ABSTRACT This paper presents a progressive scene transmission framework for massive distributed virtual environments based on a scalable two-layer P2P architecture with multi-level regions of interest (ROI). The proposed design integrates global coordination by super-peers and localised diffusion by swarm peers to achieve adaptive, perceptually prioritised transmission. A cognition-driven ROI model and predictive scaling mechanism enhance responsiveness and stability under heterogeneous network conditions. Experimental results demonstrate substantial improvements in latency, coverage, and consistency compared with classical P2P and centralised methods, confirming the efficiency and robustness of the proposed architecture for large-scale interactive systems.

ABSTRACT This paper addresses automated document assessment in blended learning environments, where evaluation requires multidimensional analysis. We propose BLAD-BERT, a neural framework that integrates BERT-based contextual representations with multi-dimensional assessment of content relevance, structure, fluency, and readability. To enhance evaluation precision, auxiliary metadata such as submission time and course context are incorporated. A semi-supervised learning strategy with pseudo-label guidance is further introduced to improve performance in low-resource settings. Experiments on real-world datasets demonstrate that BLAD-BERT outperforms baseline models in accuracy and interpretability.

ABSTRACT Federated learning-based automatic voltage control (AVC) in coordinated main–distribution networks is challenged by dynamic topologies, heterogeneous data, communication delays, and client dropout. This paper proposes DT-ADC-FedAvg, a dynamic topology-aware and delay-compensated asynchronous federated learning framework for collaborative AVC optimisation. The method adaptively handles client participation variations, compensates stale asynchronous updates, and integrates robust aggregation with drift detection to mitigate data heterogeneity and network churn. Experiments on three real-world and synthetic power system datasets show that DT-ADC-FedAvg achieves faster convergence, higher control accuracy, and stronger robustness against delays and dropout compared with state-of-the-art federated baselines.

ABSTRACT We propose a deep learning-based scheduling framework for cloud platforms handling multidimensional data streams. Our model combines temporal modelling, graph-aware reasoning, and entropy regularisation to achieve both rapid convergence and robust task assignment. Experiments on Google, Alibaba, and Microsoft datasets show that our method significantly reduces response time and queue length while improving task completion rate and concurrent throughput. Compared to SVM, RF, and MSCNet, our approach demonstrates superior performance across system-level and robustness metrics, validating its practical deployment potential.

ABSTRACT This study presents a pilot bidirectional Wireless Sensor Network (WSN) architecture designed to measure the fidelity of Digital Twins (DTw). Unlike standard systems that simply collect data, this approach creates a feedback loop from the digital to the physical twin, ensuring tighter synchronisation. We combined edge-enabled WSNs with IoT interfaces to track orientation and position with minimal delay. While Inertial Measurement Units (IMUs) proved reliable for orientation, we employed sensor fusion to correct positional drift. Achieving a maximum relative error of just 2.1%, this method provides a quantifiable baseline for predictive maintenance across the aerospace, manufacturing, and energy industries.

ABSTRACT The rapid expansion of IoT devices demands efficient task scheduling in fog-cloud infrastructures. This study presents a Heuristic-Guided Butterfly Swarm Optimisation (BSO) algorithm, integrating the Minimum Completion Time (MCT) heuristic into BSO’s initialization phase to enhance convergence and scheduling quality. A utility function balances processing time and execution cost while ensuring scalability across heterogeneous workloads. Simulations with 40-500 tasks demonstrate superiority over TCaS, MPSO, BLA, and RR algorithms, achieving up to 33.5% faster execution in fog-only settings and 64.89% higher scheduling efficiency in fog-cloud environments with minimal cost overhead. These results confirm the proposed scalable, cost-efficient solution for IoT-driven scheduling.

ABSTRACT The proliferation of cyber–physical systems (CPS) and heterogeneous multi-cloud infrastructures poses significant challenges for real-time, compliant, and adaptive resource orchestration. This paper proposes KG-ACS, a knowledge graph-based adaptive constraint scheduling framework for intelligent workload management in distributed CPS environments. KG-ACS models resource interdependencies, operational states, and constraints through a dynamic knowledge graph, enabling closed-loop, constraint-aware scheduling via semantic reasoning and hybrid optimisation across edge–cloud layers. Experiments on four large-scale public datasets show that KG-ACS reduces makespan by up to 18%, decreases compliance violations by 25%, and improves resource utilisation by 14% compared with state-of-the-art baselines.

ABSTRACT This study proposes DK-POL, integrating deep reinforcement learning with dynamic knowledge graph reasoning. Through semantic alignment, dual-channel fusion, and adaptive constraint optimisation, DK-POL consistently outperforms DQN, PPA, and SCA. Task success rates reach 91% on Freebase, 94% on CompGCN, and over 80% on MAG240M, with constraint violations nearly eliminated. Under 30% feature noise, DK-POL maintains 83.8% accuracy. Reasoning analysis reveals deeper relational traversal with low overhead (4.8 ms), demonstrating strong robustness, interpretability, and scalability across diverse decision-making scenarios.

ABSTRACT With the rapid advancement of large language models (LLMs), their application in smart grid knowledge services is increasingly promising, yet challenged by data sensitivity and output unpredictability. This paper proposes GRASP, a secure response architecture that integrates trusted execution environments, adaptive input–output auditing, and knowledge-grounded verification to enhance LLM trustworthiness in power systems. GRASP isolates inference processes, filters malicious inputs, constrains risky outputs, and reinforces factual consistency through a domain-specific knowledge graph. Experimental results demonstrate that GRASP significantly improves the safety, accuracy, and reliability of smart grid question-answering systems.

ABSTRACT Vehicle-to-everything (V2X) communication enables real-time interactions between vehicles and roadside infrastructure, supporting safety applications such as hazard detection and traffic awareness. Secure and efficient routing remains challenging due to latency, security, and privacy concerns. This paper presents a Reputation-based Awareness Routing protocol for Secure V2X communication (ReA-SR), which integrates pseudonym-based authentication and a trust scheme into clustering, selecting high-reputation vehicles as cluster heads. Analytical and simulation results show ReA-SR reduces latency, maintains packet drop below 5%, and achieves reception above 90%, improving security and efficiency for next-generation Intelligent Transportation Systems (ITS).