The integration of inverter-based microgrids (MGs) often induces variations in system characteristics and disturbances in the network. In islanded operation, voltage and frequency oscillations arise due to the lack of synchronization with the main grid. Ensuring reliable operation under such conditions requires advanced control techniques and proactive monitoring. This paper proposes an intelligent multi-agent secondary control architecture based on the Deep Deterministic Policy Gradient (DDPG) algorithm, where reinforcement learning (RL) agents dynamically adjust the parameters of proportional–integral (PI) controllers to mitigate disturbances and adapt to system uncertainties and varying operating conditions. The proposed framework employs a two-layer hierarchical structure. At the lower layer, two independent RL agents regulate frequency and voltage by tuning PI controller gains in real-time, ensuring stable operation under fluctuating loads and distributed generation. The upper layer incorporates a supervisory agent that monitors harmonic distortion and computes corrective signals sent to the droop controller, enabling coordinated adaptation and enhancing dynamic stability by reducing transient oscillations. As a proof-of-concept, the proposed multi-agent framework is implemented on a single DG unit within a four-converter microgrid, demonstrating the potential for improved dynamic performance under realistic operational scenarios. Validation in MATLAB/Simulink demonstrates substantial improvements in system performance. Frequency settling time decreased from 0.19 s (at Kp=0.3) to 0.10 s (at Kp=0.1), and voltage settling time decreased from 0.20 s (at Kp=0.3) to 0.13 s (at Kp=0.1). Total harmonic distortion (THD) significantly reduced from 2.07% (at Kp=0.3) and 2.03% (at Kp=0.2) to 0.68% (at Kp=0.1). These results highlight the scalability and effectiveness of the proposed multi-agent DDPG-based secondary control, offering a reliable and intelligent solution for advanced MG energy management under realistic operating conditions.
In the circumstance of multiple DC feed-in and high penetration of renewable energy, there is a contradiction between the continuously increasing uncertainty of power disturbances and the Automatic Generation Control (AGC) pause threshold value set based on fixed empirical values. This contradiction frequently causes Area Control Error (ACE) to exceed the pause threshold, triggering the AGC system to enter the PAUSE state. As a result, power plant output cannot respond to the dynamic demands of the power system, affecting the stability of tie-line power and system frequency. To address this problem, a dynamic pause threshold tuning strategy is proposed in this paper that considers the state switching characteristics of the AGC master station. Firstly, an interconnected AGC model incorporating an operational state determination module is established. Secondly, the concept of maximum ideal area control error is introduced, and based on its estimated values under different disturbances, the pause threshold constraints are obtained, and an adaptive tuning algorithm for dynamic pause thresholds and active support power is developed. Finally, simulation cases are conducted on a two-region interconnected power system under the Flat Tie-line load Control (FTC) mode and Tie-line Bias frequency Control (TBC) mode. The simulation results indicate that the proposed strategy effectively improves the efficiency and stability of system frequency and tie-line power regulation, significantly reducing the probability of the AGC system entering the PAUSE state.
The growing integration of distributed energy resources, particularly distributed photovoltaic systems, has placed increasing emphasis on enhancing the hosting capacity of distribution networks. Traditional optimization models or market mechanisms often face challenges from distributed energy resources stakeholders with diverse ownership or fail to realize budget-balance among distributed energy resources. To address this, this paper proposes a fair coordinated energy and voltage regulation market mechanism to improve hosting capacity of distribution system. First, a coordinated active and reactive power optimization model is developed, then in order to achieve a fair allocation of responsibilities and benefits of voltage regulation, the obligated reactive power output of distributed energy resources node is defined. Furthermore, a budget-balanced Arrow–d’Aspremont–Gérard–Varet mechanism, based on the predefined obligations, is employed to determine transfer payments among participants, ensuring that the distribution system operator bears no additional subsidy cost, with losses and profits balanced internally. Simulation results verify that the proposed scheme effectively maintains voltage quality, enhances both operational and planning-level hosting capacity, and achieves fair allocation of responsibilities and profits among stakeholders.
With the rapid expansion of grid-connected wind power systems, accurate and reliable wind power forecasting has become increasingly essential for the secure and efficient operation of power systems. To address the challenges associated with data noise reduction, feature extraction, and uncertainty estimation, a novel wind power forecasting framework integrating deep reinforcement learning and multi-objective Bayesian optimization (DRLMBO) is proposed. An improved variational mode decomposition algorithm optimized by the grey wolf optimizer is employed for data denoising, effectively overcoming the limitations of subjective parameter selection in traditional decomposition methods. Subsequently, the temporal convolutional network, Transformer, and bidirectional long short-term memory network are integrated to comprehensively extract sequence features, ensuring that local dependencies, long-term temporal patterns, and bidirectional information are simultaneously captured. Multi-objective Bayesian optimization is employed to obtain Pareto-optimal solutions, while quantile regression is incorporated for interval forecasting, thereby systematically enhancing the overall forecasting capability of the proposed model. Furthermore, a Double Deep Q-Network based reinforcement learning strategy is introduced to adaptively assign dynamic weights to multiple forecasting outputs. By learning optimal aggregation policies, the proposed strategy enables more flexible model fusion and significantly enhances forecasting stability and robustness. Comprehensive experiments conducted on multiple real-world datasets demonstrate that DRLMBO achieves superior forecasting accuracy and overall robustness compared with state-of-the-art baseline methods.
The introduction of current limiting control (CLC) enables grid-forming voltage source converters (GFM-VSCs) to operate under a mode-switching architecture between constant voltage control (CVC) and CLC modes. However, the transient stability margin of such a system has not been clearly characterized. At the same time, the time-varying nature of grid strength, resulting from dynamic changes in grid structure and operating conditions, imposes higher demands on the adaptability of mode-switching GFM-VSCs. Existing studies have predominantly focused on single parameters or specific scenarios, lacking systematic analysis of the control mode-switching mechanism and transient stability under varying grid strengths and fault depths. To address this issue, this paper first establishes a dynamic model of a grid-connected GFM-VSC system incorporating multiple control modes, accurately capturing the dynamic switching mechanisms and interaction characteristics between modes. Furthermore, a multivariate transient energy function considering power-loop coupling is constructed to provide a unified characterization of the transient energy boundaries under both CVC and CLC modes. Based on this framework, the effects of grid strength and fault depth on transient stability margins and mode-switching boundaries are quantitatively analyzed, and the dominant control parameters under different modes are further investigated from the perspectives of stability-margin variation, sensitivity distribution, and coupling characteristics. On this basis, a fuzzy logic-based adaptive transient control strategy is proposed. By monitoring the grid state in real time, the strategy dynamically adjusts the control parameters according to grid strength and fault depth, thereby improving transient stability and achieving smoother mode switching under different operating conditions. Finally, simulation results verify the effectiveness of the proposed strategy.
The increasing penetration of renewable energy sources makes it increasingly important to coordinate secure system operation with large-scale demand-side flexibility. To address this problem, this paper studies security-constrained unit commitment considering the customer directrix load for large-scale resource aggregation, which is important for improving renewable energy accommodation while maintaining operational security and economy. A coordinated bi-level framework is proposed in which the system operator determines the day-ahead scheduling strategy and the target load guidance profile under security constraints, while load aggregators respond to the published guidance signal under uncertain response behavior. On this basis, a two-stage incentive allocation mechanism is developed to allocate subsidies according to the actual contribution of each load aggregator and to encourage response behavior that is more supportive of renewable energy accommodation. Case studies show that the proposed method improves renewable energy utilization, reduces operating costs, and enhances the coordination between secure scheduling and large-scale flexible demand resources. The results also indicate that the proposed incentive mechanism improves the rationality and fairness of subsidy allocation. These findings demonstrate that the proposed framework provides an effective and practical approach for coordinating secure operation and large-scale flexible demand in renewable-rich power systems.
The increasing integration of renewable energy poses significant challenges for balancing electricity and hydrogen demand in integrated electricity-hydrogen systems (IEHS). This paper proposes a multi-timescale scheduling framework considering hybrid energy storage coordination and electrolyzer (ELZ) dynamics. The framework integrates seasonal, week-ahead, and day-ahead layers to coordinate high-pressure hydrogen storage for cross-seasonal shifting, low-pressure storage for short-term fluctuation mitigation, and battery storage for real-time balancing. A rolling horizon control strategy is embedded across all layers to continuously update decisions with new forecast information. A seasonal optimization model with Ward-clustering-based representative scenarios captures long-term supply–demand mismatches with reduced computational burden. For short-term operation, a dynamic ELZ model with start-up/shut-down behavior and hydrogen-to-oxygen safety constraints is developed, and nonlinearities are linearized via adaptive McCormick relaxation. Results show that the proposed framework improves renewable utilization, reduces grid dependence, and ensures safe and flexible operation across timescales.
In the context of power system decarbonization, energy storage has become a core flexibility resource. However, its generally low utilization rate has attracted increasing attention in the power sector. To improve the utilization efficiency of energy storage resources, a new business model named Cloud Energy Storage (CES) is proposed. This model enables multiple users to share grid-scale energy storage resources. In this paper, an optimal scheduling model for the CES operator is proposed, and various energy storage facilities are centrally controlled by the CES operator to satisfy the multitype demands from different users. First, a universal energy storage model is developed to reflect the key scheduling characteristics of multitype energy storage, including dynamic efficiency performance and degradation characteristics. Then, the proposed model is integrated into a scheduling model across multiple time scales for the CES system. Three representative energy storage technologies are considered in the scheduling model, including lithium-ion batteries, vanadium redox flow batteries, and adiabatic compressed air energy storage. Furthermore, the scheduling model is used to jointly optimize these energy storage facilities to address the demands of peak shaving, curtailed wind and photovoltaic power recovery, forecast error correction, and frequency regulation. Finally, a case study is carried out based on operational data and electricity market data from Western Inner Mongolia, China. The results show that the proposed CES scheduling method effectively mobilizes diversified energy storage resources to meet multitype demands and achieves day ahead and real time coordination across multiple time scales.
As cascade hydropower stations increasingly participate in energy and ancillary service markets, operators must determine joint bidding and scheduling strategies while accounting for strong hydraulic-electric coupling and the physical impacts of reserve deployment. Joint participation across multiple markets directly links market decisions with cascade dispatch feasibility, and actual reserve activation can substantially reshape short-term scheduling coordination, posing new challenges under market price uncertainty. To address this challenge, this paper proposes a short-term multi-market joint scheduling framework for cascade hydropower stations that explicitly incorporates reserve deployment scenarios and coupled hydraulic-electric constraints. The framework jointly considers medium- and long-term contract decomposition, day-ahead energy bidding, and spinning and regulating reserve markets. A non-probabilistic information-gap decision theory (IGDT) approach is employed to characterize price uncertainty in both energy and reserve markets. Accordingly, risk-averse and opportunity-seeking bilevel decision models are formulated to describe robustness-profit trade-offs without assuming price probability distributions, and the bilevel models are further reformulated into equivalent single-level mixed-integer linear programs. Finally, the proposed framework is applied to the cascade hydropower stations in the Lancang River Basin, China. Numerical results demonstrate that, compared with deterministic and stochastic benchmark models, the IGDT-based framework maintains feasible joint scheduling under reserve deployment uncertainty while capturing explicit multi-market risk-profit trade-offs. Moreover, the results indicate that reserve deployment probabilities significantly reshape cascade dispatch coordination and bidding behavior, highlighting the importance of explicitly modeling deployment effects.
The increasing integration of inverter-based resources into modern power systems has substantially reduced the overall rotational inertia, posing critical challenges to grid frequency stability. The AC-excitation synchronous condenser, which employs a doubly-fed induction machine with a high-inertia rotor and a three-phase inverter for excitation, has emerged as a promising device for reactive power compensation and voltage support. However, its conventional grid-following control lacks the capability to provide inertial response during frequency disturbances. To address this limitation, this paper proposes a virtual inertia injection strategy based on dual frequency-locked loops with distinct response dynamics. The fast loop rapidly tracks the grid frequency to ensure prompt inertial response, while the slow loop, incorporating a second-order low-pass filter, emulates the rotor dynamics of conventional synchronous condensers. This architecture eliminates the noise amplification and response delays inherent in derivative-based virtual inertia methods, while enabling independent adjustment of the virtual inertia constant and damping coefficient through explicit parameter mapping. The mathematical equivalence between the proposed dual-loop structure and the swing equation of synchronous condensers is rigorously established under both over-damped and under-damped conditions. Simulation results demonstrate that the proposed strategy achieves an inertial response settling time of approximately 35 ms, which is 24% faster than the conventional derivative-based method, and maintains stable operation during severe grid faults including three-phase voltage sags and phase jumps. Experimental validation on a 5.5 kW prototype demonstrates that the proposed strategy achieves a rise time of 120 ms with oscillation-free power output, representing a 68% improvement over the virtual synchronous machine method, while maintaining robust operation under harmonic-distorted grid conditions.
Artificial intelligence data centers (AIDCs) provide operational flexibility through training task time-shifting, spatial migration of inference workloads, and GPU frequency scaling, facilitating renewable energy integration and load regulation. However, these actions change IT-side heat generation, while liquid-cooling thermal inertia, heat-removal limits, and cooling power ramping constraints restrict practically available computing flexibility. Ignoring these thermal dynamics in energy storage system (ESS) planning and operation can result in a severe overestimation of AIDC flexibility and suboptimal ESS sizing. To address these limitations, we propose a thermal-aware coordinated ESS planning and operational scheduling method for liquid-cooled AIDC clusters. A dual-timescale model predictive control framework is developed to coordinate AIDC operation within temperature-state-dependent flexibility boundaries. The long-timescale layer optimizes training task allocation, whereas the short-timescale layer adjusts inference migration, GPU frequency, cooling power, and ESS operation using traffic and thermal feedback. Furthermore, a multi-objective ESS planning model is formulated to balance cost, service reliability, and low-carbon operation. Case studies show that, compared with the no-regulation operating benchmark, the proposed method reduces purchased electricity by 32.0% and decreases the service violation risk from 6.83% to 0.04%. It keeps the maximum temperature within the 50 °C safety limit, averting the severe 70.11 °C overheating under thermal-blind scheduling. In addition, coordinated computing-side flexibility reduces total ESS capacity from 8.74 MWh in the rigid-planning benchmark to 4.17 MWh. Results demonstrate that incorporating liquid-cooling thermal dynamics enables a more reliable assessment of AIDC flexibility and supports ESS sizing while maintaining Quality of Service (QoS) and thermal safety.
To address the strong conservatism of conventional N-1 security constraints, the insufficient response to minute-scale fluctuations, and the heavy computational burden of rolling optimization in power systems with high wind power penetration, this paper proposes a two-stage coordinated scheduling framework considering wind power uncertainty. In the day-ahead stage, a distributionally robust security-constrained unit commitment model is developed to account for wind power uncertainty. Bilateral chance constraints are introduced to relax the conventional rigid N-1 line-flow limits. They extend the deterministic security boundaries into probabilistic constraints with adjustable risk levels. Conditional value-at-risk (CVaR) is further employed to derive a convex reformulation of these constraints. The resulting model is solved using a column-and-constraint generation (C&CG) algorithm, thereby enabling a flexible trade-off between system security and operational economy. In the intraday stage, a Bi-LSTM-driven event-triggered model predictive control (ET-MPC) rolling optimization framework is constructed. By updating minute-scale wind power and load forecasts online and using forecast deviations as triggers, the framework links forecasting with dispatch and reduces the computational burden. Finally, case studies based on a modified IEEE 39-bus system demonstrate that, under a slight relaxation of the day-ahead security margin, the proposed method reduces the total operating cost by 1.97% and wind curtailment by 40.55% compared with the conventional N-1 checking strategy. In the intraday stage, different triggering thresholds reduce the computation time by 22.75% and 42.31%, respectively. The resulting dispatch decisions remain effective. The results verify the effectiveness of the proposed two-stage scheduling framework in improving wind power accommodation and economic operation.
The increasing renewable energy integration in active distribution networks challenges power dispatch due to heightened uncertainties. While deep reinforcement learning (DRL) shows promise for stochastic sequential decision-making, its practical implementation faces three critical issues: implementation feasibility concerns, subjective reward function design through manual tuning, and limited human–machine interaction affecting interpretability. To address these, this paper proposes a novel hybrid augmented intelligence (HAI) based stochastic dynamic power dispatch by establishing a reinforcement learning from human feedback (RLHF)-inspired framework that combines demonstration-guided policy initialization, preference-based reward modeling, and downstream policy optimization. Specifically, HAI first establishes a behavior cloning (BC) agent based on supervised learning to learn the strategy from demonstration data. Second, HAI introduces demonstration-guided early-stage training, in which the BC policy provides online actions while the DRL actor is trained in the background. Third, HAI automatically generates a set of trajectories with heterogeneous quality and ranks them using pairwise preference labels. This constitutes HAI’s self-exploration process. Finally, HAI learns a reward surrogate from pairwise trajectory preferences and subsequently uses the learned reward to improve the dispatch policy. Experimental results demonstrate that in the IEEE-33 bus case, the HAI method achieved operational cost optimization rates of 21.42%, 37.68%, and 36.06% under different expert demonstration data qualities. For the IEEE-118 bus case, the cost reduction effects of HAI exhibited corresponding variations with data quality, specifically 17.17%, 7.30%, and 5.16%. More significantly, HAI provides a methodological interface for incorporating demonstration and preference information into stochastic dispatch learning.
In power systems with high shares of variable renewables, spot prices are extremely volatile with frequent deep negatives, undermining renewable revenue adequacy and market-based integration. Fixed-price power purchase agreements (PPAs) cannot share this asymmetric downside risk, motivating contracts that link long-term bilateral commitments to real-time operation. This paper studies how a renewable generator should price a flexible PPA for a price-responsive large consumer to cap tail risk while keeping the consumer willing to participate. We model the problem as a Stackelberg bilevel optimization: the upper level maximizes the generator’s risk-adjusted revenue based on conditional value-at-risk (CVaR); the lower level is a mixed-integer linear program (MILP) in which consumers choose hourly between contract and spot purchases under intertemporal green-power quotas. Because lower-level binaries preclude standard Karush–Kuhn–Tucker (KKT) reformulations, we propose an adaptive two-stage search with explicit ɛ-optimality bounds, using exact hourly MILP evaluations. An 8760-hour case study combining 2024 Shandong spot prices with IEEE RTS-GMLC wind and load profiles shows that the mechanism raises the generator’s CVaR from −115 to ＋153 million CNY, turning worst-case daily losses into profits while preserving consumer participation, and also increases expected annual revenue relative to pure spot trading and a conventional fixed-price PPA. By shifting negative-price exposure from the generator to responsive demand, the quota gives regulators a tunable lever for allocating risk between the contracting parties.
Isolated power grids are particularly vulnerable to coordinated cyber and physical attacks because they rely on limited local resources and lack sufficient external support during disturbances. Existing dynamic defense methods commonly depend on predefined attack scenarios or system-state feedback alone, making it difficult to respond to changes in attacker objectives under incomplete observations. This paper develops a feedback-learning-based dynamic defense method for isolated power grids. A three-layer cyber-physical architecture is established, and the attack-defense interaction is formulated as a multistage sequential game under incomplete information. Multi-source feedback from critical-load operation, communication conditions, physical equipment, intrusion alarms, and defense responses is fused to recursively estimate the hidden system state and evolving attack intent. The resulting joint belief state is then used to select and coordinate key-node protection, network reconfiguration, and recovery dispatch actions. Hardware-in-the-loop experiments are conducted under cyberattack, physical attack, and coordinated cyber-physical attack scenarios. The proposed method consistently outperforms no-active-defense, static-defense, and state-feedback-based dynamic-defense schemes across the evaluated operational indicators, with the clearest advantage observed under coordinated attacks. These findings verify the effectiveness of incorporating online attack-intent cognition into resilience-oriented defense decision-making for isolated power grids.
Accurate identification of transmission line fault causes is essential for improving grid reliability and maintenance efficiency in modern power systems. This paper proposes an Expert-Augmented Fault Identification Framework, an interpretable framework that integrates an expert-guided knowledge refinement mechanism with reasoning large language models for intelligent fault cause identification. The proposed framework combines waveform feature extraction, contextual fault indicators, Bayesian probabilistic embedding, and retrieval-augmented generation to support structured and explainable diagnostic inference. Experiments on two real-world datasets demonstrate that the framework achieves a balanced weighted accuracy of 61.89% and an F1-score of 69.22%, outperforming conventional diagnostic approaches. Furthermore, field deployment in an operational transmission grid shows that the proposed system improves fault cause identification accuracy from approximately 57% to over 80%, representing a relative improvement of more than 20%. These results indicate that the proposed approach provides a practical and interpretable solution for intelligent fault analysis in modern power systems.
The growing integration of PV and energy storage systems (ESSs) poses operational challenges for power distribution grids (PDGs), mainly due to generation uncertainty and voltage-profile variations. Typical receding-horizon Model Predictive Control (MPC) fails to achieve efficient control due to forecast uncertainty. Meanwhile, it is difficult to directly improve the forecast due to limited feature availability, restricted data access, and insufficient domain knowledge. In this paper, we propose a Safe Deep Reinforcement Learning (SDRL) hybrid receding-horizon MPC to manage forecast uncertainty. Using an expert model based on historical data, the SDRL agent adjusts the receding-horizon MPC output to achieve better performance that is closer to the expert. The control objective is to minimize the impact of renewable energy on the external grid while ensuring critical grid operational safety, such as voltage levels. The proposed hybrid controller is tested under various forecast uncertainties, and the results show that it can improve receding-horizon MPC performance by 65% (reducing the PCC average daily fluctuation from 92 to 32) while ensuring voltage safety. Besides, the controller is tested over various forecast update intervals and MPC horizons to demonstrate its robustness. Furthermore, a scalability test shows that while the full SDRL suffers from scaling up, the proposed hybrid controller overcomes this issue and achieves better performance, minimizing the impact of the PDG by 42–62%.
The controllable current source converter (CSC) features self-turn-off capability and high control flexibility, making it a promising candidate for high-voltage direct current (HVDC) transmission applications. However, the lack of a systematic investigation into its DC impedance frequency characteristics limits the direct adoption of protection schemes designed for conventional HVDC systems. To address this issue, a DC impedance model of the CSC is established using the harmonic linearization method. On this basis, a directional criterion is developed according to the difference between the measured voltage-frequency response function and its theoretical values under forward and reverse faults. Moreover, an internal/external fault identification criterion is introduced to distinguish end-of-line internal faults from forward external faults. Based on these criteria, a single-ended protection scheme for hybrid DC transmission lines is proposed. PSCAD/EMTDC simulation results verify the effectiveness, selectivity and reliability of the proposed scheme under various fault conditions in complex operating scenarios with different fault locations and resistances.
Virtual power plants have emerged as a core mechanism for the efficient management of distributed energy resources. Peer-to-peer energy trading among internal aggregators provides a critical pathway to enhance renewable energy consumption and overall market economic benefits. To address the challenges of centralized coordination dependency and privacy protection in P2P trading, this paper proposes a decentralized optimization operation strategy for aggregator P2P trading based on Swarm Learning. First, an SL-based trading architecture is designed, which ensures that raw data remains local while incorporating anti-strategic properties. Second, a malicious profit-making strategy model is established to quantitatively analyze the economic losses caused by privacy leakage, underscoring the vital necessity of privacy protection mechanisms. Third, proximal policy optimization is adopted as the foundational decision-making model for each aggregator agent, and a privacy-preserving state augmentation mechanism is introduced to mitigate environmental non-stationarity in partially observable Markov games. Distribution-network constraints are embedded through grid-security boundaries, and forecast errors are considered by training with forecasted data and evaluating rewards with realized data. Case studies on an IEEE 33-bus system with eight aggregators show that the proposed method reduces the average daily total cost by 16.0 % compared with non‑communicative multi‑agent system and 20.4 % compared with conventional federated learning, while avoiding centralized training information and centralized parameter servers. Additional analyses verify the adaptability to newly joined data-scarce aggregators, the economic impact of privacy leakage, and the communication overhead of blockchain-assisted swarm learning.
