Abstract This study presents a scenario-based stochastic optimization framework for the co-planning of distributed PV and battery storage systems in distribution grids. To account for uncertainties in solar generation and load consumption, a scenario matrix is generated via the Heuristic Moment Matching (HMM) approach. This matrix is subsequently combined with deterministic power flow formulations, yielding a stochastic optimization model that determines the optimal placement and capacity of storage units, with the goal of maximizing the system’s net present value (NPV). Simulations on the IEEE-53-bus test system reveal that the proposed approach substantially enhances economic performance through effective peak–valley arbitrage enabled by energy storage. Compared to deterministic planning approaches, the stochastic model also enhances system reliability by substantially reducing bus voltage violations and unmet load rates under uncertain conditions.

Power line inspection, as a critical task of smart grid maintenance, has been empowered by Unmanned Aerial Vehicles (UAVs) and neural detectors for automation and intelligence. The inspection task requires detectors to identify diverse equipment in high-resolution aerial images rapidly and accurately. However, existing detectors are still insufficient to meet these requirements, primarily due to their inadequate dataset scale, high-resolution adaptability, and network efficiency. To address these challenges, this paper first collects 99202 inspection images with a resolution of 4939×2987 and annotates 52 categories, thereby establishing the Power Inspection Unify Dataset (PIUD) as the data foundation. Then, PIUD is further processed by the proposed Power Inspection Multi-Scale (PIMS) framework to PIUD-MS to support high-resolution training and inference. Finally, this paper develops PI-YOLO based on YOLOv11 with improved performance in terms of multi-scale capability, network structure, and receptive field. PI-YOLO is publicly released as an efficient detector for inspection or annotation applications, whose advanced performance is validated by experiments on COCO, PIUD-MS, and InsPLAD. Compared with the baseline YOLOv11 and other detectors, PI-YOLO achieves both performance improvement and latency reduction. For example, PI-YOLO-Large improves AP by 0.7 while reducing latency by 1ms on PIUD-MS compared to YOLOv11.

Gait analysis requires the integrated interpretation of plantar loading, three-dimensional ground reaction force, and kinematic information. Although recent reviews have discussed plantar pressure technologies, diabetic-foot monitoring, testing protocols, artificial-intelligence-based pressure analysis, or intelligent-shoe gait estimation, a few have integrated plantar pressure measurement, bending-induced artifacts, multisensor kinetic–kinematic fusion, and clinical applications into a unified framework. This review summarizes technological routes from both academic research and commercial systems in plantar pressure measurement, including non-wearable reference systems, wearable normal-force measurement systems, and wearable three-dimensional force measurement systems. Particular attention is given to bending-induced motion artifacts caused by sole or insole deformation, together with mitigation strategies. Recognizing that plantar pressure information alone is insufficient for comprehensive gait interpretation, this review further organizes multisensor plantar measurement systems from the perspectives of sensor types, hardware placement strategies, and data-fusion methods. Finally, the clinical translation of plantar measurement data is discussed for disease-risk identification, rehabilitation monitoring, and personalized intervention.

With the high penetration of distributed energy resources (DERs) and energy storage systems, distribution network operation faces dual challenges of increased nodal overloading risks and complex dynamic characteristics. This paper proposes a Hybrid Behavioral Cloning (HBC) algorithm integrating network reconfiguration, reactive power optimization and energy storage coordination. By leveraging a policy search-guided expert knowledge imitation mechanism, a decision-making model with both rapid response capability while meeting the system operational constraints. The proposed solution adopts a dual-layer policy architecture: the upper layer employs a discrete neural network based on expert strategies for topology optimization, while the lower layer utilizes Reinforcement Learning (RL) with correlated discrete actions to achieve multi-timescale control. The solution is assessed through simulation experiments using a modified IEEE 33-bus test network with a stochastic disturbance training environment. The numerical results demonstrate that the proposed solution can recover the load under failures with a success rate of 80.3

Abstract This paper proposes a multi-point distributed photovoltaic (PV) and energy storage coordinated control strategy based on the Safety-Constrained Multi-Agent Deep Deterministic Policy Gradient (SC-MADDPG) algorithm. The study designs voltage control and energy storage control methods within the SC-MADDPG framework, which integrates multi-agent reinforcement learning with virtual synchronous generator (VSG) technology. The approach aims to enhance voltage and frequency stability in distribution networks under high PV penetration. Experimental results demonstrate that SC-MADDPG significantly reduces voltage violations and shortens recovery time compared to conventional droop control and standard MADDPG methods, improving the adaptive capability and operational reliability of distributed PV and energy storage systems.

To address the increasing penetration of renewable energy sources (RES), battery energy storage systems (BESS) play an increasingly vital role in power systems, making it strategically important to investigate their bidding strategies for participation in electricity markets. However, traditional market models generally rely on perfectly rational agents and deterministic optimization, which limits their ability to capture the bounded rationality and risk preferences observed in real-world market behavior. To overcome these limitations, this paper proposes a novel multi-agent reinforcement learning (MARL) paradigm that integrates a risk-averse quantal response equilibrium (RQE) model into a multi-agent soft actor–critic (MASAC) architecture. The RQE formulation jointly embeds bounded rationality and risk aversion into decision-making, enabling agents to respond smoothly to uncertain rival strategies while maintaining robustness against adverse market conditions. The proposed solution is assessed based on the IEEE 30-bus system under multiple different operational scenarios. Numerical results demonstrate that the RQE-based MASAC method outperforms benchmark MARL algorithms in revenue stability, convergence speed, and robustness to uncertainty.

This work introduces a wind power fluctuation mitigation method based on real-time state of charge (SOC) feedback from a hybrid energy storage system. The approach combines SOC values of battery energy storage systems (BESS) and supercapacitors with a fuzzy control system to dynamically adjust the filtering coefficients of the first-order low-pass filter (FLF) algorithm. The SOC feedback-based strategy avoids deep charge-discharge cycles of BESS and supercapacitors, leading to improved system performance and safety. The proposed solution is assessed through simulation experiments and the numerical results demonstrated that the proposed solution can effectively smooth the wind power fluctuations at different time scales, i.e. 1-minute and 10-minute, adhering to the grid-connected power fluctuation constraints. In addition, the proposed solution is compared with the FLF-based methods without SOC feedback, it shows that the proposed solution outperforms in suppressing wind power fluctuations and optimizing energy storage operations. This work provides a promising solution for the integration of offshore wind power to the power grid, promoting stable and sustainable renewable energy generation, and paving the way for future advancements in power fluctuation mitigation.

In the decarbonization in the electricity sector, the emergence of prosumers in the medium/low voltage power distribution systems poses significant challenges for distribution system operators (DSOs), as bidirectional power exchange and carbon trading occur more frequently. This paper focuses the coordinated electricity and carbon management in a bi-level frame-work for DSOs and local distributed resources aggregators (DRAs). At the upper level, the optimal power flow (OPF) and carbon emission flow model (CEF) are solved to derive the locational marginal price (LMP) and nodal carbon intensity (NCI) for DRAs. At the lower level, an exponential-based pricing model is designed to enable peer-to-peer (P2P) carbon emission permit (CEP) trading among DRAs, which involves CEP demand-supply situation and emission assessment pressure. Chance constraints are formulated to address uncertainties in DRAs. Numerical results confirm the effectiveness of the proposed solution in coordinated electricity and carbon operation of multi-stakeholders in power distribution systems.

As an emerging type of power market entity, energy storage system can trade in both the electricity energy markets and ancillary service markets to generate revenue. With the continuous increase in new energy penetration rates, the market-oriented trading models for new energy are gradually improving, and their participation in power markets will inevitably affect the trading strategies of energy storage system. Therefore, this paper considers energy storage facilities as independent bidding entities and constructs a bi-level game model. The upper layer aims to maximize the revenue of the energy storage system and determine its bidding strategy; the lower layer realizes the joint clearing of the electricity energy market and frequency regulation ancillary service market, while considering the participation of new energy sources, e.g., wind and solar photovoltaic power in the joint market. Subsequently, the Karush-Kuhn-Tucker (KKT) conditions and strong duality theory are applied to transform the bi-level model. Finally, experiments based on a case study are conducted to assess the impact of renewable energy and its penetration proportion on the trading strategies of energy storage systems.

Accurate prediction of electricity consumption is a complex, nonlinear task, and traditional linear, fixed-parameter models often fail to capture dynamic demand patterns. To overcome these challenges, this study proposes an innovative hybrid forecasting framework that integrates Support Vector Regression (SVR) with six advanced meta-heuristic optimization algorithms—Salp Swarm Algorithm (SSA), Ant Lion Optimizer (ALO), Satin Bowerbird Optimizer (SBO), Particle Swarm Optimization (PSO), Artificial Bee Colony (ABC), and Grey Wolf Optimizer (GWO)—for intelligent hyperparameter tuning. This multi-algorithm optimization strategy enhances model adaptability and reduces dependence on a single heuristic. A K-fold cross-validation procedure was employed to ensure robustness and minimize overfitting, and a real-world electricity consumption dataset was used for evaluation using multiple statistical metrics. The results demonstrate that the proposed hybrid SVR models significantly improve forecasting accuracy compared with the baseline SVR, with SVR-SSA and SVR-ALO achieving the best performance. Specifically, the R2 values improved by up to 9.97%, and the MAPE values decreased by over 97%, confirming the effectiveness of combining SVR with meta-heuristic optimization. However, this study is limited to a fixed set of meta-heuristic algorithms and a single regional dataset. Future research will focus on extending the framework to include newer optimization techniques (e.g., MRFO, HHO, SMA) and ensemble learning approaches such as CatBoost and LightGBM, as well as testing generalizability across multi-regional and multi-seasonal datasets.

In coupled transportation-power networks, the intermittency of renewable generation and the charging demand uncertainties of Electric Vehicles (EVs) present major challenges, highlighting the importance of coordinated operation among multiple flexible and dispatchable resources. This paper developed a deep reinforcement learning based framework to coordinate the operation of photovoltaic (PV), energy storage units (ESUs) and EVs, considering the coupling interactions between the transportation and power networks. For PV inverters, a stability-constrained soft actor critic method grounded in the LaSalle invariance principle is extended for voltage control. For EVs, a two-stage scheduling scheme is proposed: a piecewise-linear user equilibrium traffic assignment model is formulated, after which a short-term revenue-oriented EV scheduling model is further developed as a mixed-integer linear program. For ESUs, a SoC-guided hierarchical control mechanism is established, with the upper layer forecasting target SoC trajectories via a hybrid model, and the lower layer embedding them into the reward function to guide ESUs. The proposed solution is validated on the Nguyen transportation network coupled with the IEEE 33-bus and the 141-bus network, respectively. The numerical results demonstrate the effectiveness of the proposed solution in maintaining grid stability and ensuring economic benefits for users.

The increasing volatility and uncertainty of renewable energy pose significant challenges to the economic dispatch of power systems. Deep reinforcement learning (DRL) has been applied to address dispatch problems under uncertainty. However, the neural network structures used in DRL are often treated as black boxes, limiting their interpretability. This paper investigates the mathematical interpretation of reinforcement learning (RL) in multi-step stochastic economic dispatch. It is shown that DRL and approximate dynamic programming (ADP) share a common mathematical foundation. Both DRL and ADP methods rely on the Bellman equation but estimate the action-value function through different approximation techniques. ADP adopts a linear function to approximate the action-value function, while DRL employs deep neural networks to capture nonlinear representations of the action-value function. This paper proves that DRL can be interpreted as a nonlinear extension of ADP using a deep learning framework, which reveals the mechanism of DRL solving multi-step stochastic optimization. The simulation experiments based on the IEEE 9-bus test system demonstrate the interpretability of DRL through the comparison with ADP.

The rapid development of data centers (DTCs) has led to substantially increased energy consumption and electricity bills. Thus, this paper aims to minimize the performance costs of geographically distributed DTCs. Firstly, an integrated electricity-heat system model for geographically-distributed DTCs is developed considering renewable energy sources (RESs) and waste heat recovery. Particularly, the differences of delay tolerances in computational tasks are fully considered in the developed model, which aligns with real-world DTCs. Further, to cope with the large-scale decision variables caused by the delay tolerance model, a RES and electricity price aware task assignment (REPTA) algorithm based three-stage energy dispatch strategy is presented, which accelerates the decision-making process. In stage I, an electricity-heat coordinated optimization (EHCO) model is constructed, which preliminarily determines the scheduling plan with the exclusion of delay-tolerant tasks. In stage II, the REPTA algorithm is designed to allocate delay-tolerant tasks according to the complementarities of electricity prices and RESs. Then in stage III, incorporating the task allocation results of stage II, the EHCO model is solved again to obtain the ultimate energy dispatch decisions. Finally, the proposed solution is assessed through comparative experiments based on the data from Parallel Workloads Archive, and the numerical results confirm its effectiveness in both environmental and economic indicators.

Cascade detectors are widely used to detect insulators’ self-blast faults from aerial images obtained by unmanned inspection systems, in which the RPN and classifier are adopted for insulator boundary extraction and fault identification. However, classifier performance can be significantly limited by the complex background and the insufficient samples. This paper proposes an augmentation method to promote the classifier based on prior knowledge and the existing RPN dataset. Firstly, a weakly-supervised framework is established to train an RRPN from the RPN dataset. Given the Oriented Bounding Box (OBB) predicted by RRPN, the extracted insulator can be combined with the background randomly to generate new samples. The proposed method is extensively assessed through experiments and the numerical results show that the RRPN can not only provide boundaries for augmentation but also replace RPN with the same detection capability. And the classifier performance can be steadily improved without any additional annotation, e.g., the F1 score of MobileNet-Small can be improved from 91.64

DCs are evolving from electricity-intensive computing facilities into flexible nodes within integrated energy systems, driven by the rapid growth of cloud services, artificial intelligence workloads, and decarbonization requirements. This review examines DCs from an integrated energy systems perspective, focusing on the coupled interactions among electricity consumption, computational workloads, and heat generation. The energy characteristics of key DC subsystems are analyzed to identify fundamental electrical thermal computational coupling mechanisms and sources of operational flexibility. Representative coordination frameworks are reviewed and classified according to economic, low-carbon, and resilience-oriented objectives, highlighting how workload elasticity, thermal inertia, energy storage, and communication infrastructure have been leveraged for coordinated operation. In addition, critical technical challenges limiting large-scale deployment are synthesized, including electricity-centric infrastructure planning, fragmented modeling and forecasting of energy data interactions, insufficient exploration of cross-sector energy reuse, and evaluation frameworks that fail to capture dynamic flexibility and system-level value. Based on these insights, future research directions are outlined to support coordinated electricity and heat computation operation and the development of sustainable and resilient DC infrastructures.

Electroluminescence(EL) imaging enables high-sensitivity detection of hidden electrical and structural defects in photovoltaic(PV) modules, revealing internal faults beyond the reach of visible-light or infrared imaging to support intelligent power plant maintenance. However, the limited number of defective samples presents a serious challenge for deep learning methods. Existing generation models can perform data augmentation, but suffer from domain shift between generated and real images, resulting in limited improvement in detection performance. To address this challenge, a novel semantic-aware data augmentation method with edge priors is proposed. First, the controllable LDM-based generation model is introduced to generate multiple types of defect images. Second, an edge-prior-guided defect recognition network is proposed, which achieves robust defect semantic extraction through the fusion of edge prior features with the backbone network. Finally, to address potential domain shift issues caused by generated data, a semantic-aware data augmentation framework is proposed, which leverages a pre-trained defect semantic estimator to guide the defect recognition process for effective data augmentation. Beyond increasing the number of negative samples, the framework captures defect-related information in the generated data, supplementing the semantic representation of rare defects and thus alleviating dataset imbalance. Extensive experiments across multiple deep neural networks demonstrate an average 7.5% improvement in accuracy compared with the baseline model, confirming the framework’s reliability for PV maintenance while revealing strong scalability and adaptability through its model-agnostic design and robustness to diverse imaging conditions.

The optimal power flow (OPF) is considered essential for the planning and operation of power distribution networks. However, with the increasing complexity of active distribution networks (ADN), conventional optimization methods for OPF encounter significant limitations. Currently, the rapid emergence and maturation of large language models (LLMs), e.g., GPT-4 and DeepSeek, have introduced new opportunities. Therefore, this paper presents the Evolutionary Algorithm Generator (EAG), an intelligent heuristic framework that integrates large language models with Evolutionary Computation (EC) to efficiently solve the ADN-OPF problem. EAG consists of a generation module and an evolutionary module. By leveraging the generalization ability and automatic code generation of LLMs, EAG automates the entire process from algorithm design to optimization, enhancing both efficiency and performance. The effects of different models and prompt strategies on ADN-OPF performance were evaluated through extensive experiments across diverse test cases. The numerical results based on the IEEE 33 bus system demonstrate superior computational performance compared to traditional intelligent algorithms. This research offers a novel perspective on the intelligent design of optimization algorithms for power systems and represents a significant application of LLMs in the field of electrical engineering.