最新刊期
卷 50 ， 期 19 ， 2026
-
Green Power Directly-connected Data Centers: Coupling Logic, Key Technologies, and Application Modes
摘要：Driven by the goals of carbon emission peak and carbon neutrality and the rapid growth of the digital economy, data centers, as critical computing power infrastructure, are under urgent pressure to transition toward green energy. However, current green power procurement modes, lacking spatiotemporal consistency and physical traceability, not only face accusations of “greenwashing” but also struggle to reconcile cost-effectiveness with 24/7 carbon-free energy supply. Green power direct connection, which enables the clear physical traceability of electricity and offers a promising solution to these challenges. On the basis of relevant Chinese policy documents, this paper systematically investigates the coupling logic, key technologies, and application modes of green power directly-connected data centers. First, it is shown that the inherent characteristics of data centers—externally rigid yet internally elastic—are highly compatible with the Chinese policy framework that emphasizes “load-determined generation, shared responsibility, and grid-friendly operation”. To address source-load mismatches under direct connection, a technical system encompassing “perception-decision-execution” is constructed, covering full-scope state prediction, multi-objective optimization decision‑making, and dual regulation via energy storage and computing power. On this basis, the most feasible application mode under current policies—the grid-connected virtual power plant (VPP) mode—is analyzed, and the future evolution toward a park-level cluster-based ecological sharing mode is explored. This study aims to provide theoretical support and practical references for the green low-carbon transition and high-quality development of data centers in China.关键词：new power system;data center;computing power;green power direct connection;virtual power plant (VPP);source-grid-load-storage integration4|1|0更新时间：2026-10-10 - 摘要：The transformation of energy structure and the advancement of the carbon emission peaking and carbon neutrality strategy have continuously increased the demand for flexibility resources in power systems. Data centers (DCs) are commonly equipped with large-scale uninterruptible power supply (UPS), and their redundant energy storage resources have the potential to participate in grid support. Given conventional UPS systems have long been considered as contingency reserve, their utilization rates under normal operating conditions are low, and the technical system for active grid support is unclear, a systematic review is conducted on DC-UPS-based active grid support (DCUGS). The application basis of DCUGS is analyzed from the perspectives of power supply architecture, redundant capacity, policy incentives, economic feasibility, and engineering practices. The core connotation of DCUGS is clarified, with reliable power supply for critical information technology loads as the prerequisite and the release of redundant UPS capacity as the boundary. A four-layer technical architecture consisting of UPS topology, flexible control, cluster interconnection, and optimization decision-making is constructed. Furthermore, key technical challenges are summarized, including bidirectional energy routing, active/reactive power decoupling control, dynamic security region characterization of UPS clusters, and multi-market coupling decision-making. Finally, future research directions are discussed, including high voltage direct current power supply, modular cascaded topology, secure operation of large-scale UPS clusters, and regulation under multidimensional constraints.关键词：data center;uninterruptible power supply (UPS);grid support;energy storage;security region0|0|0更新时间：2026-10-10
- 摘要：Against the backdrop of global digital transformation and the accelerated development of artificial intelligence, data centers have become significant loads in power systems. Focusing on high-proportion, highly interruptible general computing tasks in data centers, this paper proposes a price-signal-guided modeling and optimization method for the load flexibility of general computing tasks, while balancing task performance and energy costs. Firstly, an energy-consumption characteristic model correlating task execution power with completion time in a single-server scenario is established, using the number of threads and processor frequency as key adjustable parameters. The impact of different task allocation methods on energy efficiency in a multi-server environment is analysed, thereby providing adjustable ranges for increasing and decreasing power for general computing tasks, along with corresponding time windows, to characterise the load flexibility of general computing tasks. By leveraging electricity price signals to incentivise the exploitation of load flexibility in data centers, a collaborative optimization model for the performance and energy efficiency of general computing tasks is constructed. By jointly optimizing task allocation, the number of threads and processor frequency, and planning data center energy consumption levels across different tariff periods, operation costs are reduced whilst meeting task deadlines and server resource constraints. The results of the case studies demonstrate that the developed model accurately captures the energy-consumption characteristics of general computing tasks, while the flexibility assessment confirms its significant potential for load regulation. Optimal dispatch based on electricity prices reshapes the load and reduces costs while meeting constraints. This provides a viable pathway for data centers to participate in the flexible regulation of the power system at a granular level.关键词：artificial intelligence;general computing;data center;load flexibility;energy-consumption characteristic modeling;electricity price;optimal scheduling1|0|0更新时间：2026-10-10
- 摘要：The data center introduces renewable energy power supply to achieve green and low-carbon operation. However, the fluctuation and stochasticity of renewable energy make it difficult to be efficiently utilized by data centers. The primary thermal medium in data centers exhibits cold storage capability to absorb and release thermal energy while delaying temperature changes, providing a foundation for flexible power consumption adjustment for cooling system. To overcome the limitations of existing thermodynamic models of data centers in describing overall thermodynamic processes, a Sage-Husa Kalman filter based data center thermodynamic dynamic estimation (SDTD) model is proposed. To enable high-level renewable energy accommodation in cooling systems, a variable-temperature operation strategy is proposed based on the SDTD model to maximize the renewable energy accommodation. The model predictive control method is used to dynamically adjust the operation temperatures of cooling system, thereby adjusting power consumption to actively match the output of renewable energy. Experimental results validate the accuracy of the proposed SDTD model and the effectiveness of the proposed variable-temperature operation strategy, achieving optimal alignment between cooling energy consumption and renewable energy while improving the renewable energy accommodation rate.关键词：data center;thermodynamic model;cooling system;variable-temperature operation;Kalman filter;model predictive control;renewable energy accommodation2|0|0更新时间：2026-10-10
- 摘要：As data centers continue to grow in scale and computational density, thermal management has become one of the critical factors affecting their energy efficiency and reliability. Building accurate thermal models is essential for implementing effective thermal management. Aiming at the modeling problem for thermal state of data centers, the construction mechanism of thermal state model for Bayesian data center based on real-time temperature measurement is proposed. First, based on the thermodynamic propagation structure of data centers, a dynamic thermal model is proposed to describe the central processing unit, server chassis, and environmental temperature, in order to characterize the thermal state of data centers over time. Then, to address the strong nonlinearity and prior defects in the thermodynamic mechanism model of data centers, a layered adaptive importance sampling technique based on Bayesian inference is developed. Simulation results demonstrate that the proposed method achieves high-accuracy modeling of thermal state models for data centers and exhibits high applicability under various prior settings.关键词：data center;thermal state;thermodynamic propagation;importance sampling;Bayesian inference;load modeling0|0|0更新时间：2026-10-10
-
Multi-agent Deep Reinforcement Learning-driven Economic Optimal Scheduling Strategy for Data Centers
摘要：To realize the economic operation of data centers, this paper proposes an economic optimal scheduling method for data centers based on multi-agent deep reinforcement learning. First, a unified model covering three layers, namely chips, racks and the data centers, is established to characterize the electric power-thermal power-computing power coupling relationship among the computing system, cooling system and energy supply system, and formulate the optimization problem. Second, a multi-agent optimal operation framework is designed to define the observation space, action space and reward function for three types of agents, i.e., load scheduling agent, data center temperature regulation agent and battery energy management agent. Third, the above optimization problem is transformed into a Markov decision process, the economic operation objective is mapped into a collaborative scheduling task among multiple agents, and a cooperation mechanism among agents is constructed based on the Actor-Critic framework. Finally, the effectiveness of multi-agent cooperation is analyzed using scheduling results randomly selected from some day in July. Comparative experiments with multi-agent proximal policy optimization, proximal policy optimization and model predictive control, as well as tests under typical daily scenarios in different seasons, fully verify the effectiveness and superiority of the proposed method.关键词：data center;multi-agent;deep reinforcement learning;load scheduling;temperature regulation;battery energy management;economic optimal scheduling;data-driven0|0|0更新时间：2026-10-10 - 摘要：As an important regulable resource on the load side of the power system, the thermoelectric coupling characteristics between server computing power consumption and cooling power consumption in a data center (DC) directly affect its regulation potential and economic benefits in demand response. To ensure the accuracy of capacity reporting for DCs participating in demand response and to avoid economic penalties caused by assessment deviations, it is necessary to precisely quantify their regulation capability across various intraday time periods. To this end, this paper proposes a quantification method for data center regulation capability that accounts for variations in cooling power consumption under varying central processing unit (CPU) operating conditions. Firstly, a correlation model between workload and server operating temperature is established to calculate server temperature distribution and cooling power consumption based on CPU working condition changes. Secondly, the energy efficiency response mechanisms under spatial transfer and temporal postponement strategies are analyzed in depth, revealing the marginal diminishing returns and saturation phenomena of the contribution of cooling power consumption changes to regulation capability as the proportion of transferred or postponed workload increases. Finally, simulation based on real-world server cluster data are conducted to quantify the demand response regulation capability of DCs considering thermoelectric coupling effects. Simulation results verify that the proposed method effectively corrects regulation capacity assessment deviations, providing a precise quantitative basis for data centers to mitigate economic penalty risks and for the power grid to formulate safe scheduling decisions.关键词：data center;computing power-electric power coordination;demand response;server computing power consumption;cooling power consumption;thermoelectric coupling;regulation capability1|0|0更新时间：2026-10-10
- 摘要：The rapid development of artificial intelligence and other technologies has led to a surge in computing power demand in data centers, which in turn results in enormous energy consumption. However, the rigid computing power demand and power supply modes of large-scale power grids make it difficult for data centers to effectively reduce carbon emission and achieve green economic operation. Therefore, fully exploiting the flexibility of electric power, computing power and thermal power is crucial for improving the dynamic matching between electricity supply and computing demand, as well as enhancing the accommodation of renewable energy in data centers powered by green electricity. This paper first quantifies and characterizes the uncertainty of renewable energy based on machine learning methods, designs a two-stage distributionally robust optimization framework for coordinated operation of electric power, computing power and thermal power, and establishes multi-energy coupling modeling. Second, a batch task transfer coefficient matrix is introduced to quantitatively analyze the flexible regulation characteristics of controllable computing tasks. Finally, key indicators affecting the economic performance of data centers are investigated through coordinated optimization. The results of the case study show that the multi-energy coordinated operation can reduce costs by 49.75%, of which regulating computing power flexibility alone improves economic performance by 48.5%, indicating that the flexible regulation characteristics of computing power tasks are crucial for economic operation关键词：data center;machine learning;distributionally robust optimization;computing power;multi-energy coupling;flexibility;flexible regulation;electric power-computing power-thermal power coordination1|0|0更新时间：2026-10-10
- 摘要：Electricity market prices exhibit long-term volatility. The lack of risk-hedging means in the single spot market leads to excessively high electricity consumption cost and limited profit space for data centers, which weakens their willingness to participate in the market and restricts the coordinated electric power-computing power development. Therefore, the contract for difference is introduced as a risk hedging tool, and a game trading strategy for data center aggregators in the medium- and long-term and spot markets is proposed. First, a three-stage electricity purchase and sale strategy is designed for the data center aggregators to participate in the electricity market. The basic electricity price and electricity quantity are locked through the contract, and the electricity purchasing quantities of data center aggregators are adjusted according to the spot electricity price, realizing the hedging against price fluctuation risks. Second, considering the spatio-temporal migration penalty costs of data centers and compensation costs for users, the differential retail electricity price is set based on contributions, guiding data centers to adjust their electricity consumption strategies and balancing the economy and quality of service for users. Finally, a three-layer game trading model is constructed, with the objectives of maximizing the net profits of the upper-level power grid supplier and the middle-level data center aggregators, minimizing the total costs of the lower-level data center operator, respectively. Case analysis results show that the proposed strategy can improve the benefits of the power grid suppliers and data center aggregators and reduce the costs of data center operators, realizing the global coordinated optimization of electric power and computing power.关键词：data center;aggregator;electricity market;spot market;electricity price;contract for difference;game trading;electric power-computing power coordination0|0|0更新时间：2026-10-10
- 摘要：Under the wave of artificial intelligence, the rapid development of data centers (especially artificial intelligence data centers) has brought growth in power load and they become an important source of power system flexibility. Data center clusters can achieve spatial migration of electricity demand through computational task transfer. Its power load exhibits spatial flexibility, which holds significant value for power system security and economic operation. However, data centers and power systems belong to different operating entities, making them unable to accept unified dispatch from the power grid. How to achieve coordinated decision-making between wide-area computing power dispatch and power system dispatch faces technical challenges. First, considering spatial computational task transfer costs and constraints, a coordinated optimal dispatch model between data center clusters and power systems is constructed. Then, a feasible region projection-based spatial flexibility characterization method for data center clusters is proposed. It reduces the dimensionality of the computational task dispatch model to the electricity consumption power space to explicitly characterize feasible boundaries and price sensitivity of spatial flexibility. Finally, a spatial flexibility-incorporated non-iterative computing power-electric power coordinated dispatch method is proposed. The optimality and privacy-preserving capabilities of the proposed method are proven. Case studies verify the effectiveness of the proposed method in reducing operating costs, mitigating grid congestion, promoting renewable energy accommodation, and improving coordinated optimal computational efficiency关键词：data center;computing power-electric power coordination;flexibility;coordinated dispatch;spatial flexibility;feasible region;privacy-preserving0|0|0更新时间：2026-10-10
- 摘要：Driven by the rapid growth in computing power demand due to artificial intelligence and digital economy, the backup energy storage capacity of uninterruptible power supply systems allocated in data centers is gradually increasing. However, its utilization rate shows a declining trend with the improvement in power grid supply reliability. The centralized backup mode for clusters formed by multiple adjacent data centers provides a new idea to address this contradiction. Nevertheless, the capacity allocation problem of shared centralized backup energy storage involves a trade-off between the power supply reliability of data centers and the economic efficiency of energy storage, lacking prior knowledge. Therefore, this paper proposes an optimal allocation method for centralized backup energy storage of data center clusters under a shared framework. Specifically, considering the probability and time statistical characteristics of the occurrence and repair of multi-type component composite faults in the power distribution system of data center as well as the strong uncertainty fluctuations of source and load, a reliability assessment framework based on sequential Monte Carlo is proposed. Furthermore, a local fault management strategy for data center clusters is designed to address the absence of power supply priority and the presence of loops in the grid structure. Based on this, a quantitative assessment of the power supply reliability of the data center cluster is achieved. Further, with the constraint that the data centers meet the minimum power supply reliability demand, a cost-minimization-oriented shared centralized backup energy storage allocation model is constructed to achieve quantitative optimization of energy storage capacity and charging/discharging power. The case study results indicate that after allocating shared centralized backup energy storage, the power supply reliability of the data center cluster is improved. Under the same reliability demand, the total allocation cost of energy storage in the centralized sharing scenario decreases by 75.1%, showing stronger economic efficiency and higher energy storage utilization compared with the mode where each data center independently allocating its backup energy storage.关键词：data center;computing power;shared energy storage;distribution network;reliability;energy storage planning0|0|0更新时间：2026-10-10
- 摘要：In response to the coexistence problems arising from the rapid growth of computing power scale, including surge increment of data center loads, the local power shortages and green power curtailment, and the practical problems of the insufficient incentive for data center operators to participate in scheduling, this paper proposes a three-stage migration model of multiple data centers for green power accommodation and computing power migration profit. In the first stage, following the priority scheduling sequence of green power, the maximum flow algorithm is adopted to realize the capacity matching between green power and the power consumption of transferable load. In the second stage, a greedy algorithm is employed for secondary scheduling to further enhance economic efficiency. In the third stage, migration paths are adjusted again to support the coordinated accommodation of local and inter-regional green power. Simulation results demonstrate that the proposed model can simultaneously improve the revenue of data center operators and the economic benefits of power grids, alleviate the power supply pressure of individual data centers, and guarantee the compliance of service level agreements, realizing the win-win collaborative optimization among computing power, electric power, and green power accommodation.关键词：data center;electric power-computing power coordination;green power accommodation;demand response;spatio-temporal collaboration0|0|0更新时间：2026-10-10
Computing Power-Electric Power Coordination
- 摘要：Accurate load forecasting is of great significance for the safe and stable operation of power grids. Existing deep learning forecasting methods suffer from poor cross-entity generalization ability and struggle to adapt to few-shot scenarios. Moreover, the forecasting methods incorporating pre-trained large models often overlook the importance of aligning information from different modalities and exhibit poor model interpretability. To address these issues, this paper proposes a power load forecasting method based on interpretable multi-modal enhancement. First, an instruction-finetuned large language model is used to output textual modal information from the raw load time series, achieving modal enhancement. Subsequently, a dual-channel encoding module processes this information to obtain multimodal embeddings. Second, this paper proposes a contrastive learning strategy based on partial information decomposition to achieve interpretable alignment of multi-modal representations. Finally, the forecasting results are output based on a Transformer architecture. Test results on public datasets from various real-world scenarios demonstrate that the proposed model exhibits competitive forecasting performance surpassing commonly used baseline methods and possesses certain interpretability, thereby providing guarantees for power load forecasting scenarios.关键词：load forecasting;large language model;multi-modal;interpretability;contrastive learning5|0|0更新时间：2026-10-10
- 摘要：Guided by price signals, residential flexible loads exhibit considerable demand response potential, but their response behaviors are subject to substantial uncertainty due to bounded rationality. To address this issue, a few-shot data-driven uncertainty modeling method is proposed. By exploring the time-varying coupling characteristics between residential flexible load response and its covariates, the method dynamically updates the parameters of conditional probability distributions, thereby effectively narrowing high-probability value intervals and improving the accuracy of uncertainty characterization. On this basis, a two-stage stochastic scheduling model for new distribution systems is developed, in which the conditional value at risk (CVaR) is adopted to quantify the tail risk, and the Benders decomposition algorithm is used to reduce the model solving complexity. Finally, numerical analysis is carried out on the modified IEEE 33-bus and IEEE 123-bus test systems. Results demonstrate that the proposed method can achieve a favorable balance among economic performance, risk aversion capability, and computational feasibility, thereby providing technical support for the coordinated optimization of “source-grid-load-storage” in new distribution systems considering the uncertainty of residential flexible load response.关键词：new distribution system;flexible load;demand response;uncertainty;stochastic scheduling0|0|0更新时间：2026-10-10
- 摘要：Under the background of high-penetration integration of distributed generators, the accuracy-efficiency trade-off in dynamic simulation modeling of active distribution networks (ADNs) is a current research focus. The widely used single-bus equivalent modeling method suffers from an inherent flaw of relatively high simulation errors in transient-steady-state analysis. Firstly, this paper analyzes the core sources of error in equivalent model simulation. Based on this, it proposes a topology-based delta-shape equivalent model (TBDSEM) for ADNs. This model connects three nodes via a triangular configuration, with each bus connected to a different type of equivalent element. Subsequently, this paper analyzes the response characteristics of dynamic element current changes at the point of common coupling (PCC) and proposes a hybrid transient-steady-state equivalent impedance calculation method tailored to the TBDSEM. Finally, model validation is performed using two case studies which are an improved IEEE 33-bus distribution system and a 3-generator 9-bus main grid interconnection system from a county-level distribution network in China. The results indicate that under multi-type disturbance scenarios, the computational time of the TBDSEM is on the same order of magnitude as that of traditional single-bus equivalent models, while transient-steady-state simulation errors are reduced by 10% to 50%.关键词：active distribution network (ADN);dynamic equivalent model;distributed generator;transient simulation;point of common coupling2|0|0更新时间：2026-10-10
- 摘要：Aiming at the challenges of insufficient economy and synergy of pipeline-network layout and equipment selection in energy stations in integrated energy system planning, from the engineering practice, based on the magnitude gap between the investment costs of energy stations and pipeline networks, a two-stage station-network collaborative planning method of “master-then-slave” is proposed. In the first stage, taking the siting and sizing of the energy station considering the equipment selection as the leading factor, ignoring the impact of common path, the point weight is reconstructed into a comprehensive function including equipment investment, operation and maintenance, energy purchase and carbon emission costs by introducing the integer variable of equipment selection, and an improved P-median model considering the “equipment-station-network” coupling is established. In the second stage, the focus is placed on addressing the issue of common path in the pipeline-network layout. A flow balance-based correction model for the pipeline-network layout is developed, which introduces topological constraints between paths and edges to avoid the duplicate accounting of construction costs for common paths and its impact on the pipeline-network layout. The case study results demonstrate that the proposed method realizes the collaborative planning of equipment selection and sizing, energy station siting and pipeline-network layout, and the economy of the planning scheme is significantly improved.关键词：integrated energy system;energy station;equipment selection;pipeline-network layout;collaborative planning0|0|0更新时间：2026-10-10
- 摘要：To address the issue that small- and medium-sized flexibility resources with regulable capacity in integrated energy systems are not effectively utilized, this paper establishes a user-side integrated energy system (UIES) model. Aiming at the uncertainty in UIES solution that may arise during the utilization of flexibility resources, a hierarchical robust reinforcement learning method is proposed to enhance the robustness of solution results. First, certain flexible reserve margins are reserved for regulable devices during model construction to satisfy the regulation requirements that may occur during different time periods. Then, a hierarchical robust reinforcement learning method is designed to handle the uncertainties in the UIES model. In this method, the upper layer is a reinforcement learning model, which is intended to handle long‑term actions affecting the future states of the system. The lower layer is a robust optimization model, which focuses on solving short-term actions that only affect the real-time operation of the system. The combination of reinforcement learning and robust optimization simplifies the complexity of the agent's decision-making process, while fully taking into account the robustness of solution outcomes. Case studies demonstrate that the proposed method can effectively mobilize user-side flexibility resources and facilitate the economic, low-carbon and stable operation of the system.关键词：user side;integrated energy system (IES);flexibility resource;uncertainty;reinforcement learning;robust optimization1|0|0更新时间：2026-10-10
- 摘要：In the environment of dense grid-connection with weak short circuit ratio of offshore wind power, there are problems such as insufficient safety margin of reactive power and voltage exceeding limits at the grid-connection point. To enhance the voltage safety of the collection area while considering the construction and operation economy, this paper incorporates discrete devices as a long-term reactive power loss compensation method, assists multi-agent deep reinforcement learning algorithms in real-time decision-making of continuous devices, and proposes a multi-timesacle collaborative regulation method of heterogeneous reactive power based on region relaxation control. Firstly, from the perspectives of operation safety, energy conservation, and loss reduction, two types of regions,reactive power static security region and optimization space, are defined in the reactive power injection space at the gateway. Rolling search methods for both regions are obtained through column and constraint generation algorithm and optimal power flow-point by point method. Furthermore, based on the nested relationship of regions, a discrete device relaxation control model considering reactive power losses of stations is proposed, and a multi-agent deep reinforcement learning algorithm with segmented reward functions for real-time control of continuous devices is designed to form an online collaborative regulation system. The simulation results indicate that discrete devices can assist continuous devices in completing the collaborative task of real-time voltage control and loss reduction operation under low output conditions through low-frequency actions.关键词：offshore wind power;reactive power;voltage control;collaborative regulation;security region;region relaxation control;multi-agent;deep reinforcement learning2|0|0更新时间：2026-10-10
-
DC Collection Line Protection for Offshore Wind Power Based on Frequency-band Energy Ratio Restraint
摘要：Offshore wind power DC collection systems with multiple collection lines and complex topologies face challenges in locating faults on DC collection lines. To address this issue, this paper proposes a single-ended protection method of DC collection lines for offshore wind power based on frequency-band energy ratio restraint. First, equivalent models for single-pole grounding faults and bipolar short-circuit faults are established to analyze the amplitude-frequency characteristics and energy distribution patterns of fault transient signals. Second, complementary ensemble empirical mode decomposition (CEEMD) is employed to perform time-frequency decomposition of fault transient signals. The ratio between dominant frequency energy and DC energy from the decomposed components is utilized to adaptively adjust the setting values of energy integral criterion, enabling faulty line selection. Additionally, the proposed protection method can serve as backup protection for DC busbar. PSCAD simulation results demonstrate that the proposed method achieves rapid and accurate faulty line identification with strong noise immunity and high tolerance to transition resistance.关键词：offshore wind farm;DC collection;ensemble;empirical mode decomposition;fault transient signal;line protection;single-ended protection;backup protection1|0|0更新时间：2026-10-10
Basic Research
- Postal code：310000
- Tel：13632102679
- Technical support is provided by Beijing Founder electronics co., LTD 京ICP备09064830号-19
京公网安备11010802024621 - It is recommended to read the content of this site in Chrome&IE9+. Please switch to extreme mode in browser 360.
- Cookies We use cookies to help provide and enhance our service and tailor content. By continuing, you agree to the use of cookies.
0