The sustainable deployment of edge intelligence for cable partial discharge (PD) identification is challenged by non-independent and identically distributed (non-IID) data and limited communication and energy resources. This article proposes a dynamic event-triggered personalized federated aggregation (DET-PFA) approach for accurate and energy-efficient PD identification on edge devices. DET-PFA decouples the PD identification model into a globally aggregated feature extractor and locally retained personalized classifiers, thereby enhancing generalization across clients, preserving adaptability to local data characteristics, and reducing parameter transmission. A dynamic event-triggered mechanism further suppresses unnecessary model uploads. By coordinating parameter decoupling with a dynamic event-triggered mechanism, DET-PFA reduces communication costs in both the parameter-space and temporal dimensions. Experiments on the collected PD dataset show that DET-PFA achieves an identification accuracy of 99.87%. Raspberry Pi 4 Model B deployment experiments further demonstrate reductions of 99.23% in total communication cost and 87.81% in energy consumption compared with conventional federated learning. These results demonstrate the practical feasibility and resource efficiency of DET-PFA and highlight its sustainable computing benefits for edge-based power equipment monitoring.
Intrusion Detection System (IDS) is a vital solution in the cyber defense architecture, and its implementation should guarantee secure data transactions. This role is becoming increasingly critical as ever-increasing technological democratization has transformed the digital field into a high-risk environment for cyberattacks, leading to an increasing volume of network traffic. Under these circumstances, IDSs will be in higher demand and consume more energy. To address this challenge, the objective of this study is to introduce new sustainability metrics, including Green Feature Extraction (GFE), Redundancy Efficient Rule (RER), Waste Rate (WR), and Path Selection Energy Consumption (PSEC), to quantify the environmental impact of IDS operations. The core of our study examines data flows, algorithms, deployment architecture, configuration, and feature engineering techniques to assess their impact on IDS sustainability. The experimental validation confirms the feasibility and effectiveness of the proposed metrics: the GFE analysis shows that using only 5 features achieves 99.58% detection accuracy with optimal energy efficiency, the WR analysis reveals that Random Forest achieves the lowest waste (0.72% vs. 15.48% for Logistic Regression), and the PSEC analysis demonstrates that Gain Ratio consumes 69% less energy than Chi-Square (29.5 kW vs. 95.4 kW). These findings demonstrate substantial potential for sustainable IDS design that balances robust threat detection with environmental responsibility.
This work provides a general formulation of the task scheduling problem in HCSs by further considering the thermal issue along with the objectives of schedule length, energy consumption, and reliability with fault-tolerance. Then, it provides a novel solution approach for the general task scheduling problem in HCSs. The proposed algorithm, which is an efficient estimation of distribution algorithm-based hybrid meta-heuristic approach, is designed by employing Bayesian optimization algorithm (BOA) as a global optimizer, binary search as a local optimizer, and adaptive archive radix strategy as a parameter learner of Bayesian network (BN). Our proposed BN-based task scheduling approach aims to optimize the objectives of HCSs simultaneously. After optimization, sampling, and mutation phase, the objectives of the solutions are calculated. The solutions appropriately assigned tasks to the cores, considering the characteristics of the system. Our proposed scheduling approach exploits complex dynamic voltage and frequency scaling levels by suitably exploiting the benefits of the heterogeneity of the processing elements. We evaluate our proposed method alongside six other multi objective evolutionary algorithms (MOEAs) using a sample task graph and real-world application benchmarks, such as MiBench, modeled as directed acyclic graphs on simulated heterogeneous multiprocessor systems on chip. The performance comparison is conducted using hyper-volume and inverted generalized distance metrics. Additionally, the multi-criterion decision making method (MCDM) is employed to demonstrate the overall superiority of our proposed approach. The results of all the considered performance metrics as well as the MCDM-based comparison indicate that our proposed method outperforms all the competing MOEAs.
Accurate wind power forecasting (WPF) is critical to power system planning and operational management, thereby enhancing grid reliability under large-scale wind power integration. However, most existing methods remain confined to a one-dimensional formulation, which inadequately captures turbine-level spatial heterogeneity and consequently limits the joint modeling of positional structure, multivariate interactions, and temporal dynamics. To overcome this limitation, we reformulate WPF as a three-dimensional spatiotemporal forecasting problem by reorganizing each time step into a multi-channel spatial frame according to turbine locations. On this basis, we propose CViTRNN, a multidimensional collaborative learning framework that combines a channel-aware vision transformer with a lightweight recurrent unit to model multivariate interactions and temporal dependencies under turbine-level spatial priors. Furthermore, a two-phase autoregressive workflow is devised to extend forecasting from single-turbine settings to cross-turbine scenarios. Extensive experiments demonstrate the superior accuracy and adaptability of the proposed model, while further confirming the soundness of its design and the interpretability of its forecasting behavior.
With the escalating demand for mobile data, accurate forecasting of network traffic is paramount for resource management and Quality of Service (QoS). While Long Short-Term Memory (LSTM) networks have been applied, they often struggle with long-range dependencies and incorporating static metadata. More advanced spatio-temporal models often operate as “black boxes,” limiting their operational utility. This paper proposes GNTFT, a novel, hierarchical framework centered on the Temporal Fusion Transformer (TFT) that prioritizes both accuracy and interpretability. Our system first segments the network into behaviorally homogeneous clusters, then learns static, graph-based spatial embeddings for each sector using a Graph Neural Network (GNN). These embeddings are fused with historical KPI data within a specialized TFT model for each cluster. We conduct a rigorous benchmark against a traditional LSTM, Spatio-Temporal Graph Convolutional Network (STGCN), Diffusion Convolutional Recurrent Neural Network (DCRNN), and the state-of-the-art Graph WaveNet (GWN). The results demonstrate that while GWN achieves the highest absolute accuracy (MAE), our proposed Full GNTFT framework achieves superior performance across all metrics (MAE: 110.12, MAPE: 22.05%), compared to the baseline “TFT Only” model (MAE: 158.82, MAPE: 31.42%). Crucially, the GNTFT’s hierarchical architecture provides deep interpretability; by training specialized models for each behavioral cluster, we demonstrate that the predictive drivers of network traffic are highly heterogeneous. This work presents a transparent solution that bridges the gap between raw predictive power and the actionable, trustable intelligence that state-of-the-art models cannot provide.
Serverless computing has emerged as a cloud computing paradigm that mitigates the complexities of resource provisioning and management while charging users only for the resources they consume. However, some inherent characteristics of serverless environments, such as stateless functions and scaling to zero, make it challenging to execute workflow applications that require data transmission between functions. As a result, serverless service providers need to offer efficient workflow execution services to encourage the adoption of these platforms. At the same time, the high energy consumption of serverless cloud data centers has not received sufficient attention from the research community. Accordingly, we introduce Transportation-based Energy-efficient Serverless Scheduling (TESS), an approach designed to address deadline-constrained workflow scheduling in serverless environments while reducing energy consumption. To achieve this, we first formulate the problem as a variation of the well-known Transportation Problem and then propose a scheduling algorithm based on solutions to this problem. Evaluation results demonstrate that the proposed algorithm reduces energy consumption by 10%–31% compared with state-of-the-art algorithms.
Heterogeneous multi-core processors (such as the ARM big.LITTLE architecture) have become the mainstream platform for edge computing and real-time embedded systems due to their dynamic trade-off between high performance and high energy efficiency. Currently, heterogeneous multi-core scheduling faces the challenge of collaboratively optimizing task execution time and system energy consumption. Existing methods have the problem of insufficient heterogeneous adaptability when dealing with tasks with complex dependencies. To address these challenges, this paper proposes an innovative Federated Bundled Hybrid Scheduling (FBHS) framework, which seamlessly combines federated and bundled scheduling through a dynamic selection mechanism guided by capacity enhancement bounds, alongside a differentiated resource allocation strategy tailored for high and low utilization tasks. Experimental results demonstrate that, in task scenarios ranging from 20 to 70 tasks, FBHS outperforms pure federated scheduling (PFS), pure bundled scheduling (PBS), completely fair scheduler (CFS), and evolutionary algorithm based heterogeneous earliest finish time (E-HEFT), achieving average reductions of 18.3%, 26.1%, 31.7%, and 10.10% in energy consumption and 45.8%, 50.95%, 56.2%, and 21.18% in execution time, respectively. Furthermore, the results were verified in real data streams such as Montage and CyberShake, highlighting the effectiveness of FBHS in simultaneously enhancing real-time performance and energy efficiency on heterogeneous multi-core architectures.
Medium Access Control (MAC) is critical for ensuring reliable and timely transmission of physiological data in Wireless Body Area Networks (WBANs). However, existing MAC designs fail to address tripartite challenges: priority-agnostic contention in dense networks, rigid resource allocation under dynamic channel conditions, and disjointed retransmission and resource scheduling mechanisms. This paper proposes a motion-aware priority-driven MAC (MAPD-MAC) protocol to overcome these limitations. The protocol synergizes contention-based and contention-free channel access mechanisms through three innovations: a priority-isolated channel access framework that eliminates cross-priority collisions and mitigates energy waste through dynamic contention windows, a body-motion-aware Markov model enabling adaptive packet sizing and slot allocation aligned with predicted channel stability, and a context-adaptive retransmission scheduler that prioritizes critical data while deferring non-urgent transmissions during fading episodes. Three distinctive features further distinguish this work: urgency-driven channel access mode switching, cross-layer interference management, and clinically integrated resource orchestration. Extensive simulations demonstrate significant improvements over state-of-the-art protocols, achieving a higher packet success rate, lower energy consumption, and reduction in average Age of Information (AoI) under dynamic and dense WBANs. These advancements collectively bridge the gap between stringent clinical demands and network resource constraints, propelling WBANs toward responsive and sustainable healthcare applications.
This study focuses on analyzing the influence of the number of processors on performance and energy consumption in high-performance computers. We have developed conceptual models for potential improvement factors that could be achieved. We have applied these models and validated them using real data obtained from recent homogeneous and heterogeneous supercomputers listed in the TOP500 and Green500 rankings. Unlike previous studies, this study uses a large number of HPC systems to support the results: 237 systems for performance analysis, 96 for power consumption, and 32 for energy efficiency. The proposed models provide a better understanding of current architectural trends and allow for the assessment of achievable improvements in both computational performance and energy efficiency by increasing the number of CPUs or accelerators in each system. The study concludes that such improvements are more effective when increasing the number of accelerator cores rather than the CPU cores.
The rapid growth of mobile crowdsensing (MCS) has motivated the deployment of unmanned aerial vehicles (UAVs) for large-scale, flexible, and timely data collection. How ever, trajectory planning in multi-UAV MCS remains highly challenging due to the dynamic and uncertain activation of points of interest (PoIs), stringent energy limitations, and safety risks arising from high-density UAV operations. Existing reinforcement learning approaches can face scalability and training stability challenges in large-scale coordination under dynamic and uncertain PoI settings, while many heuristic methods rely on oversimplified assumptions such as static PoIs, leading to suboptimal performance in dynamic environments. To overcome these limitations, this paper proposes Pre-Evo, a novel hybrid Prediction-Evolution framework that synergistically integrates predictive intelligence with evolutionary optimization for efficient and reliable multi-UAV coordination. On the prediction side, long short-term memory (LSTM) networks are employed to accurately forecast the active time windows of dynamic PoIs, thereby transforming temporal uncertainty into a tractable decision making input. On the optimization side, we design an enhanced multi-agent genetic algorithm (EnhancedMAGA) equipped with three tailored mechanisms. This co-design of predictive modeling and evolutionary search enables Pre-Evo to strike a balanced trade-off among maximizing task completion, minimizing flight distance and energy usage, and reducing collision risk in dense UAV swarms. Extensive experiments on both real-world urban traffic datasets and diverse synthetic scenarios demonstrate that Pre-Evo consistently outperforms state-of-the-art benchmarks. Notably, Pre-Evo also maintains robust generalization to large scale and highly dynamic environments, highlighting its potential for next-generation UAV-assisted MCS systems.
To extend the service life of energy storage in multi-microgrid system, a distributed optimal dispatch strategy for multi-microgrid system is designed in this paper. This strategy links the battery health status to the battery resale value, and the battery resale cost is taken into account in the economic cost. According to the characteristics of this strategy, a distributed model predictive control-based accelerated distributed augmented Lagrangian (DMPC-ADAL) algorithm is proposed to solve the resulting dispatch problem. The proposed algorithm exploits the structural sparsity of the adjacency matrix to accelerate distributed optimization. Through local computation and neighbor information exchange, it reduces the dependence on global communication while coordinating the microgrids to minimize the overall operating cost. The convergence property of the proposed algorithm is theoretically analyzed. Finally, the effectiveness and feasibility of this strategy for optimizing the operation of multi-microgrid systems are verified through simulation.
Accurate short-term load forecasting is critical for the safe and efficient operation of power systems, in the context of increasing demand fluctuations and the growing penetration of renewable energy sources. However, the strong non-stationarity of load sequences poses significant challenges to existing forecasting models, making it difficult to effectively capture complex temporal dynamics and rapidly changing load patterns. To address these issues, a Two-Stage Decomposition Network (TSD-NET) is proposed as a forecasting framework, which tackles the above challenges through staged prediction, progressive refinement, and structured output fusion. The two-stage architecture first captures coarse-grained temporal trends to establish a stable forecasting baseline, followed by targeted adjustments for short-term fluctuations, thereby alleviating the adverse effects of non-stationarity. To further enhance the model's capacity to learn fine-grained load features, the network incorporates a series of refinement blocks that iteratively extract previously under-learned patterns. A structured fusion module is employed to integrate outputs from all stages, improving predictive accuracy. Extensive experiments on real-world datasets from Australia, Shanghai, and Tetouan show that our method reduces MSE by up to 46.2% and MAPE by up to 23.7%, and improves $R^{2}$ by up to 4.5 percentage points compared with the best baseline, offering a robust and generalizable solution for dynamic load prediction in complex power-system environments.
In response to the escalating demand for low-power and high-performance digital architectures, the design of energy-efficient data buffering mechanisms has emerged as a critical area of focus, particularly within edge computing, artificial intelligence accelerators, and high-speed communication systems. Asynchronous memory buffers facilitate seamless data transfer across heterogeneous clock domains. However, conventional implementations are often plagued by considerable power inefficiencies. This work introduces a novel architectural enhancement of an 8-bit asynchronous Ultra Low Power FIFO (UlpFIFO), employing a hybrid power management (HPM) strategy that synergistically integrates clock gating and power gating techniques. This dual-pronged approach effectively mitigates both dynamic and leakage power dissipation. The proposed design is rigorously evaluated through implementation on both FPGA and 32nm CMOS ASIC platforms. Empirical results reveal a 25% reduction in dynamic power consumption, accompanied by a 12% decline in leakage power, at the cost of a modest 2.5% increase in silicon area.
Large language models (LLMs) are increasingly embedded in organizational workflows, raising concerns about energy use, costs, and data sovereignty. While benchmarks often celebrate cutting-edge models, real-world deployment requires a broader question: when is a smaller, locally deployable model “good enough”? This study evaluates eleven proprietary and open-weight LLMs on ten everyday occupational tasks, including summarizing, scheduling, and drafting emails and proposals. Using a dual-LLM evaluation framework, we automated task execution and standardized assessment across ten criteria of quality, accuracy, and ethical responsibility. Results show GPT-4o delivers top performance but with significantly higher costs and environmental impact. Smaller models such as Gemma-3 and Phi-4 performed strongly on most tasks, supporting their viability where cost-efficiency, local use, or privacy matter. A cluster analysis identified three groups-High, Balanced, and Baseline Performance Models-highlighting trade-offs between quality, control, and sustainability. Task type shaped effectiveness: conceptual tasks were most challenging, while aggregation and transformation tasks yielded better outcomes. We argue for shifting from performance-maximizing benchmarks to task- and context-aware sufficiency assessments that reflect organizational priorities. Our approach offers a replicable method to evaluate AI models from a sustainability perspective and provides practical recommendations for responsible LLM deployment.
Smart grids rely on advanced metering infrastructure (AMI) for real-time usage data, yet this connectivity introduces vulnerabilities. One critical threat is grid overloading cyberattacks, where an adversary manipulates demand to exceed safe limits and trigger blackouts. Such attacks can harness compromised smart meters to drastically raise the load and destabilize the network. Despite their severity, AMI data manipulation for overloading attacks has received limited attention. We propose a novel anomaly detection framework that combines household and neighborhood-level load predictions. Using deep learning models, our system predicts the electricity usage of each customer, as well as the aggregate community load, allowing detection of sudden deviations that signal an attack. To quantify anomalies, we introduce two metrics: an Abnormality Index (capturing the magnitude of peak deviations) and a Regularity Index (measuring energy consumption pattern consistency). These indices are fed into an ensemble model to detect threats. Crucially, our ensemble model combines multiple classifiers to mitigate adversarial manipulation, making it robust to data poisoning and evasion attempts. Extensive evaluation with a real dataset demonstrates our proposed model achieves a detection rate of 93% and a false alarm rate of 9% in the most challenging baseline conditions, and maintains 81% detection even under the hardest adversarial setting using white-box AutoAttack scenarios, outperforming existing contemporary methods. This integrated solution advances smart grid cybersecurity by combining predictive load modeling, novel anomaly metrics, and robust classification.
Data center power consumption continues to grow rapidly with the increasing demand for cloud and enterprise computing. CPU active power and cooling account for a major share of this energy use, yet manufacturers provide no visibility into each CPU's unique power–thermal behavior. Process variations and inverse temperature dependence (ITD) cause individual CPUs to exhibit distinct electrical characteristics — including voltage, dynamic capacitance, and leakage — which together determine their optimal operating temperature for maximum performance-per-watt. This paper introduces a new ITD-aware CPU power modeling framework that integrates empirical voltage, leakage, and capacitance measurements to predict total power across temperature and frequency with high accuracy. The model is validated on 22 Intel Xeon server CPUs, achieving less than 3% error across operating conditions. Using the model, we identify each CPU's optimal temperature that minimizes power while sustaining performance. Analysis of real cloud servers reveals that around 50% of modern CPUs operate roughly 10°C below their optimal point, wasting approximately 15% in excess cooling energy. Classifying CPUs by their optimal temperature enables data centers to group and cool them more efficiently, reducing energy use and improving sustainability without performance degradation.
Dynamic quadratic programming (DQP) widely exists in many optimization applications. Existing recurrent neural networks (RNNs) exhibit significant limitations in solving DQP, including lagging errors in gradient-based neural networks (GNNs) and fixed scaling factors in zeroing neural networks (ZNNs). Meanwhile, the matrix singularity may cause solution crashes in ZNNs. Additionally, existing RNNs for DQP are primarily implemented on von Neumann architecture-based computers, and this architecture exhibits low efficiency in handling repetitive computational tasks. To address these issues, this paper constructs an adaptive gradient recurrent neural network (AGRNN) to solve DQP. Furthermore, compared to the von Neumann architecture, field programmable gate arrays (FPGAs) have significant advantages in handling repetitive computational tasks, and thus, the AGRNN is implemented on an FPGA. The AGRNN significantly reduces GNNs' lagging errors, has a higher flexibility than ZNNs in adjusting scaling factors, and effectively mitigates matrix singularity issues. Additionally, the FPGA implementation further improves the AGRNN's efficiency in solving DQP. Numerical experiments show that the FPGA implementation achieves at least 2.68 times the computational speed of the von Neumann architecture. Meanwhile, experiments are conducted in the dynamic economic dispatch involving two generators, where the FPGA implementation of the AGRNN utilizes 5521 LUTs, 8311 FFs, and 339 DSPs.
Modern IoT and sensor networks generate vast amounts of data, posing significant challenges for storage, transmission, and real-time processing. Traditional approaches, such as compressive sensing and machine learning-based compression, often suffer from computational inefficiencies and irreversible data loss. This paper introduces Information Density as a quantitative metric to support sensor deployment and enable AI-driven virtual sensing. We propose a framework that leverages spatial, temporal and inter-modal correlations among sensor signals to perform sensing tasks even in the absence of physical sensors. Two complementary measures: (i) Phase in Eigen Space and (ii) Mutual Information, are developed to quantify and assess information density, enabling the selection of optimal sensor configurations across both intra-modality and cross-modality scenarios. Validated using real-world data from Madrid's smart city infrastructure, this framework demonstrates the feasibility of replacing physical sensors with virtual ones under bounded error conditions (e.g., achieving <3.21% mean error with a single sensor). The results highlight the potential for scalable and energy-efficient sensing systems in smart environments.
Multi-UAV systems are increasingly essential to Internet of Things (IoT) networks, offering agile platforms for data collection, Wireless Energy Transfer (WET), and edge computing. Although significant progress has been made in UAV-enabled architectures, persistent challenges remain due to stringent energy constraints, dynamic task demands, and the high-dimensional nature of multi-UAV coordination. This work seeks to optimize deadline-aware task quality in battery-constrained IoT networks using Mobile Edge Computing (MEC)-enabled UAVs, each assigned Core Tasks with strict deadlines and Supplementary Tasks to enhance quality when additional resources are available. Nevertheless, the joint optimization of UAV 3D trajectories, WET, and task scheduling under tight energy and deadline constraints introduces a dynamic and continuous decision-making challenge. To address this, we propose a novel multi-UAV framework built upon two key pillars: (i) a Beta-distributed policy network for modeling bounded, continuous control actions, and (ii) a multi-head attention-based critic for capturing inter-UAV interactions and global context. Accordingly, we formulate the problem as a multi-agent Markov Decision Process (MDP) and develop Beta-Attention Multi-Agent Proximal Policy Optimization (BA-MAPPO), a Deep Reinforcement Learning (DRL) framework based on Multi-Agent Proximal Policy Optimization. Extensive simulations demonstrate that BA-MAPPO substantially improves energy efficiency, task quality, and deadline compliance, outperforming existing baselines in dynamic IoT environments.
This work proposes a sustainable crop and fertilizer recommendation system using eco-friendly hybrid machine learning (EHML) and explainable artificial intelligence (XAI) models. The integration of EHML with XAI is introduced for the first time in the crop and fertilizer recommendation system. This work focuses on crop recommendations solely based on the micro and macro nutrients present in the soil, an aspect that has been largely overlooked till date. Based on various hybrid machine learning (HML) models such as CatBoost and Extreme Gradient Boosting, optimized with Sparrow Search Algorithm, Phasor Particle Swarm Optimization, and Grey Wolf Optimization, crops are recommended with accuracy ranging from 99.64% to 99.87%. The state-of-the-art work is also compared with hybrid models optimized with standard optimizers, such as Genetic Algorithm and Particle Swarm Optimization. To reveal critical drivers of crop selection and enhance the interpretability of the best-performed models, XAI algorithms (Local Interpretable Model Explanation and SHapley Additive exPlanations) are employed. The best-performing eco-friendly model is also recommended based on low ${CO}_{2}eq$ emissions and electricity consumption, helping to preserve natural resources. Additionally, a farmer-friendly web application is developed to provide instant crop and fertilizer recommendations for sustainable agriculture. Finally, the recommended fertilizers are also validated using real-time field data collected via IoT sensors.
