, current building control strategies often fail to achieve these goals, as they rely on static or rule-based approaches that normally do not account for dynamic conditions. While advanced control strategies offer a more adaptive solution, their implementation is challenging due to the need for accurate thermal models, which are resource-intensive to develop. Defining building thermal zones can help to strike a balance between model accuracy and the cost of their development and implementation. However, data-driven approaches for identifying thermal zones remain scarce. This study addresses these gaps by proposing a reusable data-driven thermal zoning algorithm that employs Principal Component Analysis (PCA) and k-means clustering to define building thermal zones. This method allows for the inclusion of numerous parameters, thus increasing the applicability and consistency of the zoning process. Additionally, we propose an algorithm for zones validation, supported by qualitative criteria from literature and standards. The approach is tested in a large educational building, using time-series data from 168 rooms with a total of 262 CO2 and temperature sensors. Results show that the proposed zoning algorithm achieves over 91 % consistency score, depending on the number of selected principal components, clusters, and input parameters available. The derived thermal zones are further validated based on the synthesised qualitative criteria. Finally, the results are visualized in a DT environment, where users can explore color-coded thermal zones alongside real-time sensor data, 3D building geometry, and semantic information.

Anomaly detection in multivariate time series (MTS) from sensor data is critical in many industrial applications. The challenge lies in managing massive unlabeled datasets with complex spatio-temporal correlations, diverse anomalies, and noise. While several unsupervised methods have been proposed, they are often limited to specific applications. In this paper, we introduce a probabilistic self-supervised framework, Autoregressive Density Estimation Transformer (ADET). ADET integrates an efficient transformer for learning spatio-temporal representations with density estimation networks for multi-score anomaly detection, focusing on point-to-point, point-to-distribution, and distribution-to-distribution distances. ADET improves noise resilience using optimal truncated singular value decomposition (OT-SVD) in an end-to-end optimization process. We conducted experiments by employing several encoders and performed an ablation study to examine the effect of OT-SVD.

Offshore wind turbines face substantial challenges in operation and maintenance due to the harsh marine environment and remote locations. Predictive maintenance, encompassing fault diagnostics and failure prognostics, is a promising maintenance strategy to address these challenges. To contribute to this strategy, an integrated data-driven model is developed for probabilistic failure prognostics at the component level. The remaining useful life of a gearbox pump in an offshore wind turbine is predicted accurately based on supervisory control and data acquisition data. In this approach, light gradient boosting machines are tuned to model normal temperatures. The gated recurrent unit outperforms other neural networks and is selected to process temperature residuals with a Bayesian neural network. Results show that the prediction at the 50% percentile precedes the true failure time by 3.83 h. Moreover, there is 97.5% confidence that the true failure time falls within around +/- 5.3 h of the predicted time. Furthermore, the earliest alarm is issued at the 2.5% percentile, precisely 9.17 h prior to the true failure time. This study demonstrates the effectiveness of supervised learning and normal behavior modeling for probabilistic failure prognostics of offshore wind turbine components.

Anomaly detection is increasingly becoming crucial for maintaining the safety, reliability, and efficiency of industrial systems. Recently, with the advent of data-driven decisionmaking, several statistical and machine-learning methods have been proposed. However, these methods face several challenges, such as dependence on only real sensor datasets, limited labeled data, high false alarm rates, and privacy concerns due to centralized data processing. To address these issues, we propose a digital twin-based anomaly detection approach using federated transfer learning. Digital twins allow effective anomaly detection without requiring extensive real-world failure data. The integration of digital twins with federated learning offers a powerful solution to the challenges of anomaly detection in industrial systems. The proposed method integrates digital twin data for initial training with real physical system data for model refinement. Federated learning enhances this process by maintaining data privacy through the sharing of model updates instead of raw data. The proposed combination improves model generalization, training efficiency, and performance while ensuring data privacy. We perform an extensive analysis using publicly available datasets from real-world digital and physical asset counterparts. The results demonstrate significant improvements in anomaly detection performance, highlighting the effectiveness of integrating digital twins with federated learning.

Accurate prediction and real-time monitoring of downhole conditions are essential for the efficient and safe execution of drilling operations. Key parameters influencing these predictions include drilling fluid properties, different wellbore geometry, existing cuttings in the annulus, eccentricity of drill string and pipe rotation. Traditional hydraulics models typically utilize static equations, which may not adapt well to the varying conditions encountered during drilling. This can lead to inaccurate predictions, increased non-productive time (NPT), and potential safety risks. This paper presents an online calibration approach that dynamically updates the hydraulics model using sensor readings, ensuring continuous, real-time optimization. By doing so, the approach improves the accuracy of downhole predictions and enhances decision-making capabilities in drilling operations. The novelty of this research lies in its integration of physics-based governing equations with data-driven techniques for real-time calibration, providing an adaptive and responsive system. The proposed calibration method integrates drilling hydraulics model with model reference adaptive control (MRAC) architecture to monitor variations in drilling parameters. The calibration process is adaptive, meaning it can continuously respond to changing conditions, such as variations in flow rates, cuttings concentration, and wellbore geometry, that static models might fail to capture. This approach involves utilizing a hydraulic model, which is recursively updated using real-time measurements and a hybrid combination of physics-based and data-driven modeling techniques. The proposed modeling framework is simulated using drilling data from a case study. The results indicate that the online calibration method delivers accurate frictional pressure loss and bottom hole pressure predictions. By dynamically updating the coefficients in the hydraulic model, the system can account for real-time changes, reducing discrepancies between predicted and actual conditions. This leads to better management of drilling parameters, minimizes NPT, and enhances safety by providing more reliable data for decision-making. The approach demonstrates robustness and potential applicability across various drilling environments, including high-pressure, high-temperature (HPHT) conditions and complex well geometries.