The output fluctuations and uncertainty of distributed renewable energy, combined with load fluctuations, make overloads of power lines occur more frequently in regional power grids. Power line overloading can be divided into forward heavy overload and reverse heavy overload. Flexible resources within regional power grids have great adjustment potential and can participate in line overloading regulation. However, how to effectively utilize flexible resources to eliminate the line overloading risk and how to improve solving efficiency when there are a large number of flexible resources need to be studied in depth. To solve the above problems, a coordinated rolling optimization method for multiple flexible resources is proposed, which takes flexible resource pre-selection into account. First, the impact of different types and locations of flexible resources on the overload of the specific line, including whether they can contribute, which type of overload they can contribute, and how much they can contribute, is analyzed. In particular, this paper mainly focuses on the relevant analyses in the loop network scenario. Based on this, this paper proposes a resource preselection method considering the adjustment characteristics of flexible resources to quickly select the most economical set of flexible resources that meets the overload demand. Finally, this paper proposes a coordinated rolling optimization method for multiple flexible resources, mitigating the impact of prediction errors on power grid operations through rolling optimization. The simulation results on a loop network system and the comparison with a rolling optimization method without preselection, validate that the proposed method can effectively resolve the power line overloading problem and enhance computational efficiency.

The contribution of utility-scale wind energy to primary frequency control significantly boosts overall grid reliability. The primary frequency control capability of wind turbines is subject to constraints dictated by fluctuating wind speeds and the set load-shedding coefficient. The optimal contribution of wind turbines to primary frequency control is dictated by the collective performance of the entire power system, demanding the development of adaptive control strategies. In this paper, a deep reinforcement learning-based dynamic strategy of wind turbine generators for primary frequency control is proposed. Taking advantage of the deep deterministic policy gradient (DDPG) algorithm to deal with the state and action of multidimensional continuous system, the dynamic control strategy considering the variable wind speed, load shedding coefficient and power grid state is realized. The effectiveness of the proposed approach was confirmed through simulation studies on the IEEE 39-bus system.

Compared with traditional resources, load-side resources are diverse, and their regulatory capacity is uncertain, making it difficult for them to participate in energy and peak regulation markets simultaneously. To address the challenge that the uncertainty of load-side resources cannot be accurately described by a known probability distribution, a day-ahead joint bidding strategy of electricity energy and peak regulation markets based on the distributional robust chance constraint (DRCC) and risk expectation is proposed. Firstly, a data-driven approach is employed to characterize the uncertainty in the adjustable capacity of load-side resources. An ambiguity set based on the Wasserstein distance is constructed, which does not require prior assumptions about the specific probability distribution of the underlying random variables. Then, the bidding strategy of load-side resource integrators is proposed to minimize the risk expectation. Finally, the effectiveness of the proposed model is assessed by case studies. The proposed method overcomes the problem that the robust model is too conservative, and its computational adaptability is better than that of the stochastic model, achieving a good balance between robustness and economy.

Nowadays, the integrated energy system (IES) faces significant security challenges from both natural occurrences and intentional disruptions, even minor disruptions can impact equipment operation and the system’s energy supply. To mitigate such disruptions in IES, a protective decision optimization model (PDOM) has been proposed based on complex network theory, which combines three protective measures: backup protection, physical protection, and new lines connection to minimize the weighted betweenness loss expectation. The result of the calculation, i.e., an optimized pre-disaster protection scheme, aims to determine which nodes and lines need to be protected (and different-type protection) within the specified budget. Furthermore, EvoRein-mixed integer linear programming (EvoRein-MILP) algorithm is presented to optimize the proposed model’s scheme. Simulation results demonstrate the effectiveness and feasibility of the obtained scheme, as verified by three validation indicators (system fragmentation, energy transmission efficiency, and load losing rate). Finally, compared to the genetic algorithm (GA)-MILP algorithm and adaptive GA-MILP algorithm, the EvoRein-MILP algorithm exhibits superior optimization and operation time performance under identical conditions.

Abstract：Accurate forecasting of wind power ramp events (WPREs) is essential for improving the flexibility and operational reliability of power system dispatch. However, the variation and complexity of ramp duration significantly affect the accuracy of ramp detection and prediction. In particular, bumps and time-varying noise can cause inaccurate identification of the start and end points of the duration, and variations in the duration length can lead to inaccurate predictions. Accordingly, this paper proposes a novel WPREs forecasting method combining a dual-constrained extended locking method for duration-preserving ramp detection and a Dilated Causal Convolutional Informer (DCAInformer) model for duration-adaptive learning. First, an improved dynamic adaptive L1 trend filtering approach is employed to suppress time-varying noise and outliers while retaining the main ramp trend changes. Then, a dual-constraint extended locking detection method is proposed to identify ramp events while avoiding the mis-segmentation of continuous ramps caused by bumps. Furthermore, DCAInformer model is constructed to better capture ramp features under different durations by introducing the dilated causal convolution into encoder of Informer, which can adapt the receptive field. Case studies using data from three wind farms in China demonstrate that the proposed method outperforms benchmark models in both ramp detection and forecasting accuracy.

Integrated energy systems (IESs) inherently involve multiple energy carriers and operational objectives, making the exploration of the Pareto frontier a central task in their optimal operation. However, existing multi-objective optimization methods for IESs usually require the double-layer iterative computation to obtain Pareto optimal solutions, where the upper layer determines the weights of the objectives and the lower layer performs optimization computations for the determined weights, resulting in a considerable computational burden. To address this challenge, this paper proposes a novel balance-supervised reinforcement learning framework that accelerates the Pareto optimization of IESs by integrating dynamic weights and agent learning. First, a novel balance supervisor agent is proposed to directly endow weights with learnable uncertainty via the Bayesian-based balance algorithm. Then, the balance-supervised framework can transform the double-layer computations into single-layer computations and enable the objectives to adapt to a theoretically infinite range of weight combinations, significantly reducing the computational burden and expanding the exploration of Pareto frontier. In addition, the Bayesian uncertainty layer is proposed to enhance the model’s adaptability to uncertainty. Case study shows that, compared to the state of the art methods, the proposed framework can effectively achieve optimal performance of the IESs in multi-objective scenarios.

Modern power grids are increasingly exposed to diverse cyber-physical attacks that originate in the cyber layer and disrupt physical operation. A critical defense challenge is to rapidly determine whether an attack has occurred and identify its type. To address this challenge, this paper aims to develop a generalized detection scheme capable of identifying multiple types of attacks, particularly unknown ones. It is first observed that, under normal operation, cyber-layer data and physical-layer states are consistent across layers and evolve smoothly over time, while attacks disrupt these consistencies. Building on this insight, a self-supervised Cross-Layer Cross-Temporal Contrastive Learning (CL-CTCL) framework is proposed, which employs a Graph Attention Network-Gated Recurrent Unit (GAT-GRU) encoder to capture spatial dependencies and temporal dynamics. Two complementary contrastive objectives are introduced: Variance-Invariance-Covariance Regularization (VICReg) for cross-layer alignment and Information Noise-Contrastive Estimation with Momentum Contrast (InfoNCE-MoCo) for temporal consistency. Residual deviations from these consistencies provide reliable attack fingerprints, enabling detection and classification of both known and unknown attacks. Experiments on an IEEE 118-bus cyber-physical testbed show that the proposed CL-CTCL achieves high accuracy on known attacks and generalizes well to unknown ones, outperforming baseline methods across standard metrics.

For day-ahead security constrained unit commitment (SCUC), pre-screening transmission constraints and retaining only the active ones in the optimization process can significantly improve computational efficiency. Most existing methods assume that similar system total load or net load directly leads to similar line flows and similar active/inactive statuses. However, even when total net loads are similar, the spatial distributions of the net loads may differ, which results in different power flows and the corresponding transmission constraints statuses. To alleviate this issue, this paper proposes a hybrid data-driven method for screening active transmission constraints by matching similar hourly hybrid estimated power flow (HEPF). The HEPF integrates a deterministic component derived from the product of forecasted nodal net loads and power transfer distribution factors (PTDFs) with a data-driven surrogate model designed to comprehensively quantify the impact of net loads spatial distributions and UC schedules on power flows. Furthermore, an asymmetric Euclidean distance K-nearest neighbors (AED-KNN) algorithm shortens the similarity distance between the Historical HEPF and Current HEPF when the magnitude of Historical HEPF minus Current HEPF is positive, which effectively reduces false negatives (FNs). Numerical results on the 118-bus system show that the proposed method achieves average recall above 99% and average time saving of 47.59% compared with the original SCUC. Additional results on a 500-bus realistic synthetic system show that the proposed method still maintains high screening performance and average time saving of over 60%, further verifying its scalability.

To address the trade-off between generalization and interpretability in thermal error compensation for Vertical Machining Centers (VMCs), this paper proposes a Dynamic Physics-Weighted Gaussian Process Regression (DPWGPR) framework. In terms of Artificial Intelligence contribution, a unified parametric physical model is embedded as a Bayesian prior to impose manifold constraints. To further enhance residual learning capabilities, a physics-guided composite kernel function is meticulously designed to capture multi-scale thermal fluctuations. Crucially, a novel adaptive weighting mechanism driven by Kullback-Leibler (KL) divergence is introduced to quantify the real-time discrepancy between physical priors and data posteriors, acting as a probabilistic switch to seamlessly transition between steady-state physical consistency and transient data-driven learning. Regarding the engineering application, the method is validated on a VH800 VMC under complex non-stationary conditions to predict multi-directional deformations, including spindle elongation, headstock torsion, and column bending. Unlike traditional deterministic models, the proposed framework outputs 95% confidence intervals to quantify epistemic uncertainty, enabling risk-aware decision-making. Experimental results demonstrate that DPW-GPR reduces the Root Mean Square Error (RMSE) of the z-axis prediction to 3.44 mu m, outperforming Support Vector Regression (SVR) and Back Propagation Neural Networks (BPNN) by over 50%. Notably, the model exhibits superior data efficiency, achieving high accuracy with only 10% sparse training samples, significantly surpassing SVR and BPNN trained on full datasets. The proposed approach provides a robust, risk-aware, and data-efficient solution for intelligent manufacturing.

To accommodate the diversity and time-varying characteristics of grid regulation demands and flexible resource, efficient aggregation mechanisms are required to support grid regulation services. Dynamic aggregation adaptively adjusts resource composition according to grid demands, overcoming the limitations of the traditional fixed resource composition in virtual power plants (VPPs). However, the computational complexity of the large-scale resource composition is a new bottleneck. This paper introduces a dynamic aggregation optimization framework based on submodularity, which includes resource selection and coordination. A lazy submodular optimization method is proposed in the resource selection process to enhance the efficiency of flexible resource aggregation. Using the diminishing-return property of submodular functions, the proposed method caches previous aggregation gains as upper bounds, significantly reducing redundant evaluations. The method greatly improves computational efficiency while maintaining optimal aggregation quality. Simulations highlight the efficacy of the proposed method in dynamic aggregation.

The integration of high proportions of distributed photovoltaic (PV) systems has led to voltage violations in distribution power grid. Using the remaining capacity of PV inverters for reactive power regulation is an efficient voltage control method. Existing studies typically address PV reactive power capabilities statically, failing to fully account for dynamic variations caused by solar radiation fluctuations. This paper proposes a data-driven probabilistic assessment and hierarchical reactive power optimization method. A spatiotemporal probabilistic prediction model is constructed to perform multi-quantile predictions of PV active power and node loads, and dynamically constructs reactive power support boundaries to assess PV reactive voltage support. A hierarchical optimization strategy is developed to coordinate the adjustment of various reactive power resources in the distribution power grid. Case study results using an improved IEEE 33-bus system show that the proposed method effectively mitigates voltage fluctuations and improves the efficiency of reactive power regulation.

Rapid, early, and accurate detection of cyber intrusions is increasingly vital for ensuring the security and operational resilience of modern power systems, as cyber-attacks grow in both frequency and sophistication. However, existing approaches primarily focus on either cyber or physical data independently, ignoring critical interactions across these two layers. To bridge this gap, this paper proposes a novel Cross-Layer Graph Attention-Transformer Network (CL-GTN), which simultaneously incorporates information from both the cyber and physical layers, and effectively integrates them through a learnable physics-informed point-to-bus projection mechanism. Specifically, the projection matrix constructed from power system sensitivity factors explicitly captures the fine-grained mapping from cyber data points in communication packets to affected physical grid nodes, providing a robust cross-layer fusion basis. Graph attention networks (GAT) are employed separately at the cyber and physical layers to extract relevant local and topological features, followed by a Transformer module that dynamically aggregates cross-layer temporal dependencies for accurate attack identification. Extensive experiments on multiple benchmark systems, including the IEEE 118-bus and large-scale ACTIVSg2000 systems, confirm that CL-GTN achieves high detection accuracy (over 0.98) and strong early-warning capability, successfully detecting up to 90% of attacks before execution with an average lead of 2.4 time steps. Comprehensive ablation and robustness studies further validate the model’s stability and component effectiveness. In addition, interpretability analyses show that the proposed projection-matrix mechanism accurately identifies critical cyber–physical coupling points, thereby enhancing situational awareness and defense capability.

Renewable power forecasting (RPF) is a classic time series forecasting (TSF) problem. During the modeling process, long-term dependencies in the time series can pose significant challenges, potentially reducing the accuracy of deep learning-based predictions. Nonetheless, the power data used for prediction often presents fine-grained patterns, such as second- or minute-level resolution, which further complicates the ability of DL models to accurately capture long-term dependencies. To address this problem, we use a uniform temporal sampling strategy to convert fine-grained data into several coarse-grained groups (CGGs), encompassing a reference CGG alongside its neighboring CGGs, each standardized with consistent resolution. Subsequently, we propose a novel CGG-based model (CGGNet) for multi-step RPF. In the proposed CGGNet, we design two modules: cross-group modeling within CGGs (CGG-cross) module and granular-based attention module. The CGG-cross module is used to build connections between the reference CGG and its neighboring CGGs; the granular-based attention module uses a novel CGG-dependent attention to obtain weighted representations of the reference CGG based on its neighboring CGGs. The proposed CGGNet is used to perform RPF on three datasets, including two wind power datasets and one solar power dataset. On the three datasets, the proposed model presents average improvements by 29.24% for root mean square error (RMSE), 32.03% for mean absolute error (MAE), 16.01% for coefficient of determination (R2). The results demonstrate that the proposed model can ensure a remarkable prediction response.

Learning an efficient feature representation from wind power data is crucial for improving the accuracy of wind power forecasting (WPF). Compared to supervised learning-based models, self-supervised representation learning models, such as contrastive learning (CL), demonstrate advantages in capturing feature representations from time series data, especially for data with strong fluctuations. Consequently, we propose a CL-based model to extract more accurate feature representations from wind power data. Unlike traditional CL-based models that rely on constructing both similar and non-similar pairs of data, the proposed model only requires similar data due to the sparsity of time series data. In addition, since the similarity in time series data is primarily reflected in the trend, especially for wind power data with significant fluctuations, the proposed model incorporates dynamic temporal granularity data (TGD) to generate similar data for a given wind power data, which can strengthen trend-consistent characteristics. Subsequently, the proposed model designs two distinct yet weighting-sharing networks, referred to as the online network and the target network, to model the two different augmented views with similar trend generated from the TGD, enabling the bootstrapping of point-wise feature representations from wind power data. To demonstrate the superiority of the proposed model, we compare it with state-of-the-art (SoTA) models on two wind power datasets, showing average enhancements of 7.34% for mean absolute error (MAE), 4.75% for root mean square error (RMSE), and 3.92% for R-squared (R2), respectively. The results show that the proposed framework can generate improved performance with accuracy and robustness.

Accurate prediction of tool wear is critical for optimizing machining efficiency, minimizing unplanned downtime, and enabling predictive maintenance in intelligent manufacturing systems. However, developing computationally efficient models with high predictive accuracy remains a persistent challenge. To address this, we propose a novel pyramid-structured Long Short-Term Memory (LSTM) network for multi-step tool wear forecasting. The architecture features a hierarchical cascade of LSTM modules operating at progressively coarser temporal resolutions, enabling the model to capture both short-term fluctuations and long-term degradation trends. We further conduct a systematic ablation study to evaluate the impact of architectural depth and scale configuration on prediction performance. Experimental results demonstrate that the optimal model achieves a mean absolute percentage error below 0.5% on the validation set, a coefficient of determination close to 1, a mean absolute error of approximately 0.0005 mm, and a mean squared error lower than 5 x 10-6 mm2, indicating excellent predictive performance. These results underscore the model's high precision and robustness. The proposed framework provides a reliable, data-driven solution for tool condition monitoring and supports optimal tool replacement scheduling in smart manufacturing environments.

With the rapid economic development across various regions of China and the explosive growth of renewable energy generation, reliance solely on the generation side has increasingly proven insufficient to meet peak shaving demands. This necessitates the exploration of peak shaving strategies that involve both the generation and load sides. To address this, the paper first summarizes the sequential clearing bidding model for generation and load sides and innovatively proposes a joint clearing bidding model for generation and load sides. Subsequently, the corresponding bidding process is derived, and a day-ahead economic dispatch model is established, aiming to minimize system costs. Finally, the paper conducts performance simulations and evaluations of the two clearing modes using the IEEE118-node system. Based on the simulation results, optimization recommendations are proposed, tailored to the unique characteristics of China's electricity market.

The collaborative analysis of both cyber-layer and physical-layer data is crucial for improving detection accuracy and timeliness of cyber-attack. Cyber-layer features provide early indicators of attacks, while physical-layer features reflect the actual impact on the power system. To leverage this synergy, a cross-attention mechanism is introduced to generate cross-layer features to capture these cross-layer interactions. Furthermore, based on the traditional Mixture of Experts (MoE), a novel framework MoE-Transformer Dual Layers Detection (MoE-TransDLD) is proposed, which dynamically fuses multi-layer features to model cyber-physical dependencies. Specially, MoE-TransDLD assigns a dedicated expert to each layer, including a cyber-layer expert, a physical-layer expert, and a cross-layer expert, to more accurately model multi-layer data relationships in power systems. Notably, both the expert network and the gating network share a common Transformer architecture to extract global features, while maintaining corresponding independent feed-forward network (FFN), where each expert focuses on its respective domain and the gating network achieves adaptive and dynamic selection in decision making. The synthetic Texas 2000-bus model system is used as an experimental model and its physical-layer data and cyberlayer data are collected. The experimental results show that the MoE-TransDLD significantly outperforms the existing methods and achieves superior classification metrics and faster attack detection time.

In the context of renewable energy (RE) integration, traditional economic dispatch methods demonstrate significant limitations in addressing operational risks arising from the uncertain nature of RE. This paper proposes a dispatching strategy for hydro-thermal-wind-solar-storage complementary systems (HTWSS-CS) that balances grid flexibility and economic efficiency. Firstly, drawing from flexibility supply-demand dynamics, system flexibility metrics are established. Then, using these metrics as a foundation, a comprehensive dispatch model for HTWSS-CS is constructed to simultaneously optimize grid flexibility and economic efficiency. Verification results using an improved IEEE 39-bus system validate that the approach proposed in this paper yields marked enhancements in grid flexibility while maintaining a moderate cost escalation, effectively balancing economic efficiency and flexible regulation performance.

Load forecasting is fundamental to the operation and planning of power systems. Its accuracy is crucial for ensuring system security and reliability, as well as for reducing generation costs and improving economic efficiency. Recent studies demonstrate that large language models (LLMs) exhibit powerful capabilities in pattern recognition and reasoning for complex token sequences. The critical challenge lies in effectively aligning temporal patterns in time-series data with linguistic structures in natural language to leverage these capabilities. This study introduces a time-series forecasting approach for electrical load prediction that builds upon a pre-trained GPT-2 model, with its self-attention and feed-forward layers kept frozen during the process. Fine-tuning is applied exclusively to the input embedding layer and output projection layer. Experimental results demonstrate that the proposed method achieves performance comparable to or superior against existing approaches across multiple electrical load forecasting tasks.