Accurate acquisition of the CO2 spatial distribution in coal-fired boilers is of great importance for combustion state perception, carbon emission assessment and low-carbon operation. However, direct measurement remains challenging due to limited sensor coverage and highly complex internal conditions. Although CFD can provide high-fidelity numerical analysis results, its high computational cost limits real-time applicability. Accordingly, a multi-level information fusion-based three-dimensional spatial field prediction method is proposed for real-time CO2 spatial fields’ inference, which integrates spatial mesh regularization, a lightweight main prediction network, and a large model (LM)-driven physical knowledge guidance module. Specifically, spatial mesh regularization is applied to regularize historical spatial fields and ensure structural and scale consistency. The lightweight main prediction network employs the enhanced 3D U-Net to perform feature-level fusion of historical CO2 fields and control parameter tensors, enabling efficient extraction of multi-scale spatial features. A Qwen-7B-based LM provides decision-level fusion through supervision and physical-knowledge-guided assessment of key regional behaviors. Furthermore, the comprehensive spatial field metrics are established to enable evaluation-level fusion of key regional statistics and spatial distributions, providing a quantitative assessment of field prediction performance. Experiments across different operating conditions show that the method improves the overall prediction accuracy by approximately 10% compared with GCN, enhances prediction stability in key regions by 3%-10%, and reduces computation time by 98.3% relative to CFD.

With the large-scale integration of renewable energy equipment, substation Operation and Maintenance (O&M) is characterized by the intricate interleaving of knowledge reasoning and data analysis tasks. Traditional O&M relies heavily on manual expertise, while Large Language Models (LLMs) face the dilemma of hallucinations, making it difficult to meet the stringent requirements for high reliability. To address these issues, this paper proposes Power Ops Agent, an O&M framework that identifies user intent through a Task Routing Node to coordinately schedule different sub-agents. For knowledge reasoning tasks, the Power Ops Reasoning Agent is designed to achieve provenance tracking and reasoning of textual and graphical regulations by leveraging Multimodal Retrieval-Augmented Generation (MRAG) and related technologies. For data analysis tasks, the Power Ops Analytical Agent is constructed to harness the code generation capabilities of LLMs, incorporating Shapley Additive Explanations values to generate interpretable reports with physical significance. Experimental results demonstrate that Power Ops Agent effectively resolves the challenges of accuracy and trustworthiness in O&M decision-making.

The increasing adoption of renewable energy sources (RES) in community-level residential buildings necessitates advanced microgrid (MG) systems capable of harmonizing distributed generation (DG), energy storage systems (ESS), and electric vehicle (EV) integration. While conventional energy management systems (EMS) demonstrate efficacy in grid stabilization and operational cost reduction, their ability to balance economic objectives with occupant comfort metrics in built environments remains suboptimal. PPO-EMS addresses the problem by leveraging Proximal Policy Optimization's clipped objective to (1) evade prior modeling and achieve stable, real-time convergence in dynamic, non-stationary environments; (2) centralize load scheduling across time-flexible appliances, power-adjustable devices, and critical loads for enhanced precision; and (3) integrate electric vehicles as distributed storage while mitigating range anxiety. Empirical results demonstrate that PPO-EMS outperforms other methods in rewards, stability, and scheduling efficiency. Specifically, PPO-EMS reduces total microgrid cost slightly compared to Deep Deterministic Policy Gradient (DDPG) and Deep Q-Network (DQN), and achieves a 31-58 % cost reduction relative to fixed scheduling strategies. Moreover, it attains the highest occupant comfort reward among all evaluated deep reinforcement learning approaches, demonstrating both economic and comfort advantages.

As energy transformation and technological advancements accelerate, multi-load forecasting serves as a key component in optimizing the planning, operation, and management of smart grids. Nevertheless, conventional load prediction techniques often suffer from issues such as limited accuracy and high volatility. To overcome these limitations, this study presents a load prediction approach based on Transformer multi-model fusion, employing a Stacking ensemble learning framework to integrate multiple models, including Transformer, XGBoost, and GDBT. The method also takes into account key influencing factors, such as meteorological conditions and holidays. Specifically, it involves training and predicting from the original features, followed by utilizing Transformer’s self-attention mechanism to capture feature relationships and long-term dependencies, ultimately leading to more precise load predictions. Empirical results validate that the proposed approach achieves superior prediction accuracy and stability across multiple datasets, demonstrating significant improvements over single-model approaches and conventional techniques.

In the realm of cloud computing, effective resource allocation can significantly enhance the energy efficiency of datacenters. Task scheduling and Virtual Machine Placement (VMP) are two pivotal aspects of resource allocation. However, in current research, they are often treated separately, overlooking the potential for integrated optimization. In this paper, we propose an integrated solution for task scheduling and VMP in energy-efficient datacenters, based on queueing theory and Deep Reinforcement Learning (DRL) methods. This novel and comprehensive approach provides an alternative perspective for resource scheduling strategies in datacenters. We construct a queueing theory model for task scheduling, aiming to minimize the number of VMs that need to be instantiated, while ensuring that Service Level Agreement (SLA) violation remains at a low level. Furthermore, we design a VMP algorithm based on DRL for real-time selection of Physical Hosts (PHs) for deploying VMs. Finally, we conduct a simulation evaluation using a small-scale datacenter. The experimental results demonstrate that our method consistently ensures a lower rate of SLA violation. Compared to existing algorithms, the DRL-based VMP algorithm enables a more balanced utilization of the various resources in the PHs and reduces the total power consumption of the datacenter by more than 10% on average.

The communication protocol is an important support to realize the communication between equipment and Internet. And it covers all aspects of the IoT (Internet of things) system. To address the security problem of forging or tampering of key data in traditional IoT protocols, this paper designs an improved MQTT (message queue telemetry transmission) protocol that uses blockchain technology to ensure the security of transmitted data in the process of data transmission. Because the information in the blockchain is not tamperable, which in turn ensures that data stored in brokers are not maliciously tampered with. Through simulation experiments, it is proved that this scheme is lightweight, efficient and easy to implement, which helps to protect the security of IoT data.

Power Internet of Things (IoT) is an important support for digital innovation service of power energy internet, covering all aspects of power system. Power IoT security defense system may have customer data information leakage during transmission because of the use of traditional means of isolation. This paper proposes a reliable transmission and application security architecture for power smart IoT based on energy interconnection, aiming to solve the reliable transmission and security authentication problems existing in power systems. The paper first analyzes the security risk of the grid wise IoT system, proposes an effective power IoT security transmission scheme, and evaluates the safe and reliable transmission of the grid wise IoT system. Then designed a safe and reliable transmission of the grid smart IoT system, to deal with the traditional power network transmission security and communication security problems. The final application is in the construction of the security transmission platform of the wisdom park of Shanxi Electric Power Company, which provides the corresponding security protection capability in the power IoT through the situational awareness security measures of each layer, and realizes the reliable transmission and security application of the source network load storage and other links in the power IoT environment.

传统云端电能质量扰动识别模式下,海量扰动数据对云端的服务器造成了较大的存储、计算压力,且云端扰动识别存在延迟,实时性较差;边缘侧扰动识别可以缓解云端压力,降低延迟,但边缘侧之间无法实现数据的跨域共享.针对以上问题,文章提出了基于联邦学习的边缘计算框架,首先,边缘侧使用本地数据训练模型,然后将模型参数上传至云端进行聚合,更新模型并下发至边缘侧进行部署,在边缘侧对电能质量扰动进行识别分类.实验结果表明,相比云端扰动识别模式,基于联邦学习的边缘侧扰动识别对云端的存储需求下降了97.58%,数据通信成本下降了53.68%,单次扰动识别的传输速率需求下降了99.994%,满足扰动识别实时性的要求;优化后的联邦学习算法与传统的联邦学习算法相比,扰动识别准确率提升了1.72%～3.64%.

In order to study the standard security access authentication mechanism of intelligent sensing terminals of massive power Internet of Things, In order to study the standard secure access authentication mechanism of intelligent sensing terminal of massive power Internet of Things, a new privacy protection method widely used in block chain is proposed to prove identity. The traditional power IoT cloud-side interaction security access MQTT protocol still has a lot of room for adaptation and optimization. First, the proposed non-interactive zero-knowledge proof identity authentication method reduces the time of traditional standard secure access authentication process; Second, it reduced the computing resources consumed in a large number of intelligent sensors access authentication. The comparison results show that, the access authentication time of this method is 30%∼50% less than that of the traditional secure access authentication process. The computing resources consumed during authentication are reduced by 20% to 30% compared with traditional security and secrecy mechanisms.

Third-party eavesdropping is a unsolved problem in the process of data transmission in the physical layer of IoT (Internet of Things) in Power Systems. The security encryption effect is affected by channel noise and the half-duplex nature of the wireless channel, which leads to low key consistency and key generation rate. To address this problem, a reliable solution for physical layer communication security is proposed in this paper. First, the solution improved the key consistency by dynamically adjusting the length of the training sequence during feature extraction; Second, using an iterative quantization method to quantify the RSS (Received Signal Strength) measurements to improve generation rate of the key. Finally, based on the short-time energy method for the extraction of wireless frame interval features, by monitoring the change of inter-frame interval features, we can quickly determine whether there is an eavesdropping device into the link. Simulation results show that the reciprocity of legitimate channels R (R will be explained in detail in the following) is improved by 0.1, the key generation rate is increased by about 70%, and the beacon frames are extracted from the wireless link with good results compared to the methods that do not use dynamic adjustment of the pilot signal during the channel probing phase. The result shows that this method can effectively prevents third-party eavesdropping, effectively improves the key consistency and generation rate, and effectively implements beacon frame detection.

利用BBR(Bottleneck bandwidth and RTT)算法虽可以实现在复杂网络中带宽的充分利用,但该算法对网络噪音所造成的丢包现象敏感,且该算法因存在协议内部不公平的问题而无法实现物联数据的实时高效获取.针对以上问题,提出了1种基于深度强化学习的起搏增益优化算法(Deep reinforcement learning of BBR,BBR-DRL).首先,通过获取数据传输的往返时延、发送窗口大小和网络带宽等环境参数来实时感知网络状态;然后,结合环境参数,利用起搏增益进行动态调整,使得BBR算法能够及时与外部动态网络环境进行交互,从而降低丢包敏感度、提高不同往返时延(Round-trip time,RTT)流之间的公平性.实验结果表明,与经典BBR算法相比,所提出的BBR-DRL算法协议内部的公平性提高到了98.2％,丢包敏感性明显降低.

传统的云端电能质量扰动识别方式下,海量分布式电能质量数据会给网络负载带来巨大压力.为降低云端识别延迟,采用边缘侧扰动识别的方式,但是边缘侧智能终端计算资源有限,无法部署大规模深度神经网络.文章针对边缘侧智能终端计算资源有限和扰动识别准确率降低的问题,提出了一种基于知识蒸馏的边缘侧电能质量扰动识别方法.首先将云端训练好的性能稳定但复杂度高的深度神经网络模型进行知识蒸馏,生成一个结构简洁且运算量小的模型;然后再将蒸馏后的优化模型下发并部署在配电物联网边缘侧,直接执行电能质量扰动分类识别计算.实验结果显示,相比于现有的知识蒸馏算法,经过本方法优化过的小模型准确率提高了 1.4%～2.46%.同时,与传统的云端识别方式比较,边缘侧扰动识别的数据传输速率需求降低了 99.993%.表明在边端计算资源有限的前提下,基于知识蒸馏的边缘侧电能质量扰动识别方法能够满足准确率和实时性的需求.

随着电网数字化转型程度不断加深,电力物联网的建设面临高接入、高并发、高交互的挑战.文章在现有电力物联网技术研究基础上,从终端接入、边缘计算和平台应用3个方面,对电力物联网关键技术展开研究综述.终端接入方面,介绍了负载均衡技术在电力物联网终端接入中的应用,提出一种用算法实现海量终端的灵活接入和智能感知的思路;边缘计算方面,讨论了边缘计算技术缓解数据传输压力实现电力物联网业务实时性的可能性,并详细阐述了边缘计算缓存技术发展现状;平台应用方面,研究了电力物联网平台应用层实现多业务快速集成的相关技术,分别从自适应系统、微服务体系架构和可扩展网络架构3个方面进行讨论.最后,总结电力物联网关键技术研究意义并对未来研究方向进行展望.

Environmental factors such as channel noise and hardware fingerprints affect the encryption effect of physical layer key generation techniques, resulting in low consistency of generated keys. Feature pre-processing is a common means of improving consistency of keys. However, most of the existing feature pre-processing algorithms improve key consistency by sacrificing key generation rate, which is not very usable. Therefore, it is proposed a physical layer key generation method based on SVD pre-processing. This method uses the SVD feature processing algorithm to pre-process the channel features extracted from both sides of the communication before quantization, in order to simultaneously improve key consistency and key generation rate. The simulation results show that when the channel SNR is greater than 10 dB, the BER of the SVD scheme is significantly lower compared to the scheme without pre-processing and the DCT and PCA pre-processing schemes; when the SNR is greater than 20 dB, the SVD scheme KGR can reach a level of 10bit/s, which is significantly higher than the other three schemes. The results show that this scheme can significantly increase the key generation rate while effectively improving key consistency.

A reasonable and efficient scheduling strategy does not only help ensure the safe and stable operation of battery energy storage system, but also extend the battery cycling life and reduce the system overall costs. In this paper, a novel rule-based dual planning strategy is proposed to achieve refined management for the hybrid battery energy storage system, including lead-acid battery storage system and lithium iron battery storage system. Aiming at different characteristics and SOC value of different batteries, the upper layer of the proposed scheduling strategy is designed to determine whether the hybrid battery energy storage system is enabled and grid-connected and in the lower layer, the battery scheduling controller will cooperate the operation of different batteries to realize shallow circulation of the lead-acid battery storage. Various simulations are conducted in the environment of Matlab/Simscape, and the simulation results have proved the feasibility and effectiveness of the proposed scheduling strategy.

随着电力物联网的建设,电力物联网平台面临海量异构终端接入的挑战以及数据爆发式增长的压力.文章面向电力物联网海量终端接入技术,探讨当前电力物联网终端接入在感知层、网络层、平台层和应用层面临的问题,并由此从信息感知、网络传输、实时计算和安全防护4个方面对终端接入技术展开研究综述.在感知层,针对电力物联网信息感知的关键作用,研究信息感知技术体系以及相关终端接入算法;在网络层,电力物联网产生大量数据流量,信息交互面临挑战,分别从软件定义网络、数据传输与流量优化2个方面对电力通信网络作技术探究;在平台层和应用层,针对当前电力物联网实时计算方面的问题,引出边缘计算技术缓解云端计算压力,并详细介绍当前计算卸载算法.最后,对终端接入的安全防护相关技术进行分析总结.

Due to the volatility and randomness of the photovoltaic power generation, it is difficult for traditional models to predict it accurately. To solve the problem, we established a model based on the self-attention mechanism and multi-task learning to predict the ultra-short-term photovoltaic power generation. First, we selected the data with the optimal timing length and input the data into the Encoder-Decoder network based on the self-attention. The validity of features extracted by the encoder was checked by the decoder. Then, we added a restriction to the middle layer of the Encoder-Decoder network to prevent the autoencoder from copying the input to the output mechanically. This condition is used to predict the photovoltaic power generation, so a multi-task learning model was established. Finally, to take full advantage of the features that are efficiently expressed and allow our main task, the prediction task, to learn some unique features autonomously, we proposed a step-by-step training method and have validated the effectiveness of this view in experiments. Through experimental contrast, it is found that compared with the Encoder-Decoder network based on CNN and LSTM, the performance of the proposed method has been increased by 14.82% and 8.09% respectively. The RMSE and MAE of the Encoder-Decoder model based on the self-attention mechanism using step-by-step training are 0.071 and 0.040 respectively.

Our present work allows efficient detection of COVID-19 from chest X-Rays at a level exceeding practicing radiologists. The algorithm uses the architecture EfficientNet extended and named K-EfficientNet. The K-EfficientNet is associated with progressive resizing, which resizes the images from 112×112 to 224×224 during the training process. By combining six publicly available and independent datasets, we get a large dataset named K-COVID containing 14,124 X-Rays images of patients affected by Pneumonia or COVID-19 and patient with Normal X-Ray images. The application of transfer learning on the ImageNet dataset and data augmentation allows us to achieve 97.3% accuracy and 100% sensitivity, and 100% Positive Predictive Value on COVID-19 detection.

The wind power industry continues to experience rapid growth worldwide. However, the fluctuations in wind speed and direction complicate the wind turbine control process and hinder the integration of wind power into the electrical grid. To maximize wind utilization, we propose to precisely measure the wind in a three-dimensional (3D) space, thus facilitating the process of wind turbine control. Natural wind is regarded as a 3D vector, whose direction and magnitude correspond to the wind's direction and speed. A semi-conical ultrasonic sensor array is proposed to simultaneously measure the wind speed and direction in a 3D space. As the ultrasonic signal transmitted between the sensors is influenced by the wind and environment noise, a Multiple Signal Classification algorithm is adopted to estimate the wind information from the received signal. The estimate's accuracy is evaluated in terms of root mean square error and mean absolute error. The robustness of the proposed method is evaluated by the type A evaluation of standard uncertainty under a varying signal-to-noise ratio. Simulation results validate the accuracy and anti-noise performance of the proposed method, whose estimated wind speed and direction errors converge to zero when the SNR is over 15 dB.