The nature of cloud cover over photovoltaic (PV) arrays causes significant variability in power generation, posing challenges for grid stability and energy management. Tracking cloud motion from sky images offers a promising solution for short-term PV forecasting; however, noisy images, unexpected cloud transitions, high computational burden, and temporal leakage in large datasets remain critical barriers. To address these challenges, we propose the multi-relational Attention-guided Mean-Max Spatiotemporal GraphSAGE Neural Network (AM2-STGNN). The framework begins by extracting spatiotemporal motion features from cloud images. As finding the connection between PV and Cloud motions is a challenge, the Apriori rule mining algorithm identifies interrelationships between cloud motion and spatiotemporal features, validating their correlation with high lift and confidence score. Subsequently, a multi-relational graph is constructed from the mined rules and spatiotemporal features for processing by the model. Within AM2-STGNN, the mean-max aggregation replaces the conventional Laplacian matrix, reducing time complexity. Experimental results demonstrate that the proposed model outperforms conventional machine learning, deep learning, and baseline Graph Neural Network (GNN) models, achieving superior R2 scores and surpassing the studies on the same dataset in short-term PV forecasting. In addition, temporal leakage during training is explicitly addressed through a three-stage temporal holdout evaluation. Finally, node-level interpretability analysis is illustrated to show the model’s decision-making process.
This paper presents an integrated framework for reliable ultra-short-term wind power forecasting and uncertainty quantification. The framework integrates robust data preprocessing, two-stage feature selection, and lag selection to construct quality inputs. At its core, it combines a bidirectional long short-term memory (BiLSTM) model with whale optimization algorithm (WOA)-based hyperparameter optimization to identify an optimized architecture and training configuration. To assess generalizability, the forecasting core is validated across two scales: a multivariate wind turbine dataset from Tenerife, Spain, and a univariate regional wind power dataset from Belgium. Results show an 80.02% reduction in normalized root mean square error (nRMSE) relative to the persistence model, while maintaining competitive accuracy against state-of-the-art deep learning models. Robustness is assessed through multi-step forecasts up to 60 min and statistical significance tests. Distribution-free prediction intervals generated by bootstrap resampling provide well-calibrated uncertainty information across both sites, supporting operational decision-making in wind-integrated power grids.
The rapid growth of digital healthcare systems and Internet of Medical Things (IoMT) devices increases the risk of unauthorized access and the leakage of sensitive patient data. Conventional authentication and encryption methods are inadequate for distributed resource-constrained healthcare environments, such as 5G edge networks. Traditional biometric systems are also vulnerable to template compromise, whereas centralized and decentralized Artificial Intelligence-based security solutions suffer from privacy breaches and computational latency. Addressing these challenges, this work proposes a Deep Learning based Zero-Knowledge Proof (ZKP) authentication framework for secure patient identity verification and protection of medical data. Therefore, the approach for behavioural biometric patterns is to first preprocess them to remove noise and normalize variations. Further, an Orca Spatial Attention Transformer Analysis (OSATA) model is employed for extracting robust spatial-temporal features representative of unique user behaviour. Such optimized features are further used to develop a zero-knowledge proof for user authentication without revealing raw biometric or medical information. Further, controlled access to healthcare servers is guaranteed through a secure verification and token-based authorization mechanism. The proposed system is implemented in Python, investigational results demonstrate improved security, efficiency, and privacy preservation compared with conventional authentication methods.
The Random Vector Functional Link (RVFL) framework provides a simple and effective classification solution using a single-layer feedforward structure with randomization. The ensemble and deep variants of RVFL utilize multiple layers to improve performance, whereas the generalized eigenvalue-based deep RVFL (edGERVFL) solves a generalized eigenvalue problem integrated with the deep RVFL architecture. Multiview learning methods use multiple data views for enhanced generalization, but existing deep RVFL-based models do not incorporate multiview information. We formulate a Multiview Ensemble Deep Generalized Eigenvalue RVFL (MV-edGERVFL) and its variants — Multiview Kernel edGERVFL, Multiview Improved edGERVFL, and Multiview Kernel Improved edGERVFL — to achieve improved classification performance with multiple views of complex and non-linear data. The proposed model integrates multiview feature extraction with a deep ensemble of randomized networks based on generalized eigenvalue-based classification. To enhance applicability, we introduce three specialized variants that address non-linearity and singularity while retaining a closed-form solution. Multiview Kernel edGERVFL performs kernel mapping for non-linear separability. The Multiview Improved edGERVFL reduces singularity issues in correlated or sparse data. The Multiview Kernel Improved edGERVFL handles non-linearity and singularity together for improved stability. Experiments on UCI and AWA datasets offer consistent improvements when multiple data views are available.
Randomization-based learning offers an efficient alternative to fully iterative training by fixing part of a model’s internal transformation and estimating only a small set of trainable parameters. This paradigm can substantially reduce training time and computational demand while retaining competitive predictive performance. It also provides a distinctive scientific lens through which the architectural bias and representational capacity of neural systems can be investigated in the absence of iterative learning. This editorial introduces the Special Issue on Randomization-Based Deep and Shallow Learning Algorithms, which brings together twelve contributions covering theoretical and architectural developments, systematic benchmarking, hybrid deep-randomized models, uncertainty quantification, and applications in computer vision, neuroimaging, renewable energy forecasting, industrial process modeling, intelligent transportation, maritime safety, and network science. Collectively, the papers demonstrate that randomization-based learning is evolving from a family of shallow, computationally efficient learners into a broader design principle for deep, ensemble, and uncertainty-aware systems. The Special Issue also exposes open challenges concerning principled random-feature design, reproducible benchmarking, scalable linear solvers, robustness, calibration, interpretability, and deployment under streaming and resource-constrained conditions.
Long-horizon multivariate time-series forecasting aims to predict multiple correlated variables over extended horizons from historical observations. This task is challenging because non-stationary series can contain overlapping trend, periodic and disturbance components whose spectral distributions vary across input windows. Existing frequency-domain models can reveal periodic information, but fixed filters, predefined frequency partitions or globally shared band boundaries may underuse predictive components when dominant periods, disturbances and cross-variable relationships change between samples. We propose Adaptive Frequency-Band Mixture of Experts (ABF-MoE), a frequency-domain mixture-of-experts model built around sample-adaptive spectral-band construction and expert routing. ABF-MoE learns band widths from each input spectrum, forms differentiable soft masks, refines the resulting complex spectra with band-specific residual experts and fuses them through an energy-informed learnable gate. The refined history is then passed to a frequency residual forecasting module that iteratively removes historical backcasts and accumulates future forecasts. Experiments on six standard public long-horizon forecasting benchmarks show competitive performance against recent frequency-domain, mixture-of-experts and general forecasting baselines. These results support adaptive spectral partitioning for non-stationary and multi-periodic multivariate forecasting under standard benchmark protocols.
This survey presents a comprehensive study of error correction and mitigation techniques in Solid-State Drives (SSDs) based on NAND flash memory, a high-density nonvolatile memory technology implemented as arrays of series-connected floating-gate memory cells, motivated by the escalating error vulnerability resulting from aggressive flash memory scaling. To ensure data integrity, contemporary SSDs employ advanced Error Correction Codes (ECC), supplemented by a wide array of auxiliary methods. The study systematically categorizes mitigation strategies according to the specific error types they target, thereby offering fine-grained insights into their design and application. In addition, it examines recovery techniques for uncorrectable errors and considers system-level approaches that contribute to improved device endurance and performance. This organized and multi-layered perspective provides a thorough understanding of the mechanisms that underpin reliability in modern SSD architectures.
Low altitude intelligent transportation systems (LITS) coordinate Autonomous Aerial Vehicles (AAVs) across shared airspace, but their wireless drone to ground links remain exposed to Global Positioning System (GPS) spoofing, radio frequency (RF) jamming, command injection, and future quantum attacks. Existing Internet of Drones authentication schemes usually rely on elliptic curve or symmetric primitives, while National Institute of Standards and Technology (NIST) transition guidance targets deprecation of quantum vulnerable public key mechanisms beginning after 2030, they also rarely connect authentication with privacy aware intrusion detection. We propose Post Quantum Low Altitude Intelligent Transportation Security (PQ-LITS), a multi layer framework that combines post quantum mutual authentication with federated telemetry anomaly detection. The authentication layer uses the Module Lattice Based Key Encapsulation Mechanism (ML-KEM-768, Federal Information Processing Standards (FIPS) 203) and the Module Lattice Based Digital Signature Algorithm (ML-DSA-65, FIPS 204) to establish a session key in three messages, with 0.82milliseconds (ms) latency over 1000 runs. The detection layer trains autoencoders across three ground control stations with non independent and identically distributed (non IID) attacks using Federated Averaging (FedAvg) and Gaussian differential privacy (σ=0.01). On 15,000 synthetic telemetry records, PQ-LITS reaches F1 = 0.7549, 95.7% of the centralised upper bound, while reducing communication by 58.2%. We give a security proof in the Quantum Random Oracle Model, reducing session security to Module Learning With Errors (Module LWE) and Module Short Integer Solution (Module SIS), and release the dataset, code, and benchmarks.
OAuth 2.0 is an authorization protocol standard widely adopted by mainstream service platforms, providing users with controlled authorization. In recent years, vulnerabilities have been continuously discovered within this protocol, making a comprehensive security review of the existing deployments extremely important. The existing security analysis methods require access to the source code which are typically not available from closed-source providers. To address this, we have designed an innovative end-to-end black-box verification framework that seamlessly integrates behavioral models with formal verification tools. To mitigate the inherent reliance on human expert analysis, we employed large language models to assist in intermediate steps, thereby automating the entire process. Crucially, the existing methods also fail to detect complex logical flaws involving multiple parties. To overcome this challenge, we have designed a distributed role modeling technique. It accurately reconstructs complex multi-party interactions, thereby capturing suspicious protocol behaviors. We implemented a prototype of MAFOAuth and conducted experiments on seven major OAuth platforms, successfully detecting all real-world vulnerabilities supplemented into the state machine path. Additionally, we uncovered three previously unknown robustness issues in widely used instances. The results demonstrate that our framework developed for analyzing black-box OAuth protocol implementations is effective.
The rapid integration of Renewable Energy Sources (RES), particularly solar photovoltaic (PV) and wind energy systems, has significantly transformed the operational dynamics of modern microgrids. Despite their environmental and economic benefits, the intermittent nature of RES and the widespread use of power electronic converters introduce major Power Quality (PQ) challenges, including voltage sags and swells, harmonic and suprachiasmatic distortion, frequency instability, voltage imbalance, and reactive power fluctuations. Conventional PQ mitigation approaches based on fixed-gain controllers and passive filtering techniques often exhibit limited adaptability under the nonlinear, stochastic, and time-varying conditions of renewable-energy-based microgrids.This paper presents a comprehensive and systematic review of recent Artificial Intelligence (AI)-based techniques developed for PQ improvement in renewable energy microgrids. The review methodology is based on a structured literature analysis of publications from 2021 to early 2026 collected from major scientific databases, including IEEE Xplore, Scopus, Web of Science, and ScienceDirect. The selected studies are critically analyzed according to AI methodology, application domain, computational complexity, real-time feasibility, and PQ mitigation performance.The review covers a broad range of AI approaches, including Machine Learning (ML), Deep Learning (DL), Reinforcement Learning (RL), Deep Reinforcement Learning (DRL), Graph Neural Networks (GNNs), Federated Learning, and hybrid intelligent optimization techniques. Furthermore, the paper examines recent advances in AI-enabled harmonic mitigation, reactive power optimization, voltage stability enhancement, edge intelligence, lightweight AI deployment, and autonomous microgrid control. Comparative analyses are also provided to evaluate the strengths and limitations of different AI architectures in terms of accuracy, adaptability, scalability, inference latency, and embedded implementation capability.The findings indicate that AI-based methods provide substantial improvements in adaptive PQ monitoring, disturbance classification, predictive control, and real-time mitigation compared with conventional techniques. However, several critical challenges remain unresolved, including data scarcity, model interpretability, cybersecurity vulnerability, computational burden, and reliable deployment under real-world operating conditions. Finally, future research directions are discussed, highlighting the growing importance of lightweight AI, Transformer-based architectures, hybrid cloud-edge intelligence, and physics-informed AI frameworks for the development of next-generation self-healing and autonomous smart microgrids.
