In recent years, meshless methods have been increasingly applied to the simulation of various engineering problems due to their inherent advantages over traditional mesh-based approaches, including greater flexibility, independence from predefined meshing, simpler adaptive analysis, improved automation, and suitability for complex problems. Several meshless methods have been used for porous media simulation, and are broadly categorized into collocation, global weak form and local weak form methods. In this study, a comprehensive comparison of the applicability of these three categories of meshless methods for simulating coupled flow and transport problems in porous media is presented. The Radial Point Collocation Method (RPCM) (strong form), the Element Free Galerkin Method (EFGM) (global weak form) and the Meshless Local Petrov Galerkin (MLPG) method (local weak form) are implemented and systematically compared. These methods are applied to the analysis of flow in a synthetic regular domain aquifer, flow and non-reactive contaminant transport in a synthetic irregular boundary porous media problem and groundwater flow in a field aquifer located in India. The simulated groundwater heads are compared with analytical solution, observed field data and results obtained from widely used MODFLOW-MT3DMS models. The deviation of the solutions from the analytical solution is in the range of 0.67% to 0.16% for the hypothetical case study. For the field-scale case study, mean absolute error of 0.183%, 0.181% and 0.188% are obtained for the RPCM, EFGM and MLPG models, respectively, outperforming MODFLOW, which exhibits a deviation of 0.254% from observed values. Overall, the present study reaffirms the practical applicability of these meshless methods for real-world groundwater problems and provides valuable insights into the utilization of each category of meshless method, with respect to problem type, computational efficiency and accuracy requirements.
The construction industry's substantial carbon footprint, primarily attributed to the production of Ordinary Portland Cement, necessitates a transition toward more sustainable alternatives. Geopolymer concrete (GPC), an innovative binder synthesized from industrial by-products like fly ash (FA), offers a promising low-carbon solution but is hindered by performance variability and a lack of standardized design protocols. This research addresses this critical barrier by developing robust predictive models for the compressive strength of FA-based GPC. Six machine learning algorithms, including Bagging, Categorical Boosting (CatBoost), K-Nearest Neighbors (KNN), LightGBM, Random Forest Regressor (RFR), and eXtreme Gradient Boosting (XGBoost), were developed and evaluated. The results demonstrate that the XGBoost model achieved superior predictive accuracy, with the lowest average errors (Mean squared error of 4.66, Mean absolute error of 1.42) and the highest average coefficient of determination (0.962). To enhance model interpretability, a Morris sensitivity analysis was conducted. The analysis quantitatively identified the coarse aggregate as the most influential parameter governing compressive strength, followed by key chemical precursors such as silicon dioxide (SiO2) and aluminum oxide (Al2O3). These findings not only align with established material science principles but also validate the physical realism of the machine learning model. This study provides a reliable computational framework for predicting the performance of FA-based GPC, facilitating mix design optimization and accelerating the adoption of this sustainable material in modern construction.
The mechanical behavior of nonwoven fabrics as reinforcement in cementitious composites remains insufficiently explored, particularly from a numerical modeling perspective, despite their growing interest as sustainable alternatives to conventional textiles. This study presents a simplified, engineering-oriented numerical modeling framework for reproducing the flexural mechanical response of cementitious composites reinforced with flax nonwoven fabric. Four-point bending (flexural) behavior of nonwoven fabric-reinforced cementitious composites was numerically simulated using ANSYS software. The model is developed using Finite Element Analysis (FEA) and incorporates a Representative Volume Element (RVE) approach to account for the heterogeneous fiber-matrix interaction. The required material properties were iteratively calibrated using existing experimental data for three composite configurations comprising 4, 5, and 6 layers. The proposed model demonstrated agreement with experimental results within the investigated configurations, achieving normalized root mean square errors (nRMSE) of 5.2%, 4.6%, and 2.07% for the respective configurations. Furthermore, correlations between material parameters and geometric factors were identified, providing preliminary insights for estimating model input properties from easily measurable variables. Finally, sensitivity analyses were performed to evaluate the influence of key geometric and material parameters on the structural response, offering valuable insights for the optimized design of nonwoven fabric-reinforced cementitious composites.
The introduction of the Open Radio Access Network (O-RAN) architecture enhances network flexibility but introduces novel security threats targeting open interfaces and the RAN Intelligent Controller (RIC). Particularly in the Near-RT RIC environment, an effective Intrusion Detection System (IDS) that satisfies strict near-real-time constraints of within 1 s is essential to defend against cyber attacks. This paper proposes an Artificial Intelligence (AI)-based IDS xApp designed for real-time cyber attack monitoring in the O-RAN Near-RT RIC environment, and quantitatively analyzes its anomaly detection performance and inference latency characteristics against multi-layer security threats utilizing Open RAN Centralized Unit(O-CU) network layer data and Open RAN Distributed Unit (O-DU) radio telemetry data. Evaluation using a public dataset (NetsLab 5G O-RAN IDD) on four deep learning models (LSTM, CNN, Transformer, Autoencoder) showed that supervised learning-based models achieved high F1-scores (reaching up to 0.99) on both datasets. Furthermore, their performance variation remained highly stable at approximately the +/- 0.1 pp level upon transition from the training environment (the Service and Management Orchestration, SMO) to the deployment environment (Near-RT RIC). In the inference latency analysis, the system's scalability was evaluated by increasing the number of prediction instances up to 80,000. The results confirmed that the latency follows a highly predictable linear time complexity (O(N)). Specifically, the LSTM, CNN, and Autoencoder models successfully maintained a response time within 1000 ms even under the maximum load of 80,000 instances across both datasets, whereas the computationally heavy Transformer model experienced resource exhaustion in the KServe inference pod at approximately 20,000 instances, causing the inference process to terminate and rendering further measurement infeasible. Comprehensively, the LSTM model demonstrated the most outstanding balance between performance and operational efficiency by recording stable detection performance, short tail latency (approximately 140 ms at P99), and low training resource consumption. This study experimentally demonstrates the anomaly detection performance of the IDS xApp in the O-RAN near-real-time control environment, and comprehensively verifies its practical effectiveness by considering both inference latency and resource consumption.
Solving differential equations (DEs), including ordinary differential equations (ODEs) and partial differential equations (PDEs), is fundamental to scientific computing and engineering. The development of deep learning has led to Physics-Informed Neural Networks (PINNs), in which physical laws are embedded directly into the loss function. However, PINNs inherit the intrinsic instability of deep neural networks (DNNs) and lack an effective mechanism for Uncertainty Quantification (UQ). This paper proposes a stochastic ensemble framework to address these limitations. The proposed method is a double-stochastic ensemble framework that combines bagging (via bootstrap resampling and randomized collocation points) with Monte Carlo dropout (MCDO) applied at inference time to provide consistent UQ. This double-stochastic design yields two main contributions. First, it reduces the intrinsic variance of the predicted solution and improves accuracy, particularly in small-data and/or noisy-data regimes. Second, the variance of the ensemble outputs serves as a reliable and computationally feasible proxy for the solution uncertainty. Numerical experiments on ODEs and nonlinear PDEs (e.g., Burgers' equation) under impulse noise demonstrate that the proposed method outperforms standard baselines in predictive accuracy. Furthermore, the obtained Prediction Interval Coverage Probability (PICP) and Mean Prediction Interval Width (MPIW) confirm that our framework yields well-calibrated prediction intervals and effectively mitigates the severe performance degradation observed in single PINNs.
Computer vision has been widely adopted in intelligent construction monitoring; however, existing studies primarily focus on identifying individual construction elements or isolated activities, with limited capability for integrated monitoring of complete construction workflows. Such workflow-level automation is a prerequisite for intelligent construction and unmanned job sites. To address the challenge of reliable visual recognition in drill-and-blast tunnel environments characterized by uneven illumination, localized glare, and dust interference, this study proposes a methodological framework for construction workflow recognition at the tunnel face using computer vision and context reasoning. The framework consists of three components: (1) a construction workflow model with a sequence library database, (2) a robust construction element recognition model combining an enhanced YOLOv11 with Segment Anything Model 2 (SAM2), and (3) a hierarchical workflow reasoning mechanism driven by domain knowledge. A hierarchical workflow model embedding procedural logic is established through field investigation and normative analysis. SAM2 is employed for automated dataset annotation, while YOLOv11 is structurally enhanced with Convolutional Block Attention Module (CBAM), Adaptive Feature Enhancement (AFE), and Swin Transformer modules to improve feature representation and adaptability to degraded visual conditions. Workflow identification is finally achieved by integrating visual perception outputs with hierarchical context reasoning. Validation in an active drill-and-blast tunnel shows that the proposed method attains an average detection precision of 91.1% across 11 construction element categories, exceeding 95% for large equipment, and an average workflow recognition accuracy of 94%. The results demonstrate the effectiveness of the proposed framework for monitoring the tunnel construction workflow and supporting construction management.
Malware has evolved from the early Creeper virus into highly sophisticated and organized cyber threats. Over time, it grew in sophistication, adopting advanced techniques, stealth tactics, and autonomous propagation. Modern malware leverages encryption, obfuscation, zero-day exploits, and AI-assisted techniques to conduct stealthy and persistent attacks. Classification of its exact family is the end goal to defend and mitigate the latest attacks. Researchers have contributed significantly and introduced many techniques to tackle malware threats. Binary detection is performed at a large scale, but very little in multi-class classification. In this research, a hybrid technique is proposed by combining a sandbox with AI models to extract hidden patterns and classify its category and family with high accuracy. A dataset (AU-PEMAL-2025) is prepared, which includes 10,839 records of 26 malware families. Five ML and three DL models are trained on the newly created dataset to validate its effectiveness. The ML classifiers achieved the highest accuracies of 0.9945, 0.9788, and 0.9485, while the DL models achieved 0.9932, 0.9591, and 0.9286 accuracies with minimal losses in detection and multi-class classification of category and family, respectively. Our findings reveal that the proposed approach can efficiently detect the obfuscated malware variants and safeguard organizations from unseen malware threats.
The development and testing of autonomous agricultural robots requires realistic simulation environments that accurately represent field conditions and terrain features. Traditional manual scenario creation is time-consuming, expensive and limits the diversity of testing conditions. This paper presents an integrated two-stage system for semi-automated generation of realistic 3D simulation scenarios. The first stage transforms publicly available geospatial data into high-fidelity 3D terrain models, supporting 23 discrete levels of detail (LoD), from 0 to 22, and generating simulation-ready models compatible with the Gazebo robotics simulator. The second stage provides a web-based tool that enables users to populate generated terrains with crop elements, configuring crop distributions, row patterns, and field geometries through a map interface. Detailed performance evaluation across multiple LoD levels identifies the optimal balance between visual fidelity and computational efficiency, with levels 19-20 providing sufficient geometric detail for accurate sensor simulation while maintaining real-time performance on standard hardware. The complete system integrates with the Robot Operating System (ROS) 2 and Gazebo, significantly reducing scenario creation time by eliminating the need for manual modelling of terrain and agricultural elements. Validation experiments were conducted using a Summit-XL robotic platform equipped with an additional RealSense depth camera. The results demonstrate the system's capability for developing and testing autonomous navigation algorithms in the generated scenarios.
Cerebral palsy is a prevalent neurodevelopmental syndrome that disrupts motor development in children, making early detection vital for effective intervention. Traditional clinical assessments rely on subjective observations, often missing minor motor abnormalities until they become severe, typically after 12 months of age. This article presents a novel deep learning model, TransCP-Net (Transformer-based Cerebral Palsy Network), designed for early detection of infant cerebral palsy through spatiotemporal pose representation learning. The architecture employs hierarchical spatial and temporal attention to analyze complex motion patterns in video sequences, integrating multi-modal data for improved accuracy. TransCP-Net incorporates specialized preprocessing, including temporal smoothing and trajectory encoding, to enhance feature learning. Tests on 1370 infant movement videos yielded impressive results: 94.7% sensitivity, 92.3% specificity, and an AUC-ROC of 0.968, outperforming ten state-of-the-art methods. Notably, it achieved a sensitivity of 96.3% within the critical 9-15 weeks range of fidgety movements, enabling timely interventions. Attention visualization highlights key areas such as the hips and shoulders, reinforcing clinical relevance. TransCP-Net demonstrates effectiveness across diverse clinical settings, serving as a viable, non-invasive tool for early cerebral palsy detection.
Tunnel environments often suffer from GPS denial, uneven illumination, and structural uniformity, which lead to feature degradation, loop closure failure, and long-distance drift in SLAM systems. To solve these problems, this study aims to propose a high-precision SLAM method suitable for tunnel structural health monitoring. Firstly, an ABA-CLAHE image enhancement algorithm is proposed, which adopts cascaded processing of nonlinear brightness adjustment in HSV space and CLAHE local contrast optimization to improve low-light image quality and enhance feature stability. Then, SURF feature matching combined with the RANSAC algorithm is used to ensure feature matching accuracy. Finally, a factor graph model is constructed by integrating IMU pre-integration, laser odometry, visual odometry, and loop closure constraints, and iSAM2 incremental optimization is employed to achieve globally consistent mapping. Municipal tunnel tests show that the loop closure error is reduced to 0.096 m and the global reprojection error is 1.10 pixels, and the structural continuity of the constructed dense 3D map is significantly improved. This method provides a technical solution with centimeter-level accuracy for tunnel structural health monitoring, which is demonstrating strong practical potential for engineering applications.
The current study comprehensively investigates Williamson nanofluid flow and transport in an asymmetric porous tapered channel under varying slip conditions, using both analytical and supervised machine learning approaches. This mathematical model integrates thermophoresis, Brownian motion, the Soret and Dufour effects, thermal radiation, and a transverse magnetic field to accurately describe thermosoluble transport phenomena relevant to biomedical contexts. The non-Newtonian Williamson formulation is used to explain how fluids, such as blood, dilute when sheared. Darcy resistance is used to describe porous structures in tissue scaffolds, capillary networks, and dialysis membranes. A perturbation method is used to find analytical solutions that show how key dimensionless parameters affect the profiles of velocity, temperature, concentration, Nusselt number, Sherwood number, skin friction, and pressure gradient. Supervised machine learning models, including artificial neural networks, are also used to predict heat and mass transfer properties and confirm analytical trends with a high degree of accuracy. The results show that increasing the Hartmann number reduces fluid motion due to Lorentz force resistance by approximately 14%, while the Williamson parameter increases shear-thinning and increases velocity by approximately 9%. Thermal radiation significantly broadens the temperature distribution, increasing heat transfer by 12%. The combination of perturbation analysis and supervised machine learning models demonstrates strong predictive power and makes the results more reliable. The integrated analytical-machine learning framework provides essential insights for enhancing nanoparticlemediated drug delivery and advancing hyperthermia cancer treatment through regulated thermosolute transport in porous biological tissues.
Metasurface design often requires solving field distributions across varying structural parameters and frequencies, where neural operators offer a promising avenue for fast prediction. However, conventional neural operators have problems with degradation of the accuracy in multi-scale structural analysis. In this work, we propose a Generative Residual Enhanced Neural Operator (GRE-NO) framework that introduces a generative residual network to model the systematic bias of the main predictor. The core model retains the DeepONet architecture with both branch and trunk networks implemented using Fourier Neural Operators, combining strong generalization and efficient global representation. To handle the complexity of unbounded acoustic scattering problems, we integrate the Boundary Element Method (BEM) into data modeling and field computation, which reduces the problem dimensionality and enables training with samples at the 104 scale. Numerical experiments on some 2D and 3D acoustic metasurface problems demonstrate that the developed GRE-NO achieves excellent accuracy in results with relative errors under 1% in this study, outperforming conventional neural networks in accuracy of prediction.
As large language models (LLMs) become increasingly integrated into enterprise decision-making processes, structural pressures such as version drift, cross-source evidence integration, and regulatory accountability have shifted the primary challenge from isolated generative performance to system-level consistency, traceability, and governability. This paper systematically reviews key technological developments relevant to enterprise requirements, attribution evaluation, and multi-agent coordination. The analysis demonstrates that the main obstacle to enterprise LLM adoption is not model capability, but rather the structural gap between fragmented technical modules and the need for high-reliability decision-making. In response, a risk-controlled data flywheel architecture is proposed that integrates perception, reasoning, verification, and governance layers. By converting reasoning outputs into observable risk signals and feeding them back into retrieval and structural components, this architecture establishes a continuous improvement loop. This approach provides a systematic deployment blueprint for enterprise-grade LLM systems, emphasizing traceability, accountability, and sustainable optimization in high-risk and long-term operational contexts.
Classical image denoising methods remain relevant in practical scenarios where training data or noise models are unavailable, yet their performance is highly sensitive to parameter selection. Non-Local Means (NLM) is a representative example whose effectiveness depends critically on smoothing strength, patch size, and search window configuration. This paper formulates NLM parameter selection as a black-box optimization problem under unknown noise conditions and employs adaptive metaheuristic optimization strategies for this task. We propose an adaptive optimization framework that integrates rank-based perturbation, opposition-based learning, L & eacute;vy-flight exploration, and noise-aware parameter constraints to improve robustness and convergence. The proposed method is evaluated against fixed-parameter NLM and NLM optimized using standard evolutionary algorithms under identical protocols. Experiments on three sets of datatset demonstrate consistent improvements in PSNR and SSIM, highlighting the continued relevance of adaptive optimization for classical denoising.
Hybrid CNN-Transformer models are widely used in medical image segmentation because they combine CNN-based local feature extraction with Transformer-based global context modeling. Despite their popularity, these models face several challenges, including computational complexity, noise blurring, and information loss. This paper proposes an enhanced convolutional attention network (ECANet) for liver segmentation. ECANet uses a U-shaped architecture with efficient channel-attention-based skip connections. Both the encoder and decoder are constructed using enhanced convolutional Transformer (ECT) blocks, where group convolution is integrated into the convolutional attention module for efficient Token embedding and channel disentanglement, and a Token-wise multi-layer perceptron (MLP) branch is incorporated into the wide-focus module to improve feature representation across channels. Deep supervision and a hybrid of Binary Cross-Entropy (BCE) and Dice loss are used to improve boundary accuracy. We evaluate the proposed model on the publicly available LiTS17 dataset. Experiments show that ECANet outperforms the compared CNN-based and CNN-Transformer baseline models on both quantitative and qualitative measures.
Background: Examining peripheral blood smears is a vital diagnostic tool in hematology. Although deep learning-based automated systems have improved the accuracy of blood cell detection, most current methods depend on fully supervised learning and need extensive annotations to identify unusual cell shapes. Clinical practice often lacks these complete annotations, limiting the ability to generalize to rare or unseen abnormalities. To address this issue of incomplete annotated data, this paper introduces a hybrid framework that recognizes anomalies for reliable and clear analysis of peripheral blood smears. Methods: The proposed framework combines supervised blood cell detection with unsupervised one-class anomaly detection. We use a YOLOv8 detector to accurately identify and categorize red blood cells, white blood cells, and platelets in peripheral blood smear images. The detected blood cells are then assessed further using a lightweight Cell Reconstruction Autoencoder, which is trained only on normal-looking cells. To enhance sensitivity to minor variations in the model's construction, we also introduce a Wasserstein-based method for latent distribution regularization. This method scores anomalies based on how well the reconstruction matches and the deviation in the latent space, without needing any abnormal training samples. Results: We evaluate detection performance on the TXL-PBC dataset and validate it externally with pathological samples from the ASH Image Bank. The results show high accuracy in detecting abnormalities and in distinguishing normal from abnormal cells (F1Score/Anomaly = 0.926; AUC = 0.963). Our experiments and statistics confirm that using the hybrid approach with Wasserstein regularization significantly improves sensitivity and robustness compared to individual or reconstruction accurately identifies cells and detects previously unseen morphological abnormalities within a one-class learning framework. This approach combines efficiency with anomaly awareness and clear clinical interpretation, making it suitable for practical and scalable analysis of peripheral blood smears in clinical settings.
A global health concern, neurodegenerative disorders like Parkinson's and Alzheimer's impact both mental and physical functioning. The complex interplay among immunological response, protein accumulation, and brain health necessitates sophisticated mathematical modeling. This study introduces a fractional-order mathematical model using the Mittag-Leffler derivative to describe the dynamics of neurodegeneration, incorporating key biological factors such as functioning and infected neurons, extracellular alpha-synuclein, microglia, and T-cells. A fundamental assumption of the model is that neuronal deterioration is influenced by memory effects, where past states impact current disease progression, making fractional-order calculus more suitable than traditional integer-order models. The model accounts for the secretion and clearance of alpha-synuclein, the activation of immune responses, and the role of microglia in mitigating or exacerbating neuronal damage. Sensitivity analysis emphasizes the crucial role of factors like neuronal cells production Pi(N), infection prevalence gamma, and stimulation of microglial cells Theta. Numerical simulations support the long-run neuroinflammatory feedback mechanism, revealing that smaller values of fractional order eta < 1 reduce disease progression. This is based on the premise that increased memory (eta values less than one) leads to slower transmission of pathological protein aggregation. The study demonstrates that building a surrogate machine learning model of the NARX-BRBNN type, calibrated using numerical solver output, not only decreases computing complexity but also accurately replicates the dynamics of the fractional equation. This comparison underscores the necessity of employing fractional-order numerical schemes for accurately modeling complex neurobiological systems. The study proposes focused treatment approaches and provides insightful information on the course of neurodegenerative diseases.
Maintaining overhead distribution facilities inherently involves high risks for operators, where ensuring worker safety and operational efficiency remains a paramount challenge. In particular, automating the positioning of aerial work platforms is crucial to mitigate electrocution hazards during live-line maintenance tasks. This paper proposes a novel autonomous framework for detecting jumper lines that could be employed to estimate the optimal bucket position in live-line maintenance of overhead distribution systems. The proposed framework comprises three core modules to form a unified pipeline for autonomous field inspection: a 4D multi-modal map, Sparse-dense fusion network (SDFNet), and Rotational multi-pyramid Transformer with texture and augmentation (RoMP-Tax). The 4D multi-modal map aims to establish an accurate spatial-temporal representation of the maintenance area by integrating light detection and ranging (LiDAR), camera, inertial measurement unit (IMU), and global navigation satellite system (GNSS) measurements. The SDFNet detects telegraph poles from the 4D multi-modal map through geometry, pseudo, and fusion streams, which effectively extract both geometric and optical features. The RoMP-Tax, designed with a hybrid CNN-Transformer architecture enhanced by LBP-based texture encoding and Mixup augmentation, identifies insulators under complex textures and varying illumination. Extensive evaluations on field measurements and benchmark datasets demonstrate the high accuracy and consistent performance of the proposed framework with respect to multiple quantitative metrics, validating its robustness and generalizability. The proposed framework, deploying core technologies of the fourth industrial revolution, provides a reliable and efficient solution for estimating optimal bucket positioning, thereby contributing to the establishment of safe, data-driven live-line maintenance of distribution facilities.
Hybrid fiber reinforced plastic (HFRP) composites, especially intra-layer carbon/glass hybrids, offer a promising balance of specific strength, impact resistance, and cost efficiency for thin-walled energy-absorbing structures. This study investigates the low-velocity impact response and energy absorption of intra-layer carbon/glass hybrid hat-shaped beams. Tensile and impact tests evaluated the effects of hybrid ratio and fiber orientation. A multiscale damage model based on micromechanical damage and failure criteria was established via Abaqus/VUMAT, integrating stress amplification factors to bridge micro-meso-macro scales. Experimental results show that carbon fibers aligned with the loading direction yield hybrid composites with superior tensile properties to carbon fiber reinforced plastic (CFRP). Under impact loading, the hybrid beams possess energy absorption values intermediate between those of single fiber reinforced composite beams, among which the B-[C0/G90]6 hybrid beam attains the maximum costeffectiveness, exceeding the pure CFRP beam by 27%. The established multi-scale model accurately predicts laminate tensile and hat-shaped beam impact responses, with all relative errors within 10%. Based on the validated model, orthogonal discrete optimization was performed using mesoscopic carbon fiber bundle fraction and macroscopic ply angle as design variables. Results indicate ply angle most significantly influences cost-effectiveness, leading to a 28.74% improvement post-optimization. This work provides an integrated multiscale modeling and optimization framework for performance-cost balanced intra-layer hybrid HFRP thin-walled structure design.
We present QFedFormer, a federated transformer for dynamic electric vehicle (EV)-charging price prediction that combines quantization-aware training, SHAP-guided explainability, and blockchain-based incentives. The framework trains across distributed charging stations without centralizing user data, and programmable contracts set tariffs from forecasted demand and user-declared flexibility, while token rewards are derived from SHAP-based utility scores and anchored on-chain via Merkle proofs. On a real-world dataset, QFedFormer attains an energy-demand RMSE of 1.82 +/- 0.02 kWh and a tariff RMSE of 11.83 +/- 0.10 KRW/kWh (MAPE 2.7 +/- 0.2%) in the non-private baseline, outperforming FedAvg and Block-FeDL by 14.1% and 9.5%, respectively. Under client-level differential privacy (DP) with (a = 1.6, C = 1, p = 0.1, SDP = 10-5), QFedFormer achieves (e = 2.0, SDP = 10-5) after 50 rounds under a R & eacute;nyi accountant, with forecast accuracy degrading modestly to 1.95 kWh RMSE (similar to 7.1% relative increase vs. non-private baseline). Blockchain evaluation shows an average audit latency of 58 ms per audit round, while a permissioned Ethereum-compatible deployment sustains more than 500 client updates per minute with gas costs of similar to$0.039/client per audit round. These results indicate that QFedFormer enables accurate, privacy-preserving, and auditable coordination of EV-grid interactions, offering both regulators and service providers a practical deployment pathway.
