ABSTRACT Procuring balancing power economically in accordance with the integration of variable renewable energy sources (VRE) presents a significant challenge. To meet this challenge, control reserves for balancing power have been traded through markets in many countries. Given the experience of European countries in this regard, there are two mainstream auction mechanisms in the balancing power market: uniform‐price and pay‐as‐bid auctions. For balancing power suppliers, one of the most important issues is how to determine the proper bidding strategy in the balancing power market to satisfy their profit‐maximising goals. On the other hand, it has become a trend for various players to participate in the balancing market through Virtual Power Plants (VPPs). In this paper, the fundamental characteristics of the above two auction mechanisms for tertiary control reserves are examined using a numerical simulation for VPP players based on the recommended practice for automatic generation control (AGC30) developed by IEEJ. The bidding strategy of each VPP player in the balancing power market is modelled using the Q ‐learning algorithm to maximise the player's expected profit under the uncertainty of imbalance. The clearing points in the balancing power market with the two auction mechanisms are compared based on simulation results.
ABSTRACT This research study proposes a physics‐constrained nature‐inspired (PCNI) reinforcement learning (RL) and model predictive control (MPC) strategy that simultaneously incorporates a trigonometric Lyapunov function into both RL's training loss functions and reward shaping, that is, dual role of physics. The proposed control synergy features a unique structure, that is, a physics‐informed RL supervisor acts as a meta‐optimiser, dynamically adjusts the MPC's cost function weights and innovatively its prediction horizon in real‐time. The RL agent is trained with a proposed custom loss function, incorporating Lyapunov constraints, system physics, and operational limits for innovative, safe, and physically plausible learning. This means that, unlike employing heuristic reward shaping, this paper introduces a concept where physics is explicitly integrated into the training process. Inspired by the natural species' adaptability, such as birds adjust their flight trajectories and reaction distance under disturbances, the proposed framework dynamically reshapes the prediction horizon and cost function weights to preserve system stability. The robustness of the proposed control architecture is validated on a multi‐machine nonlinear microgrid model under severe and recurring disturbances through a rigorous simulation environment with random perturbations every 4 seconds across seven distinct initial conditions. A conventional droop, MPC, and RL controllers are included as a baseline reference for comparative analysis. Simulation results illustrate consistently bounded frequency and voltage responses, and stringent compliance with operational limits for the proposed control scheme. The proposed control methodology achieves: (a) stability rate compared with for droop, for MPC, for RL, (b) average frequency RMSE of 0.041 pu compared with 0.029 pu for droop, 0.057 pu for MPC and 0.062 pu for RL, (c) voltage RMSE of 0.0062 pu compared with 0.018 pu for droop and 0.020 pu for both MPC and RL, and (d) average final Lyapunov value of 0.022 compared with 3.059 for droop, 2.036 pu for MPC, and 2.820 for RL, illustrating convergence to the stability criterion V(x) < 0.06 in all seven test conditions with PCNI RL‐MPC control only. These research findings highlight the potential of a hybrid learning based and adaptive model‐based control framework for stable operation of multi‐machine microgrids under consistent time‐varying and severe disturbances.
ABSTRACT With the increasing integration of electric vehicles (EVs) and renewable energy in industrial parks, achieving coordinated low‐carbon operation of park‐level energy systems has become a critical challenge. Meteorological conditions introduce significant uncertainty in renewable generation, making its effective management pivotal to system operation. This paper proposes a meteorology‐informed price‐based bilevel optimisation framework. In the upper level, the operator minimises the total operating cost and carbon emissions of the park by considering fuel expenditure, grid transactions and carbon trading. The optimised price is endogenously determined to reflect the marginal supply cost together with the carbon shadow price. In the lower level, the aggregated EV cluster and energy storage system (ESS) respond to this price by scheduling their charging and discharging behaviours to achieve both economic efficiency and carbon reduction. The model jointly optimises economic dispatch, carbon mitigation and price coordination among multiple energy resources. Case studies based on an actual industrial park verify the effectiveness of the framework. The results demonstrate that it significantly reduces total carbon emissions, enhances renewable energy utilisation and improves flexibility in energy scheduling. Overall, the proposed framework provides a practical and scalable solution for intelligent and low‐carbon operation of modern industrial parks.
ABSTRACT To mitigate substantial wind and solar curtailment, several regions in China have adopted electrolytic aluminium load to absorb local renewable energy. However, integrating high proportion of renewable energy into the electrolytic aluminium system tends to induce significant tie‐line power fluctuations. To address this challenge, this paper proposes a multi‐time scale model predictive control (MSMPC) strategy aimed at mitigating tie‐line power fluctuations. The proposed strategy enables hierarchical and coordinated control of the electrolytic aluminium system by simultaneously considering control actions at both the minute‐ and second‐level time scales. A tie‐line power fluctuation control model is developed for the grid‐connected electrolytic aluminium system, which facilitates effective coordinated control among electrolytic aluminium load (EAL), thermal power units and energy storage station (ESS). Furthermore, to enhance the control performance of the MSMPC strategy, a renewable energy hybrid prediction model is introduced, which significantly improves the prediction accuracy of the MSMPC disturbance input. On this basis, an RT‐LAB experimental platform for the electrolytic aluminium system is established to validate the effectiveness of the proposed strategy. The validation results demonstrate that the proposed control strategy effectively smooths tie‐line power fluctuations under various operating conditions of the electrolytic aluminium system.
ABSTRACT Probabilistic power flow (PPF) analysis characterises the probability distributions of power system states, such as bus voltages and branch power flows, under stochastic inputs arising from load variability and renewable generation, playing a key role in the planning and operation of modern power systems. However, the practical adoption of PPF is limited by the high computational cost of existing methods at inference time. This paper proposes an accelerated PPF solution method based on distribution‐to‐distribution (D2D) deep neural network (DNN) regression, which directly maps input probability distributions to the corresponding output distributions. Unlike Monte Carlo simulation (MCS)–based PPF methods, the proposed approach requires no sampling at inference time. As a proof‐of‐concept case study, the proposed method is evaluated on the benchmark 33‐bus radial distribution feeder with solar photovoltaic generation uncertainty modelled using beta distributions. The results demonstrate that the D2D DNN closely approximates the MCS ground truth distributions, with average relative errors below 0.5% across all output quantities, while reducing average computation time from 36 to 0.03 s per PPF scenario.
ABSTRACT To address voltage instability caused by power fluctuations and state‐of‐charge (SOC) imbalance of the hybrid energy storage systems (HESS) in grid‐connected DC microgrids, a comprehensive control strategy for lithium battery‐superconducting magnetic energy storage (SMES) cooperative operation is proposed. First, a SMES control method based on zero steady‐state error voltage tracking is designed, utilising its millisecond‐level response capability to achieve zero‐overshoot compensation for instantaneous power disturbances. Second, to overcome the limitations of traditional low‐pass filter (LPF) methods in energy management, a two‐dimensional fuzzy logic controller is integrated with an LPF to establish a nonlinear mapping between SOC and power references. This mechanism dynamically optimises power weighting, maintaining the SOC of HESS within healthy ranges. Furthermore, a system impedance model is established and verified via frequency sweep simulations, the influence of the strategy on impedance characteristics is analysed to assess small‐signal stability under multiple operating conditions. Finally, MATLAB/Simulink simulation results demonstrate that the proposed strategy reduces voltage settling time by over 43.3% compared with methods such as improved droop and variable–time constant filtering. Additionally, the SMES‐based scheme outperforms supercapacitor‐based alternatives with a 57.5% reduction in settling time. The proposed approach effectively ensures the sustained stability of HESS while providing superior voltage support.
ABSTRACT This paper presents a novel optimal sizing model for a grid‐tied microgrid operating under real‐time pricing (RTP) for electricity trading with the main grid. The model determines the optimal capacities of solar photovoltaic (PV), wind turbine (WT), battery energy storage (BES) and inverter using a deep reinforcement learning approach. A double deep Q‐network (DDQN) is proposed to solve the complex sizing problem efficiently. The sizing model incorporates a rule‐based energy management strategy, where the average of day‐ahead electricity price forecasts is used to guide the battery's state of charge (SoC) decisions. Although the model is designed to be generic, a residential building in Australia is used as a case study to validate its practical applicability. Numerical results demonstrate that the proposed method under RTP conditions achieves a lower net present cost (NPC) of electricity compared to existing sizing models from previous studies. The effectiveness and robustness of the proposed deep learning‐based approach are further confirmed through comparative analysis with other machine learning techniques and metaheuristic algorithms.
ABSTRACT Accurate short‐term forecasting of electric vehicle (EV) charging demand is crucial for demand‐side management and grid stability in modern power systems, especially within the context of virtual power plant (VPP) operations. However, EV load profiles exhibit strong stochasticity and multi‐scale variability, making traditional single‐model predictors prone to overfitting, mode mixing, or degraded performance under shifting operating conditions. This study proposes a hybrid decomposition–clustering–adaptive forecasting framework that integrates variational mode decomposition (VMD), Louvain community detection and a lightweight adaptive model pool. First, VMD decomposes the raw load signal into mode components with reduced frequency overlap. Second, a correlation‐based similarity graph is constructed and processed by the Louvain algorithm to automatically group modes with coherent temporal characteristics. Finally, an adaptive prediction mechanism selects or refines models for each community based on a normalised MSE threshold. Experiments on three real‐world charging‐station datasets show that the proposed method significantly improves forecasting performance, achieving a 60.78%–75.18% reduction in MAPE compared with conventional single‐model baselines, and a 40.55%–53.70% improvement compared with nonadaptive VMD–LSTM schemes, while maintaining manageable computational cost. These results demonstrate the framework's robustness and its potential applicability to real‐time EV charging management and other nonstationary energy forecasting tasks.
ABSTRACT State‐of‐charge (SOC) estimation for lithium iron phosphate (LFP) batteries is a challenging task, owing to the wide plateau region in the open‐circuit voltage (OCV) characteristic, over which the low sensitivity of voltage to SOC renders voltage‐based correction methods unreliable. This paper proposes an ampere‐hour integration and adaptive extended Kalman filter (AH‐AEKF) fusion strategy within a region‐dependent estimation framework: in the plateau region, AH integration serves as the primary estimator to avoid errors induced by low‐voltage sensitivity. In the steep region, the AEKF is activated to perform closed‐loop correction via voltage residuals. Under operating profiles with sufficient current excitation, the dynamic parameters of the two resistor‐capacitor (2RC) equivalent circuit model (ECM) are identified online using an adaptive forgetting‐factor recursive least‐squares (FFRLS) algorithm, enabling real‐time tracking of ECM parameter variations under dynamic operating conditions. For low‐excitation constant‐current (CC) validation, offline HPPC‐identified parameters are used to avoid unreliable online updates. Furthermore, the identified parameters are incorporated into the AEKF to predict the terminal voltage, whose residual with respect to the measured voltage is used to perform closed‐loop correction of the SOC estimate. Finally, validation is conducted on a 280‐Ah LFP battery under CC tests and dynamic stress test (DST). Under correct initialisation, the proposed method maintains the maximum absolute SOC estimation error below under the HPPC, CC and DST profiles. Additional tests demonstrate improved convergence and robustness under initial SOC uncertainty, sensor disturbances and parameter perturbations.
ABSTRACT This paper proposes an MTL‐TCN‐Transformer model based on multi‐task learning in response to the challenges including climate change faced by multi‐load forecasting in integrated energy systems. To tackle the issues in multi‐load forecasting, such as the increased volatility due to climate change and the complex coupling relationships among loads, the proposed model is designed with an integrated architecture that effectively extracts long‐term temporal dependencies, captures correlations between loads and establishes global dependencies. By using the maximum information coefficient method for feature screening, nine features representing the correlation between loads, such as the ratio and difference between loads, were constructed to capture the complex correlations among multiple loads. The application of an adaptive weight update strategy suppresses fluctuations during the training process, enhancing the robustness and convergence speed of the model. The experimental results show that the model proposed in this paper exhibits better performance than the traditional methods in the four‐season multi‐load forecasting, with an average reduction of 67.952% in the MAPE of the electrical load. This paper provides a solution for high‐precision multi‐element load forecasting and points out future expansion directions such as enhancing cross‐regional generalisation capabilities.
ABSTRACT With the rapid expansion of electric vehicles (EVs), effective charging‐service‐fee (CSF) pricing is increasingly important for coordinating charging‐station operation and user‐side charging decisions. This paper proposes an operator‐side dynamic CSF pricing model within a simulation‐driven sequential leader–follower framework. The charging operator acts as the leader and jointly adjusts the CSFs of its self‐operated electric vehicle charging stations (EVCSs), whereas EV users respond through route‐ and queue‐informed station selection. The operator objective coordinates revenue, charging‐pile utilisation and a station‐level power distribution network (PDN) imbalance proxy. Standard soft actor‐critic (SAC) is retained as the pricing‐policy backbone and is equipped with three task‐specific stabilisation modules: the demand‐ and congestion‐aware stratified experience pool (DCA‐SEP), the critic‐disagreement‐regularised joint Bellman loss (CDR‐JBL) and the constraint‐triggered adaptive delayed update (CT‐ADU). In a Wuxi case study based on five independent experiment‐level runs, the proposed model increases the mean revenue of self‐operated EVCSs by 45.70%, reduces the PDN‐imbalance proxy by 18.40% and lowers total charging time under the reference demand setting.
ABSTRACT In recent years, peer‐to‐peer (P2P) energy trading has attracted significant attention with the goals of making optimal use of local generation, reducing costs, increasing the share of renewables in energy consumption and enhancing the resilience of the local grid. One aspect that has received less attention in designed settlement mechanisms for P2P energy trading is the prioritisation of generation and consumption by energy sellers and buyers. Energy sellers/buyers typically own different devices for generation/consumption and incur different rates/costs depending on their production/consumption level; therefore, prioritisation of generation and consumption in P2P markets is always a fundamental element. In this paper, this issue is modelled in detail: sellers and buyers prioritise their generation and consumption devices in each trading time interval and enter the P2P market according to those priorities. The prioritisation is based on price and the amount of energy produced/consumed, effectively giving prosumers full control over managing their resources. On the other hand, the determination of settlement and transaction prices is entirely dependent on supply and demand, which results in trades being settled fairly. In fact, the settlement price for transactions leads to satisfaction for both buyers and sellers, because it fully reflects their preferences. Additionally, a negotiation‐based approach between buyers and sellers is proposed, relying on direct negotiation between them. In this approach, buyers and sellers interact with the aim of increasing social welfare and thereby can achieve their energy trading objectives. The designed market‐settlement mechanism offers technical advantages such as fast response and low latency, computational scalability, ease of implementation, compatibility with variable resources and storage and reduced need to exchange sensitive information. Simulation results show that, besides improving metrics such as simulation time and social welfare, the proposed mechanism is also highly effective in aspects like fairness in energy distribution.
ABSTRACT As inputs to wind power forecasting models, meteorological factors are vulnerable to adversarial attacks, leading to deviations in forecasted wind power. Existing research on adversarial attacks against renewable energy forecasting models mostly focuses on a single destructive objective, neglecting attack stealthiness. This makes it challenging to achieve actual destructive effects in scenarios where example detection mechanisms are deployed. Therefore, this paper proposes a Dual‐Objective Adversarial Attack (DOAA) algorithm that balances attack destructiveness and stealthiness to achieve the joint optimisation of both. First, a wind power forecasting model based on deep neural network is constructed, which accurately captures the complex coupling relationships between meteorological factors and wind power. Second, a graph autoencoder (GAE) model suitable for time‐series example detection is designed, which fuses the temporal correlations of data from adjacent time periods using graph structures and realises the detection of adversarial examples through reconstruction loss. Finally, the above models are integrated to construct the DOAA algorithm based on Projected Gradient Descent. By adjusting the coefficient of the reconstruction loss term, the preference for attack destructiveness or stealthiness can be achieved. Experimental results verify that the proposed DOAA algorithm achieves a favourable balance between attack destructiveness and stealthiness.
ABSTRACT The studies pertaining to scheduling of plug‐in electric vehicles (PEVs) and distributed energy resources (DERs) in unbalanced AC microgrids (MGs) usually consider time varying lumped load models. This may lead to unrealistic values and inappropriate system support as unbalanced MGs have different loading levels for each phase. Therefore, in this paper optimal scheduling of DERs and PEVs is performed in 3‐phase unbalanced AC MG whilst considering different per‐phase loading levels of time‐varying voltage dependent (TVVD) load models. At first, the impact of lumped loading and per phase loading approach on DERs scheduling in unbalanced AC MG is investigated for TVVD load models. Then, PEVs are connected in presence of DERs whilst considering per phase loading in case of all TVVD loads to minimise the generation cost and system losses. Enhanced grasshopper optimization algorithm (EGOA) is utilised to evaluate the optimum performance of unbalanced AC MG. The results demonstrate that the values obtained through a per‐phase loading approach are more realistic and PEVs integration has a significant impact on cost reduction.
ABSTRACT This paper proposes a Learning‐Augmented Distributionally Robust OPF (LA‐DROPF) framework for coordinated transmission‐distribution dispatch under deep uncertainty. The framework embeds GNN‐derived power flow sensitivities within Wasserstein‐metric DRO, where surrogate errors are absorbed by an inflated ambiguity radius. A hybrid Benders‐ADMM decomposition maintains inter‐operator privacy with convergence guarantees, and a CVaR layer controls tail voltage‐security risk. Finite‐sample guarantees ensure exponential decay of out‐of‐sample constraint violations, and an online mirror‐descent mechanism adapts the ambiguity set in real time. Experiments on 217‐bus and 630‐bus coupled systems—including tests with heterogeneous distribution feeder topologies—show that LA‐DROPF reduces DS‐level worst‐case cost by 33% relative to stochastic OPF and eliminates voltage violations that afflict deterministic and GNN‐only methods. A controlled ablation isolating the CVaR and radius‐inflation effects—including a fair comparison against exact‐sensitivity DRO augmented with the same CVaR layer and equivalent total radius—confirms that the combined mechanism yields up to 7% worst‐case cost reduction under heavy‐tailed uncertainty beyond what the Wasserstein radius alone provides. Benders‐ADMM convergence to within 1% optimality gap and AC power flow feasibility are verified explicitly across all test systems.
ABSTRACT A rapid integration of distributed energy resources and electrification has increased the need for intelligent energy management in microgrids under grid‐connected and islanded operation. The variability of renewables, energy storage system cycle limits and market interactions are primary drivers for learning‐based coordination in microgrids. Therefore, this papers presents a comprehensive review of reinforcement learning and deep reinforcement learning methods for microgrid energy management. The comparison of RL and DRL methods is provided for MG EMS applications. Multiagent DRL frameworks and constrained safe RL architectures are also discussed. The states, actions, rewards, evaluation metric and grid‐connectivity are also presented for reinforcement learning based microgrid energy management systems. Finally, limitations of existing methods are identified, including challenges related to reward function design, partial observability, uncertainty management, cybersecurity and real‐world deployment along with future research directions.
ABSTRACT Cross‐provincial electricity trading in China expands the spatial scope of resource allocation and allows the adequacy benefits provided by provincial capacity mechanisms to extend beyond local jurisdictions. However, the costs of such mechanisms are still largely recovered within the provinces where they are implemented, giving rise to cross‐provincial cost leakage and a mismatch between the beneficiaries of capacity adequacy and the entities bearing the associated costs. This mismatch distorts market incentives and weakens the effectiveness of long‐term supply adequacy signals in cross‐provincial trading. To address this issue, this paper reveals the formation mechanism of the rights–responsibilities mismatch in capacity mechanisms under coupled cross‐provincial electricity trading and analyzes its critical impact on market incentive signals and power supply security. Based on this analysis, an improvement scheme for fairer capacity cost allocation is proposed, which is compatible with the current market operation environment. The proposed scheme corrects cross‐provincial market incentive signals by allowing capacity costs to be transmitted to external beneficiaries that actually receive adequacy support through cross‐provincial transactions, thereby achieving a cross‐provincial match between costs and benefits. Simulation results demonstrate that the proposed method improves the fairness of capacity cost sharing and mitigates distorted trading incentives. This study provides a useful reference for the design and operation of cross‐provincial electricity markets in the presence of non‐localised capacity adequacy benefits.
ABSTRACT Hyper‐scale data centers, with rapid load modulation capabilities, introduce a novel cyber‐physical attack vector for power systems. This paper investigates coordinated resonance attacks wherein adversaries exploit data center load flexibility to excite inter‐area oscillation modes. We formulate the attack as a Markov Decision Process and develop a Deep Reinforcement Learning framework using Proximal Policy Optimization that enables model‐free adversaries to discover mode‐matched forcing patterns through local frequency observations alone. We extend the analysis to multi‐agent scenarios using Centralised‐Training‐Decentralised‐Execution, demonstrating that geographically distributed data centers can learn anti‐phase coordination without real‐time communication. Simulations on a reduced‐order dynamic equivalent (tuned to the IEEE 39‐bus 0.64 Hz mode) show that: (1) a single agent learns mode‐targeted forcing using 200 MW of controllable load; and (2) two coordinated agents achieve comparable amplification with only 100 MW each. Sensitivity analysis reveals attack severity scales linearly with load magnitude and inversely with system inertia, and that the learned forcing pattern degrades gracefully under mode frequency mismatch (> 50% effectiveness at 6% mismatch, still ∼3 × random at 25%). High‐fidelity validation using ANDES on IEEE 39‐bus and WECC 179‐bus benchmarks confirms that narrowband forcing produces consistently larger frequency deviations than random modulation—specifically, 1.20 × and 2.20 × the random baseline on the respective benchmarks—while remaining bounded (tens of mHz) under standard control loops. The active‐power focus of the attack is justified by the predominantly real‐power nature of data center loads and the electromechanical coupling governing inter‐area modes; secondary reactive power effects are analysed and shown to be an order of magnitude smaller. These findings highlight an emerging security concern as data center penetration increases and grid inertia declines.
Impedance-based approaches can be employed to address harmonic stability issues in inverter grid-connected systems. The system stability can be analysed and improved by constructing an impedance model for the inverter and then performing impedance reshaping. The Lyapunov-Function-based Control (LFC) strategy ensures global stability of the inverter grid-connected system under large-signal transient conditions. However, research on the impedance modelling and stability analysis for inverters using this method remains insufficient. To address the harmonic stability issue of the inverter based on LFC under weak grid conditions, this article first establishes a mathematical model for the inverter's output impedance. Then the influence of various parameters on the stability of the inverter grid-connected system is analysed based on the impedance model. Furthermore, a method for impedance reshaping is proposed to improve system stability by tuning the active damping coefficient. Finally, simulations and experiments are conducted to validate the theoretical output impedance model and the proposed stability enhancement approach.
ABSTRACT The rapid expansion of renewable energy and the growing role of electricity market trading have created an urgent demand for ultra‐short‐term probabilistic forecasting of wind and solar generation. Existing studies often emphasise deterministic accuracy, yet lightweight and deployment‐friendly approaches for calibrated uncertainty modelling remain scarce. This paper proposes a practical framework for ultra‐short‐term probabilistic forecasting based on a Swin Transformer backbone combined with a learnable perturbation strategy. Unlike conventional ensemble methods that require training multiple models, the proposed approach introduces adaptive sample‐level disturbances into a single deterministic predictor, enabling efficient uncertainty quantification without additional training overhead. The framework is validated on 15‐min resolution wind and solar power data from a region with diverse meteorological and spatial conditions. Results show that the disturbance‐based ensemble achieves reliable predictive intervals while preserving high deterministic accuracy, supporting risk‐aware scheduling and energy management in renewable‐rich power systems. The proposed method offers a scalable and computationally efficient alternative for uncertainty‐aware forecasting in large‐scale zero‐carbon energy applications.
