During low voltage ride-through (LVRT), multiple wind farms (WFs) are prone to transient synchronization instability due to power flow coupling and control interactions, which seriously threatens the security and reliability of grid-connected wind power operation. To address this issue, this paper investigates the transient synchronization stability of multiple WFs during LVRT from the perspective of active power imbalance caused by network impedance and synchronous equilibrium point existence. First, a dynamic grid-connected model of multiple WFs is established by considering the coupling effect of the common transmission line impedance, and its influence on q-axis terminal voltage dynamics is revealed. Then, based on the equilibrium point existence condition and PLL transient synchronization behavior, the impacts of line parameters and fault severity on synchronization stability are analyzed. Furthermore, a coordinated active power control strategy is proposed to regulate the d-axis active current reference of each WF, compensate for network power losses, and maintain stable synchronous equilibrium points during LVRT. The proposed strategy can significantly enhance the transient synchronization stability of multiple WFs under severe grid faults. Finally, the validity and physical feasibility of the proposed control strategy are verified through both MATLAB/Simulink and hardware-in-the-loop (HIL) experiments.
Intra-hour photovoltaic (PV) power forecasting is critical for reliable grid and market operations, but fast-moving clouds cause rapid irradiance changes that resist prediction. Existing image-based and multimodal models are brittle to camera and site heterogeneity and domain shift; late fusion weakens cross-modal interactions and site priors are underused, limiting generalization. To address this limitation, we propose PV-VLM (vision-language model), a parameter-efficient forecasting architecture built on three lightweight algorithmic designs. A vision projection alignment layer maps frozen VLM contrastive embeddings into regression-ready tokens, preserving macro-scale cloud structure without backbone fine-tuning. A dual-prompt conditioning mechanism couples learnable soft prompts with dynamically generated hard statistical prompts, conditioning the frozen LLM on both explicit real-time context and implicit dataset-specific priors. A temporal-query cross-modal attention module uses patchified power-history embeddings as Query to selectively retrieve relevant visual-textual cues from the LLM output, enforcing temporal selectivity over static concatenation. Two-stage fusion first aligns heterogeneous modalities via the LLM backbone, then fuses them with temporal dynamics through the query-driven attention. Across three horizons, PV-VLM reduces error relative to the strongest baseline by an average of 5.13% RMSE and 5.46% MAE; in cross-domain transfer it achieves average reductions of about 9% RMSE and 13.5% MAE with strong $R^{2}$ values. The framework trains 0.6% of parameters and achieves sub-second inference latency on consumer GPUs, enabling real-time edge deployment.
Growing penetration of inverter-based resources (IBRs) reduces effective grid inertia and complicates frequency regulation. This paper proposes a non-intrusive, end-to-end inertia estimator that integrates a differentiable short-time Fourier transform (STFT) front end with a compact 2D convolutional neural network. By learning the STFT window parameters directly from raw phasor measurement unit data, the model adaptively emphasizes inertia-sensitive frequency bands, while Grad-CAM provides spectral interpretability. On the IEEE 9-Bus system, the proposed method consistently outperforms CNN-, recurrent-, and Transformer-based baselines across multiple IBR penetrations and inertia levels, achieving MAPE at the level of 1–2% with consistently lower RMSE in both individual and mixed-regime evaluations. Under strong measurement noise (20–30 dB SNR), the adaptive window further improves robustness compared with fixed-window designs. Continuous tests demonstrate accurate tracking of piecewise-constant inertia, including values beyond the training range. Scalability is validated on the IEEE 39-Bus system using the same unified training protocol, where the estimator maintains high accuracy (RMSE ≈ 0.02–0.04) with low variance across operating conditions. The average inference latency is about 1.5 ms per 60-s data window, indicating suitability for real-time deployment. Field-PMU validation on real-grid measurements with SCADA derived kinetic-energy references further demonstrates practical observability beyond simulation. Overall, the results demonstrate accurate, noise-robust, computationally efficient, and interpretable online inertia estimation under varying IBR conditions.
Hydro-renewable dominant sending-end grids face unique seasonal stability challenges: inertia scarcity in dry seasons and water hammer effects in wet seasons. This paper proposes a comprehensive framework for risk assessment and coordinated control. A refined system frequency response (SFR) model is established to reveal the seasonally differentiated high inertia yet low-stability risk driven by hydro-turbine non minimum phase characteristics. A dual-boundary Inertia Security Margin Index (ISMI) is formulated to quantitatively decouple rate-dominant risks, represented by |RoCoF|max reaching 0.920 Hz/s in the dry season, from amplitude-dominant risks in the wet season. Based on ISMI, an adaptive coordinated strategy dynamically schedules wind and HVDC resources. Case studies show that the proposed strategy identifies nadir-dominant risks missed by a RoCoF-only trigger and raises the nominal nadir from 48.9244 Hz to 49.0302 Hz. It maintains the nadir above 49.0 Hz over the investigated 0.040–0.070 p.u. disturbance range. At 0.060 and 0.070 p.u., HVDC modulation further improves the nadir from 49.0233 Hz to 49.0317 Hz and from 49.0134 Hz to 49.0198 Hz, respectively. A secure nadir of 49.0232 Hz is also maintained under a 400 ms communication delay.
The increasing penetration of renewable energy has strengthened the coupling between power systems and weather conditions, making annual hourly joint wind-photovoltaic (PV) power scenario generation essential for adequacy assessment and flexibility planning. However, existing methods are constrained by limited training samples, weak physical consistency, and difficulties in modeling long sequences. To address these challenges, a hierarchical conditional flow matching framework (H-CFM) is proposed for annual hourly joint wind-PV power scenario generation. The framework formulates the task as a two-stage process, consisting of hierarchical meteorological sequence generation and deterministic physical mapping. A year-solar term-day-hour hierarchical structure is introduced, where solar ecliptic longitude defines a circular manifold prior to characterize solar-term phases and encode seasonal periodicity. An overlap conditioning mechanism with boundary constraints is developed to improve continuity across adjacent days. The proposed method is validated using ERA5 data from Inner Mongolia, China. Results show that H-CFM outperforms benchmark models in terms of probabilistic distribution, temporal characteristics, and extreme event representation, achieving improved reproduction of annual energy levels, seasonal resource patterns, and extreme scenarios. Ablation studies further demonstrate the effectiveness of the solar-term phase prior and the overlap conditioning mechanism in capturing seasonal structure and maintaining long-horizon continuity.
The high penetration of inverter-based resources (IBRs) creates weak grids with pronounced spatial frequency dynamics that conventional optimal power flow (OPF) and control schemes fail to address. This paper proposes a model-free, real-time control framework to coordinate grid-forming battery energy storage systems (GFM-BESSs) by integrating deep reinforcement learning (DRL), time-frequency analysis, and edge computing. A dual-pathway deep deterministic policy gradient (DDPG) architecture is developed: an actor network ensures OPF-feasible dispatch and virtual inertia tuning, while a critic network evaluates long-term spatial-frequency outcomes using wavelet-based scalograms. Enhanced by a hybrid CNN-LSTM backbone and a modality-separated dual replay buffer, the framework is deployed via edge computing for decentralized, low-latency execution. Validated on a modified IEEE distribution system, the proposed approach frequency. substantially outperforms baseline methods, achieving a 42.7% reduction in frequency nadir deviation, 58.3% lower rate of change of frequency (RoCoF), and 30.2% improved state-of-charge management. The framework offers a scalable paradigm for resilient, low-inertia power systems.
Hydrogen blending in natural gas networks forms hydrogen-enriched natural gas (HENG) systems, facilitating renewable energy integration and decarbonization. However, due to the complex dynamics of hydrogen-enriched natural gas, current dispatch strategies still struggle to manage hydrogenfraction exceedance risks. To bridge this gap, this paper proposes a constraint-learning-based framework for the optimal dispatch and risk management of integrated electricity and hydrogen-enriched natural gas (E-HENG) systems. Specifically, the methodology employs a data-driven surrogate to approximate hydrogen-fraction dynamics and exactly reformulates the inference process of the fixed, trained surrogate as mixed-integer linear constraints for integration into the optimization model. To enhance computational efficiency, a customized compact graph neural network (CGNN) is developed. By incorporating topological and operational correlations, the CGNN provides a physically structured surrogate for hydrogen fraction dynamics, while reducing the computational burden of the embedded optimization problem. Furthermore, a rapid risk-mitigation strategy is devised based on an adapted graph neural network (GNN) risk model to proactively manage transient hydrogen-fraction exceedance. This strategy quantifies transient exceedance risks and coordinates source-load adjustments within milliseconds, thereby supporting risk mitigation during the tested high-risk intervals. Case studies demonstrate favorable trade-offs among modeling accuracy, operating cost, and computational efficiency relative to the considered benchmarks.
Addressing the lack of inertia caused by the rapid rise in the penetration of offshore energy bases (OEBs), virtual inertia (VI) can be provided for frequency support via grid-forming technology. However, the provision of VI changes the converter's dynamic frequency-domain characteristics, resulting in severe oscillation risks. As a result, an economic dispatch strategy considering both frequency and oscillation stability is urgently required. Therefore, this paper develops a combined stability-economic dispatch model (CSED) embedded with primary-frequency-response-based frequency safety analysis and polynomial-chaos-expansion-Kriging (PCK) surrogate-based oscillation stability analysis. To fully characterize the synergies and uncertainties of offshore renewable energy, wind-solar wave complementary scenarios are joint-probability modeled by explicitly expressing the impact of sea conditions on components of the OEB. Building upon the synergistic scenarios, the CSED integrates VI support and its oscillation characteristics to strike an optimal trade-off between economic, frequency safety and oscillation stability. To address the computational burden caused by frequent simulations and analyses during the solution to the CSED, this paper adopts a non-intrusive PCK model that employs the Kriging residual to correct the polynomial chaos expansion model, thereby predicting the overall trend while compensating for intense oscillations through a small sample set. Experimental results show that the CSED reduces the comprehensive stability penalty by 35.66 at a 1.60 operating cost increase compared with conventional strategies, achieving an efficient optimization with the PCK.
Quasi-dynamic multi-energy flow (QMEF) calculation can accurately describe the operation states of integrated energy systems. However, the existing multi-stage holomorphic embedding method (MSHEM) struggles with the trade-off between the truncation order and effective time step of its semi-analytical solutions (SASs), which makes it difficult to balance accuracy and efficiency. Aiming at these challenges, this paper proposes a framework for selecting strategies for fast QMEF calculation of electricity and heat integrated energy systems (EH-IESs) based on an improved MSHEM. The improved MSHEM adaptively searches for both the effective time step and the truncation order using the proportional-integral control strategy, and automatically identifies the time and type of disturbance by the coefficients of the SASs. This ensures accurate QMEF calculation without the awareness of disturbance time. Furthermore, the proposed framework selects strategy from the improved MSHEM and its variants based on the average rate of change of variables to avoid unnecessary computations. The proposed methods have been applied to the E9-H32 IESs in Barry Island and an E64-H105 IESs from a northeast city in China. Experimental results demonstrate that these approaches achieve outstanding accuracy, robust convergence performance, and high computational efficiency.
The growing renewable energy penetration increases the risks of sub- and super-synchronous oscillations (sub/super SOs) in power systems, making accurate identification of oscillation frequencies critical to grid security. This paper proposes a continuous wavelet transform (CWT)-inspired deep learning (DL) method for dominant frequency identification in sub/super SOs. The designed sub/super-SO-oriented architecture leverages multi-scale representation and scale-selective feature aggregation, adaptively emphasizing informative temporal characteristics under diverse oscillation conditions. The proposed method offers several advantages over existing approaches. First, unlike conventional DL methods that employ generic network architectures, it adopts sub/super-SO-oriented structure inspired by CWT based analysis. Second, in comparison with the CWT and other related spectral analysis techniques, it demonstrates low sensitivity to parameter settings and robust performance even in the presence of data loss. Moreover, the dominant convolutional kernel scales exhibit adaptive switching behavior under complete, isolated-missing, and consecutive-missing data scenarios, thereby enhancing the model's effectiveness and credibility. Experimental validation confirms the method's superiority over existing algorithms. It also exhibits strong generalization to real-world data with grid structures, noise conditions, and oscillation causes that differ from their counterparts in the training set, highlighting its practical applicability.
Existing multi-agent deep reinforcement learning (MADRL) methods for local photovoltaic (PV) inverter voltage control face a critical dilemma: direct online learning risks safety violations, while simulator-based training requires accurate system parameters that are difficult to obtain in practice. This paper proposes Orchestra-SAC, an offline MADRL algorithm that learns policies solely from static historical operational data. The key challenge lies in achieving effective multi-agent coordination under offline settings, where agents must extract coordination patterns from static historical data based solely on local observations, without real-time interaction feedback. To address this, a conductor-musician framework is designed, where a conductor actor with global network observations guides musician actors with local PV measurements, while musician assistant networks enable the conductor to adapt to musicians' local observational capabilities and adjust guidance strategies for effective coordination. Combined with improved sampling mechanisms for offline learning, this method learns highly coordinated control strategies from limited historical data that adapt to stochastic PV and load fluctuations. Experiments on IEEE standard systems demonstrate that the proposed method significantly outperforms baselines, achieving a 94.72% optimality rate on the IEEE 33-bus system. Furthermore, the proposed framework exhibits strong generalizability, seamlessly extending to other mainstream off-policy algorithms such as MADDPG, MATD3 and MASAC.
Online tracking of inertia in AC microgrids with high renewable integration is a challenging yet critical task. However, traditional methods typically rely on known large disturbances, which are infrequent in actual microgrid operations, thereby limiting the applicability of online inertia estimation. To address this issue, we propose a Bayesian online inertia estimation framework utilizing ambient measurements. This framework enables continuous inertia tracking without relying on specific disturbances, substantially alleviating the limitations of conventional approaches. Moreover, to prevent significant estimation bias caused by the high noise level in ambient rate of change of frequency (RoCoF) measurements, the proposed framework extracts inertia directly from frequency measurements. By evaluating the Fisher Information and deriving the Cramér-Rao Lower Bound (CRLB), we theoretically prove that this frequency-based approach inherently yields a substantially lower estimation error bound compared to the RoCoF-based methods. Furthermore, a modified importance sampling method is introduced to accurately characterize the posterior distribution of inertia. Simulation studies on an AC microgrid with two diesel generators and a wind turbine generator demonstrate that the proposed method can efficiently and accurately track system inertia online across ambient conditions.
This paper proposes a novel data-knowledge hybrid-driven method for damping torque estimation in ultra-low frequency oscillations (ULFO), enabling accurate and robust identification of hydropower unit damping characteristics during early-stage ULFO, even with low-quality datasets. First, a multi band knowledge-guided block (MB-KGB) is proposed. In this block, the different frequency band information of the measurements is processed and reconstructed from a novel perspective, based on knowledge guidance. Specifically, for the ultra-low frequency (ULF) band, an improved dissipating energy flow (DEF) method is proposed as theoretical knowledge to guide the extraction of high information entropy features. Subsequently, a steady-state value distribution-based spatial temporal graph attention network (SD-STGAT) is proposed. Topology knowledge is embedded to capture the spreading patterns of ULFO in the grid. Meanwhile, steady-state distribution is incorporated into the attention mechanism to holistically adjust network structural parameters. Finally, a coalitional game theory-based model interpretation method is utilized. This method enables an interpretable analysis of the correlation between the proposed method and physical mechanisms, revealing new phenomena. Results demonstrate that the proposed method exhibits excellent accuracy and robustness against insufficient label quality. At the same time, the proposed method demonstrates better adaptability to non-ideal measurement conditions and diverse operating conditions.
This paper proposes an energy-carbon synergistic configuration approach for hydrogen-ammonia storage in renewable energy bases (REBs) with variable time-scale decoupling, aiming to enhance renewable energy utilization and power supply stability. Firstly, multi-dimensional performance trade-off metrics are designed from the dimensions of coordination, controllability and stability to comprehensively quantify the power supply stability of the REB, to provide a guidance for supporting the secure grid integration and efficient operation of large-scale REB. Secondly, a coordinated configuration model of hybrid hydrogen-ammonia energy storage considering the electric network topology of REB and ammonia transmission pipeline network is formulated to determine the optimal capacity and location of the storage and conversion equipment. Finally, to address the high-dimensional 8760h configuration problem, an asynchronous distributed parallel solution (ADPS) algorithm using accelerated conjugate gradient direction is developed to decompose it hierarchically in time into sub-problems with different optimization horizons. This algorithm can reduce time dimensionality while preserving temporal couplings, and enables asynchronous optimization of sub-intervals for efficient computation. Comparative case studies have validated the superiority of the proposed methodology on renewable utilization, power supply stability and carbon reduction performance of REB.
The increasing penetration of renewable energy significantly reduces system inertia and degrades frequency stability. Grid-forming wind farms (WFs) exhibit the capability to actively respond to frequency fluctuations, whereas the regulation capability is influenced by the time-varying nature of wind power, as well as by regional differences arising from the uneven spatial distribution of regulation resources. Different from thermal power units (TPUs), the frequency support characteristics of WFs vary dynamically with regulation parameter settings. To address the challenges, this paper proposes an optimization method for dispatching frequency regulation parameters in high-penetration wind power systems. An active support model for WFs is developed to delineate the participation boundaries for inertia response and primary frequency regulation. Meanwhile, a nodal frequency response model is established to represent spatial dynamics of system frequency. Furthermore, a robust disturbance scenario set is constructed based on nodal flexibility index, and the optimization is efficiently solved using Bayesian optimization with a Kriging surrogate model. Simulation results show that the proposed method minimizes the frequency regulation cost while ensuring operational security and fully exploiting the frequency support capability of WFs, increasing the frequency nadir from 49.5 Hz to 49.7 Hz and improving the frequency security margin by more than 10%.
The increasing penetration of renewable energy sources reduces system inertia and weakens primary frequency response, making frequency security a critical issue in long-term energy storage (ES) planning. However, most existing ES planning studies still treat ES primarily as an energy shifting and operational flexibility resource for reducing operating costs and promoting renewable accommodation, while insufficient attention has been paid to incorporating the transient frequency support capability of grid-forming (GFM) ES into security constraints at the long-term planning level. To bridge this gap, this paper proposes a frequency security constrained GFM ES planning framework for renewable power systems. First, the frequency support characteristics of GFM ES under virtual synchronous generator control are modeled, and three frequency security constraints are formulated using rate of change of frequency, frequency nadir, and quasi-steady-state frequency deviation. To improve tractability, the nonlinear constraints are transformed into linear forms through Taylor expansion and conservative approximation. Second, a robust planning model with a targeted extreme-scenario-based uncertainty set is developed to capture critical low-inertia operating conditions. Finally, a column-and-constraint generation algorithm is adopted for efficient solution. Case studies on a modified Garver 6-bus system, the IEEE 39-bus system, and a practical provincial system validate the effectiveness of the proposed framework.
Three-phase adaptive auto-reclosing is crucial for rapid power restoration after high-voltage AC line faults. However, in islanded wind farms integrated via modular multilevel converter-based high-voltage direct current (MMC HVDC) systems, a three-phase trip on the single-circuit AC outgoing line deprives the wind farm of voltage support, risking immediate disconnection. Furthermore, the rapid decay of line electromagnetic energy renders passive fault identification methods ineffective. This paper analyzes the wind farm's overvoltage and frequency instability mechanisms during grid loss operation, proposing a short-term uninterrupted operation strategy. Subsequently, an active signal injection method is proposed, utilizing the open-loop control of the sending-end MMC to generate a small-amplitude probing voltage. To analyze this injection process, the positive- and negative-sequence equivalent models of the MMC within the submodule capacitor discharge loop are derived to determine the safe probing voltage boundary and construct a transient fault identification criterion based on discharge current amplitude. Additionally, a coordinated strategy integrating the auto-reclosing process with the control of power electronic sources at both ends is developed. Simulations verify that the scheme prevents injection-induced overcurrents during severe faults while maintaining high sensitivity for high impedance faults, thereby facilitating rapid power restoration and significantly enhancing system reliability.
The rapid growth of renewable energy integration in modern power grids introduces significant challenges in resource planning, operational reliability, and cost optimization. Existing microgrid design approaches often rely on static or heuristic methods, which struggle to handle high-dimensional, uncertain energy data and extreme operating conditions. To address these gaps, this paper proposes a data-driven microgrid design framework that combines machine learning-based data profiling with a multi-layered hybrid optimization scheme. High-resolution energy data are preprocessed using a fusion of random, seasonal, and AI-based clustering techniques to enable adaptive partitioning into design and validation sets. The design phase employs a hybrid optimization algorithm integrating swarm intelligence, evolutionary strategies, and structured programming to simultaneously optimize cost, reliability, and uncertainty. A bifurcated architecture separates system sizing from validation, leveraging piecewise linear formulations and Bayesian reliability metrics to rigorously evaluate renewable energy sources (RES) and energy storage system (ESS) adequacy. Comparative analysis across five design mechanisms demonstrates that the proposed artificial intelligence-enhanced method achieves perfect reliability, minimal error indices, and superior resource allocation, outperforming conventional algorithms. The proposed system achieves minimal demand–supply energy imbalance ($RMSE = 10.58, MAPE = 1.65$) while exhibiting zero reliability violations ($MOIP = 0, MER = 0, MED = 0$, and $MMD = 1$). In addition, it maintains zero loss-of-power-supply probability (LPSP = LOLP = 0), negligible energy curtailment ($CR \approx 0$), and maximum energy utilization ($EUR \approx 1$), indicating strong scalability across high-dimensional datasets. Stress-testing under extreme operating scenarios, including cyclone-induced renewable scarcity and heatwave-driven demand surges, further confirms the robustness of the proposed framework. Even under severe generation and demand extremes, the system effectively suppresses loss probabilities and rapidly restores high energy utilization, demonstrating its suitability for resilient microgrid planning in climate-vulnerable and uncertainty-dominated environments.
High renewable penetration has led to declining system inertia and reduced frequency support, resulting in increased spatial heterogeneity of nodal frequency dynamics and emerging challenges to frequency security. However, how nodal frequency dynamics are influenced by renewable control characteristics, including virtual inertia and damping, remains insufficiently understood due to their complex interactions. Existing studies lack systematic methods to quantify their contributions, hindering the identification of vulnerable buses and the optimal allocation/tuning of renewable resources. This paper establishes a theoretical framework to fill this gap. First, the nodal frequency response is decomposed into a global component representing the system-wide smooth trend and a local component representing the bus-specific high-oscillatory behavior in time domain. Then, an envelope extraction method is proposed to capture and separate the core dynamics of the local component. Finally, an anlytical quantification approach is developed to trace the contribution of renewable resources to nodal frequency dynamics by successively linking the response to global and local components, system modes, and ultimately the underlying device control parameters. Several insights into frequency dynamics are also observed through theoretical analysis and case studies
