Accurate temporal monitoring of power-system operating states is increasingly important in renewable-rich and measurement-intensive grids. This paper presents an interpretable Temporal Fusion Transformer (TFT)-based framework for multivariate power-system state prediction from sequential measurements. The framework combines variable selection, gated residual processing, recurrent temporal encoding, multi-head attention, and post-hoc feature-attribution analysis. The learning task is explicitly formulated as temporal state prediction: a fixed window of historical measurements is mapped to a future state vector, rather than reconstructing the instantaneous state through a conventional static estimator. The method is evaluated on Monte Carlo operating scenarios generated for the IEEE 118-bus benchmark. In response to robustness and reproducibility requirements, the revised study includes five independent random-seed trials, paired significance testing, component ablation, Gaussian-noise and missing-data experiments, hyperparameter sensitivity analysis, and computational-complexity measurements. Across five seeds, the proposed TFT obtains an RMSE of 0.03486±0.04756 and MAE of 0.01988±0.02692 , compared with RMSE values of 0.12508 , 0.13178 , and 0.22107 for GRU, LSTM, and Transformer baselines, respectively. Controlled corruption experiments show progressive degradation as noise and missing-data severity increase. The interpretability analysis is reported conservatively as model-level explanatory evidence, and its limitations are quantified through an independent permutation-importance comparison.
Industrial predictive maintenance (PdM) requires models that remain computationally manageable while handling heterogeneous sensor data and changing operating conditions. This study evaluates FusionXNet-Edge, a compact multi-branch architecture that combines CNN, BiLSTM, GRU, and MLP components, with an auxiliary branch that receives physically motivated engineered features. The methodology assesses the model on the AI4I, MetroPT3, and Engine datasets using leakage-controlled preprocessing, a common single-step input dimensionality, fixed source-domain operating thresholds, and both in-domain and cross-dataset evaluation. The results show competitive in-domain performance while keeping the parameter count at approximately 16 000. Multi-seed ablation experiments show that the auxiliary physics-guided features yield only small, dataset-dependent changes in predictive performance, indicating that most of the model’s capability comes from the compact fusion backbone. Additional component ablations show that simpler variants can remain competitive in selected datasets, particularly under the single-step input configuration. Cross-dataset experiments reveal strongly asymmetric transfer behaviour, and additional schema-alignment diagnostics show that a common dimensional interface does not imply equivalent physical feature semantics across datasets. These findings position FusionXNet-Edge as a compact, deployment-oriented PdM architecture rather than a physics-constrained or universally transferable model. The study also identifies source-target distribution shift, operating-threshold mismatch, and feature-semantic differences as important limitations for cross-dataset deployment.
Developing reliable automated medical image classification systems remains challenging because annotated medical datasets are often limited, lesion categories are highly imbalanced, and high-confidence errors can occur even when conventional classification performance is strong. This study presents a reliability-oriented skin lesion classification framework based on pretrained DINOv2 representations, Multi-Head Self-Attention (MHSA) feature refinement, a lightweight multilayer perceptron classifier, and uncertainty-aware inference. DINOv2 is used as a pretrained visual backbone, while the downstream classification components are trained in a supervised manner on a lesion-disjoint partition of the HAM10000 dataset. Monte Carlo Dropout is employed during inference to quantify predictive uncertainty and support selective prediction, while post-hoc temperature scaling is used for probability calibration. On the fixed 992-image test set, the representative DINOv2 + MHSA configuration achieved an accuracy of 82.96%, weighted F1-score of 0.8342, macro F1-score of 0.7502, macro AUC of 0.9601, and weighted AUC of 0.9551. Comparisons with ImageNet-pretrained ViT-B/16, Swin-Tiny, and EfficientNet-B0 provided additional context for conventional classification performance. The reliability analysis showed that uncertainty-based error detection achieved an AUROC of 0.8293 and risk–coverage analysis obtained an AURC of 0.0502. Temperature scaling reduced the Expected Calibration Error from 0.0911 to 0.0369, while the Brier Score decreased from 0.2651 to 0.2467. The MHSA ablation showed improved feature-space organization and ranking-based discrimination, although threshold-dependent classification metrics did not improve uniformly. Limited-data experiments further examined performance when only 10%, 25%, 50%, and 100% of the training data were available. Overall, the study evaluates DINOv2-based skin lesion classification from both predictive and reliability perspectives and emphasizes uncertainty estimation, calibration, selective prediction, representation analysis, and data efficiency rather than claiming uniformly superior classification accuracy.
The advent of next-generation networks, such as Sixth Generation (6G, brings unprecedented demands for high security, ultra-low latency, exceptional reliability, and enhanced capacity. A critical enabler of these networks is reconfigurable wireless network slicing, which allows operators to deploy multiple virtualized network instances on a shared infrastructure, thereby enhancing Quality of Service (QoS). Achieving optimal QoS requires dynamic reconfiguration and optimization, which can be effectively realized through Artificial Intelligence (AI) and Machine Learning (ML) techniques. AI-driven solutions are pivotal in ensuring intelligent decision-making for network management and mitigating network slice failures. In this paper, we introduce an ensemble machine learning model to predict the optimal network slice. The proposed model’s performance is validated across diverse applications, including Smartphones, Smart Transportation, Smart Cities and Homes, Public Safety, Industry 4.0, Healthcare, and AR/VR/Gaming. Achieving an overall accuracy of 98.78%, the model demonstrates its potential to revolutionize network management in next-generation wireless systems.
Energy-efficient operation in multi-user MIMO–OFDM networks is increasingly constrained by redundant subcarrier power allocation and inefficient guard-interval handling under dynamic traffic demand. Existing carrier-adjustment and optimization approaches often overlook the coupling between instantaneous subcarrier activity and effective transmission demand, leading to avoidable energy expenditure and interference. This paper proposes a K-Demand–based Alignment Process (K-DAP) that enables demand-aware carrier alignment through FFT-assisted non-overlapping frequency identification, allowing energy to be allocated only to effective (active and non-replicated) subcarriers while suppressing replicated or unused frequencies. In addition, K-DAP restructures consecutive guard-time intervals to prevent allocation overlap and stabilizes the number of aligned carriers across transmission periods. A detailed analytical formulation and simulation framework are developed to evaluate K-DAP under varying user loads and channel conditions. Compared with state-of-the-art baselines (AOT-GA, APCO, and AF-DT), results show that K-DAP improves energy efficiency by 8.82%, increases average sum rate by 10.63%, and reduces average error rate by 10.48% under maximum-user scenarios. Overall, K-DAP provides a scalable carrier-level resource coordination mechanism for energy-aware multi-user MIMO–OFDM networking without sacrificing throughput.
