ABSTRACT This paper proposes a data‐driven building energy management (DDBEM) framework for smart microgrids that enhances economic efficiency and promotes fairness‐aware energy allocation. Utilising nonintrusive load monitoring (NILM) with a multitask gated recurrent unit (M‐GRU) model, the framework disaggregates smart metre data into appliance‐level consumption profiles to extract consumer behavioural features during peak periods. Consumers are subsequently classified using the k‐means algorithm based on appliance‐level characteristics, and a quantitative consumption score is assigned to each consumer. These scores are integrated into a dual‐objective microgrid scheduling model that minimises operational costs and energy not supplied (ENS) while dynamically adjusting demand response programs (DRPs) at the individual appliance level. Simulation results demonstrate that the proposed framework reduces total operating costs by up to 10.1% compared to the baseline scenario while enabling fairness‐oriented resource allocation. The energy supplied to energy‐efficient consumers increases by an average of 36%, reflecting the proposed fairness‐oriented allocation strategy. Additionally, the framework decreases outage frequency for low‐consumption consumers and mitigates peak demand through targeted DRPs. These findings indicate that integrating NILM‐based behavioural analysis with dynamic appliance‐level DRPs offers an effective and scalable solution for energy management in smart microgrids.
ABSTRACT Multiple pulsed loads in islanded DC microgrids cause operating‐point migration and load‐impedance contraction, reducing source‐load stability margins and limiting fixed VI designs. This paper proposes a stability‐margin‐driven adaptive virtual‐impedance (VI) method. An operating‐point‐dependent impedance model captures how pulsed‐load superposition and source variations affect minor‐loop interaction, whereas a worst‐case margin index identifies stability deficiency. Load‐migration and margin‐deficiency signals coordinate VI parameters to reshape source impedance across the critical interaction band. The method is implemented as a bounded software correction within cascaded controllers, with voltage, current and electrothermal constraints enforced. Simulations under source variations and mixed pulsed loads verify stability‐margin recovery, improved voltage quality and load accommodation, reduced electrothermal burden and feasible operation without passive components or power‐stage modification.
ABSTRACT Accurate identification of network topologies in low‐voltage (LV) electrical distribution networks is critical for grid management, yet remains challenging due to incomplete geographic information system (GIS) records and limited advanced metring infrastructure (AMI) coverage. This paper proposes an ILP optimisation framework that infers customer topological paths solely from static data available in DSO databases, including the GIS coordinates of network elements and the assignment of each customer to a feeder terminal junction within the MV/LV substation. An algorithm is introduced to reduce the search space to a tractable set of candidate paths, which are then optimally selected subject to network radial constraints. The framework is validated on both an academic test network and a real Belgian low‐voltage network, achieving an 88% customer connection rate. A robustness analysis under corrupted feeder assignment data demonstrates a fault detection rate exceeding 95% and a false connection rate below 5% at realistic error levels.
ABSTRACT Aiming at the problems of inaccurate clustering of massive small and medium‐sized users, inaccurate characterisation of individual differences within clusters and insufficient credibility of flexibility potential assessment, a method for flexibility potential assessment of small and medium‐sized users based on the interval estimation of price elasticity is proposed. The proposed method first constructs a price elasticity factor matrix of users and identifies the principal component of price elasticity based on entropy to achieve user clustering. Then, a price elasticity interval estimation model is constructed for each user cluster, and entropy and quantile regression are integrated to estimate the prediction interval of price elasticity for each user cluster. Finally, based on chance‐constrained programming, the flexibility potential of massive small and medium‐sized users is assessed. Simulation results show that the proposed method can effectively characterise the individual differences and uncertainties within each user cluster, and has higher robustness and credibility compared with traditional deterministic models.
ABSTRACT Gigawatt‐scale AI data centres are increasingly viewed as grid‐edge cyber‐physical flexibility assets for low‐inertia power systems. While modern data‐centre power trains (metering, controllers, UPS inverters and batteries) can technically provide fast frequency response (FFR) via active power injection, relying solely on electrochemical storage creates a critical conflict between grid support and internal backup reliability (state‐of‐charge depletion and micro‐cycling). This paper proposes a cyber‐physical plant controller (CP‐PPC) for data centres that coordinates millisecond‐scale IT load modulation via dynamic voltage and frequency scaling (DVFS) with UPS battery energy storage (UPS‐BESS) to deliver FFR. A supervisory, SoC‐aware agent allocates frequency‐support effort between DVFS‐based “fast primary response” (fast load relief) and battery injection, dynamically adapting the split to workload‐dependent DVFS capability and battery health. Time‐domain simulations on a simplified Thevenin‐equivalent grid, aggregate swing model and the IEEE 39‐bus system demonstrate that the hybrid scheme extends FFR duration beyond 3.7 times that of a battery‐only strategy (300 s vs. 81 s) and outperforms a fixed‐ratio coordination benchmark (115 s), while reducing battery cycling by 20%. We further quantify sensitivity to supervisory communication latency, model the impact of time‐varying IT workloads on DVFS headroom and estimate both battery degradation and IT throughput costs to characterise the operational cost of FFR participation.
