To mitigate voltage violations and grid losses induced by the high penetration of distributed energy resources (DERs), while addressing the limitations of single-level flexible interconnection in cross-level supportive capabilities, this paper proposes a multi-level closed-loop flexible interconnection distribution architecture with embedded DC (EDC-MLFID), along with a corresponding power flow optimization method. To tackle the deficiencies of single-level interconnections in source-load coordination, the proposed architecture establishes an embedded medium-to-low voltage DC subnetwork via Voltage Source Converters (VSCs) and employs Dual Active Bridge (DAB) converters to enable vertical power transfer, thereby forming a resilient closed-loop topology that spans multiple voltage levels. To optimize the operational performance of this architecture, a steady-state power flow model accounting for the multi-mode control characteristics of VSCs and DABs is established. Subsequently, a Long Short-Term Memory (LSTM)-based Particle Swarm Optimization (LSTM-PSO) algorithm is proposed to solve the nonlinear optimization problem, with the objectives of minimizing active power losses and voltage deviations. By leveraging LSTM to learn the historical evolutionary patterns of the system, the algorithm markedly enhances global search capability. Validation conducted on a modified IEEE 33-node system demonstrates that, compared with single-level schemes, the proposed architecture achieves maximum reductions of 41.37% and 37.72% in DC-side network losses and voltage deviations, respectively, while the corresponding maximum reductions on the AC side reach 9.00% and 7.57% under representative high-PV and peak-load conditions. These results validate the effectiveness of the proposed multi-level closed-loop approach in enhancing the flexibility and economic efficiency of modern distribution networks.
DC microgrids can operate in either standalone or grid-connected modes and typically employ hierarchical control schemes to ensure voltage regulation and power sharing among distributed generation units. The primary control layer operates locally and without communication, while the secondary and tertiary layers rely on communication protocols that are often subject to delays and packet losses. The stability of a microgrid depends on properly tuned controller parameters, which must be adapted to the characteristics of the communication network. However, in most distributed secondary-control approaches, the selection of controller gains is still based on fixed parameters, heuristic criteria, or trial-and-error procedures, with limited connection between communication delays and the desired dynamic response. To address this gap, this paper presents a dynamic-response-based tuning method for consensus-based distributed secondary control in DC microgrids. The proposed approach uses the experimentally observed maximum communication delay as a tuning constraint for selecting the secondary-control parameters. Although the linearized microgrid model does not explicitly include delay operators, the tuning procedure enforces a separation between the communication time scale and the closed-loop voltage and power dynamics, improving tolerance against communication-induced perturbations. A comprehensive dynamic model of the microgrid, including power converters, control loops, and communication-delay-aware tuning constraints, is developed to support analysis and parameter optimization. The effectiveness of the tuning method is validated through MATLAB simulations and Hardware-in-the-Loop testing with CAN Bus communication.
To address the challenges posed by extreme weather conditions and the uncertainty of distributed photovoltaics (DPV) to power restoration (PR) in traditional distribution systems (DS), this paper proposes a multi-level PR method for flexible distribution systems (FDS) based on distribution-microgrid coordination (D-MC). Firstly, an optimization model for PR in FDS is constructed, with the objectives of maximizing the load restoration amount and minimizing network losses and switch operation costs. Secondly, considering the uncertainty of DPV output and the uncertainty of DPV scenario probabilities, a min–max-max–min two-stage multi-level robust optimization (RO) model for PR in FDS is established. To tackle the difficult-to-solve coupling relationship of max–min in the multi-level RO problem, a column and constraint generation algorithm based on alternating iteration strategy of uncertain variables (U’AIS-C&CG) is proposed. Finally, simulation verification is conducted on the modified 78-node D-MC system. The results show that the proposed method can effectively improve the PR capability of the DS, while achieving a good balance between economy and robustness.
Reliable high-impedance fault (HIF) detection is important for the secure operation of active distribution networks with increasing renewable energy integration. However, weak fault zero-sequence current (ZSC) signals may closely resemble non-fault switching disturbances in both waveform morphology and time–frequency characteristics. Historical measurements are also often recorded at different sampling frequencies, producing unequal sequence lengths over the same observation interval and limiting cross-frequency generalization. A practical detector should therefore accommodate heterogeneous sampling conditions, distinguish HIFs from multiple operating disturbances, and provide confidence information for each decision. To this end, this paper proposes an HIF detection method combining an Energy-Enhanced Markov Transition Field (EE-MTF), an Adaptive Sampling-Aware architecture (ASA), and uncertainty quantification. EE-MTF converts bus ZSC signals into two-dimensional representations and enhances weak fault characteristics. ASA incorporates sampling frequency information into multiscale feature encoding and aligns image patches according to physical time, thereby establishing consistent temporal correspondence among heterogeneously sampled signals. Monte Carlo dropout is further employed to estimate predictive entropy and variance, allowing low-confidence outputs to be identified for manual verification instead of treating all predictions as equally reliable. Experiments using industrial-grade RTGeneral simulation data covering sampling frequencies from 1 to 100 kHz and multiple switching disturbances show that the proposed method achieves an accuracy of 99.53 % with an inference latency of 2.8 ms. The results demonstrate robust cross-frequency HIF detection and confidence-aware decision support under diverse sampling and disturbance conditions.
Post-contingency power flow dispatch in renewable-rich power systems requires rapid response and operational security. Conventional analytical methods have limited adaptability to dynamic post-contingency operating states and feasible regions, whereas reinforcement learning methods still struggle to embed unstructured dispatch knowledge into policy learning. To address these issues, this paper proposes a large-language-model-enabled knowledge-mask reinforcement learning (LLM-KM-RL) framework. The framework integrates a knowledge graph with LLM-based knowledge reasoning to convert unstructured dispatch-rule reasoning into action-space-aligned constraints for policy learning. Multi-hop retrieval over the knowledge graph extracts a local knowledge subgraph containing contingency-related topology and operating-procedure information. Based on this context, a Power LLM Agent Skill transforms dispatch-rule constraints into knowledge masks aligned with the action dimensions. The masks are embedded into both policy generation and value estimation in proximal policy optimization, enabling decision-making within a knowledge-constrained continuous action subspace. Case studies show that the proposed method reduces exploration in invalid action regions, avoids regulation behaviors that conflict with equipment states and dispatch rules, and improves compliance with physical constraints and operating procedures.
