Reducing the carbon footprint of data centers (DCs) is increasingly important, and renewable energy can lower reliance on carbon-intensive grid electricity. However, uncertainty in renewable generation and workload demand creates spatiotemporal supply–demand mismatches, leading to renewable energy underutilization. Effective scheduling therefore requires long-horizon coordination to exploit workload flexibility across time slots and geo-distributed DCs, while remaining robust to uncertainty. Forecast-based planning enables such coordination by providing future information but is sensitive to prediction errors, whereas end-to-end reinforcement learning (E2E-RL) can adapt to runtime states but faces a large combinatorial decision space and lacks future-oriented guidance. To reconcile long-horizon coordination with runtime adaptation, we propose ReBase, a prediction-driven two-stage scheduling framework for renewable-aware DCs. In the first stage, decomposition-enhanced multi-step forecasting supplies future wind power and workload sequences to an offline mixed-integer linear program (MILP), which constructs a prediction-based baseline schedule. In the second stage, this baseline guides RL-based localized residual correction under realized deviations. This design reformulates online RL from complete schedule generation to localized residual adjustment while preserving long-horizon coordination. Experiments show that decomposition-enhanced forecasting generally improves multi-step prediction over the original long short-term memory (LSTM) predictor, with gains varying across data characteristics and prediction horizons. ReBase outperforms all non-Oracle schemes in scheduling performance. Compared with E2E-RL, ReBase improves wind utilization by 15.56%–22.86% and reduces carbon emissions by 12.58%–23.08%. ReBase approaches the Oracle benchmark with a wind utilization gap as low as 1.03%. These results highlight the potential of ReBase for low-carbon DC operation under renewable and workload uncertainty.
The increased penetration level of renewable energies into power systems brings along new challenges of intermittency and stability, for which appropriate energy storage solutions are needed. The increasing reliance on renewable energy resources brings with it the problems of variability and stability of power systems, which requires more advanced and effective energy storage planning methods. The optimization of heterogeneous energy storage planning is complex and intricate, which cannot be fully reflected by the conventional optimization method. To overcome this, this work proposes a two step hybrid optimization method in which the first step is a Linear Programming optimization (LP) to generate feasible solutions, and the second step is Simulated Annealing (SA) optimization for global optimization. The structure is two-stage, which results in a more optimal and feasible solution. Its novelty is based on using LP-SA for fulfilling different constraints while simultaneously trying to minimize costs worldwide to satisfy goals related to planning multiple aspects. The experimental results have been found effective by having a Net Present Value (NPV) of 0.345, Levelized Cost of Storage (LCOS) of 0.98, Return on Investment (ROI) of 0.179 and Energy Storage Efficiency (ESE) of 0.75. The proposed approach is more economically viable and efficient than other approaches such as LFSCOE, PEM RFC and HSS models. The key advantage of this model is that it can integrate conflicting constraints and deliver technically and economically optimal solutions, making it an extremely promising tool for (strategic) siting of energy storage systems in future smart grids powered by renewables.
Coastal ecosystems play a critical role in climate adaptation, biodiversity conservation, and sustainable environmental management. This study presents an integrated geospatial framework for evaluating climate adaptation capacity, coastal vulnerability, Nature-Based Solutions (NbS) suitability, and regenerative landscape development in the Yellow River Delta National Nature Reserve, China. Multi-source datasets (2020-2025), including Landsat 8/9 imagery, SRTM Digital Elevation Model data, climate variables, and coastal ecosystem information, were analyzed. Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI), Modified Normalized Difference Water Index (MNDWI), and Bare Soil Index (BSI) were combined with a Hybrid Random Forest–Extreme Gradient Boosting (RF–XGBoost) model for Land Use and Land Cover (LULC) classification. Analytical Hierarchy Process (AHP) and GIS-based weighted overlay analysis assessed coastal vulnerability, NbS suitability, and climate adaptation capacity. The Hybrid RF–XGBoost model achieved 96.10% classification accuracy with an F1-score of 96.05%. The framework identified vulnerable coastal zones, ecological infrastructure networks, and priority areas for wetland restoration, living shorelines, and ecological corridors, supporting resilient coastal planning and sustainable ecosystem management.
Mixed-criticality systems (MCS) are widely adopted in safety-critical embedded platforms where tasks with different assurance levels share limited computational resources. With the increasing power density of modern multicore processors, energy consumption and thermal management have become critical challenges in mixed-criticality scheduling. Since energy consumption and processor temperature are tightly coupled, optimizing energy alone is insufficient to effectively control peak temperature. While existing studies have investigated energy-aware scheduling or thermal-aware techniques independently, their joint optimization in semi-clairvoyant mixed-criticality systems, particularly on heterogeneous multiprocessor platforms, remains largely unexplored.This paper proposes a two-stage thermal-aware and energy-efficient scheduling framework for imprecise mixed-criticality tasks. In the offline stage, an energy-aware task partitioning algorithm is developed to map tasks to heterogeneous processors, minimizing energy consumption while ensuring schedulability. In the online stage, semi-clairvoyant execution information is exploited to reclaim dynamic slack time and allocate it to tasks with higher steady-state temperatures, thereby reducing peak processor temperature through dynamic voltage and frequency scaling. Extensive simulations on heterogeneous mixed-criticality task sets demonstrate that, against the strongest competing baseline WFOPTRA (CU-WFD partitioning combined with our online slack-reclamation algorithm OPTRA), the proposed framework reduces system energy consumption by up to 11.75% and peak processor temperature by up to 6.49°C; against the unmanaged First-Fit EDF baseline (FFEDF), the corresponding reductions reach 13.76% and 7.61°C. These gains are achieved with negligible online overhead. We further provide formal schedulability-preservation arguments for slack reclamation under heterogeneous frequencies and semi-clairvoyant mode transitions.
The dense proliferation of Internet of Things (IoT) devices with heterogeneous quality-of-service (QoS) requirements creates substantial operational-energy challenges for wireless data acquisition in edge-assisted sensing and monitoring systems. This paper develops an operational energy-efficiency (EE) framework for reconfigurable intelligent surface (RIS)- and unmanned aerial vehicle (UAV)-assisted heterogeneous IoT access. The framework jointly optimizes quasi-stationary three-dimensional UAV placement, RIS passive beamforming, device admission, subchannel scheduling, and transmit-power allocation. The system model distinguishes mission-critical and massive-access tiers, accounts for cross-tier interference and minimum service ratios, and incorporates amplifier inefficiency, circuit power, RIS controller and element consumption, UAV hovering, repositioning energy, and a finite UAV battery. The resulting mixed-integer nonlinear fractional program (MINLFP) is addressed using an inexact Dinkelbach–block coordinate descent–successive convex approximation framework. Locally tight rate minorants, semidefinite relaxation, adaptive trust-region updates, feasibility-controlled binary recovery, and post-quantization resource re-optimization are employed to obtain a locally stationary feasible solution without claiming global optimality. Over 200 paired Monte Carlo realizations, the proposed method achieves 7.20 Mbits/J under the nominal configuration, improving upon the No-RIS and literature-adapted Max-Rate benchmarks by 22.03% and 46.94%, respectively. The maximum observed EE is 12.80 Mbits/J at 240 RIS elements, while increasing the surface to 320 elements reduces EE by 10.16% because of the additional hardware-power overhead. At 80 massive-access devices, the method achieves 7.60 Mbits/J while satisfying the prescribed admission and minimum-rate requirements. Additional results quantify finite-resolution phase losses, UAV-repositioning break-even distances, imperfect-channel-state-information sensitivity, relaxation gaps, feasibility-recovery success, and runtime scalability. These results show that RIS–UAV cooperation provides an operational EE advantage only within identifiable hardware, mobility, and channel-accuracy conditions.
