Key Takeaways & Executive Findings
- •• • Two-stage VF2-based graph clustering reduces effective execution sequence variants from 20 to 1, cutting DRAM-related access cycles from 18,595 to 1,502 and raising L2 cache hit rate from 16.09% to 84.54%. This directly mitigates thread warp divergence in SIMT architectures, enabling predictable memory access patterns essential for scaling to thousands of heterogeneous power electronic devices. • • Static specialization via automatic code generation eliminates dynamic graph parsing, reducing instruction count per control step by 91.8%. This translates to a 2.7x kernel-level speedup, demonstrating that removing runtime interpretation overhead is critical for converting GPU raw compute throughput into actual EMT simulation acceleration. • • The integrated method achieves a nearly 3x end-to-end speedup over conventional CPU-based simulation for a grid with thousands of renewable energy devices. This performance level is necessary for practical wide-frequency oscillation analysis and stability studies where traditional CPU serial or multi-core parallelism fails to meet efficiency requirements. • • The approach exhibits excellent scalability, as validated on large-scale scenarios. By discarding idealized homogeneity assumptions and automatically aggregating structurally isomorphic control topologies, it addresses load imbalance that plagues simple grouping methods, making it viable for real-world grids with diverse wind, PV, and storage systems from multiple manufacturers.
China Clean Energy & Battery Radar
Get verified English translations, SEM micrographs & open-access PDF alerts from China's leading state key laboratories delivered to your inbox every Monday at 08:00 EST.
Abstract
The integration of high-penetration renewable energy sources imposes severe computational bottlenecks on electromagnetic transient (EMT) simulation of large-scale power grids containing thousands of power electronic devices. Graphics processing units (GPUs) offer high-throughput parallelism, but direct application is hindered by three obstacles: thread warp divergence caused by heterogeneous control topologies, non-coalesced memory access from graph-based computation, and instruction redundancy from dynamic parsing. This paper proposes a heterogeneous parallel acceleration method for large-scale renewable energy grids. By introducing graph isomorphism, a two-stage graph clustering algorithm identifies and aggregates control systems at the topological level, eliminating instruction flow divergence from heterogeneous controllers. A GPU-oriented automatic code generation framework transforms aggregated models into branchless, memory-coalesced computational kernels. Experimental results on a scenario with thousands of renewable devices demonstrate a nearly 3x speedup over conventional simulation methods, with excellent scalability. Specifically, compressing effective execution sequence variants from 20 to 1 reduces DRAM-related access cycles from 18,595 to 1,502 and raises L2 cache hit rate from 16.09% to 84.54%. Static specialization eliminates dynamic graph parsing logic, reducing instruction count per control step by 91.8% and achieving a 2.7x kernel-level speedup. The method provides a high-fidelity, highly versatile, and high-performance EMT simulation solution for large-scale renewable energy systems.
1. Introduction
Existing commercial EMT simulation tools for power grids with high renewable penetration rely on CPU-based serial or multi-core parallel execution, which cannot meet computational efficiency demands when thousands of power electronic devices are involved. While GPU acceleration has been explored, prior methods often assume model homogeneity or use general-purpose graph interpreters. These approaches fail in real grids where control topologies are strongly heterogeneous across device types and manufacturers, leading to severe thread warp divergence, non-coalesced memory access, and instruction redundancy. Full-GPU strategies struggle with sparse, stiff network equations, and CPU-GPU-FPGA co-designs suffer from high development cost and inflexibility for frequently changing topologies.
This work introduces a GPU-oriented heterogeneous parallel acceleration method that fundamentally addresses these bottlenecks. A two-stage graph clustering algorithm based on VF2 identifies and aggregates isomorphic control topologies, eliminating instruction flow divergence at the mathematical level. An automatic code generation framework then transforms aggregated models into branchless, memory-coalesced CUDA kernels with consistent topological ordering. Experimental validation on a scenario with thousands of renewable devices demonstrates a nearly 3x speedup over conventional simulation, with DRAM access cycles reduced from 18,595 to 1,502 and L2 cache hit rate improved from 16.09% to 84.54%. Static specialization cuts instruction count per control step by 91.8%, yielding a 2.7x kernel-level speedup. The method provides a high-fidelity, scalable solution for EMT simulation of large-scale renewable energy grids.
Loading authentic research manuscript (Pages 1–5)...
YU Zhitong, YANG Minghao, SONG Yankan, HUANG Shaowei, CHEN Ying, SHEN Chen (2026). GPU-Based Parallel Acceleration Method for Electromagnetic Transient Simulation of Large-Scale Renewable Energy Power Grids. Power Automation Equipment. https://doi.org/10.16081/j.epae.202606011
Research & Educational Purpose Only: The translations, structured abstracts, analytical annotations, and data reports provided by SinoGreenTechare intended exclusively for academic research, internal corporate R&D, and educational benchmarking. They do not constitute formal engineering, chemical safety, legal, or professional advice.
Copyright & Intellectual Property Notice: Original copyright of the underlying source articles and experimental data remains with the respective authors, institutions, and original publishing journals. SinoGreenTech claims intellectual property only over its proprietary translations, analytical syntheses, and AEO structured enhancements in accordance with international fair use and academic citation principles.
Frequently Asked Questions
What specific hardware bottlenecks in GPU-based EMT simulation does the proposed method address, and what are the quantified improvements?
The method targets three bottlenecks: thread warp divergence from heterogeneous control topologies, non-coalesced memory access from graph-based computation, and instruction redundancy from dynamic parsing. Quantitatively, graph clustering reduces effective execution sequence variants from 20 to 1, cutting DRAM-related access cycles from 18,595 to 1,502 and raising L2 cache hit rate from 16.09% to 84.54%. Static specialization reduces instruction count per control step by 91.8%, yielding a 2.7x kernel-level speedup. End-to-end, a nearly 3x speedup is achieved over conventional simulation for thousands of renewable devices.
How does the two-stage VF2-based graph clustering algorithm handle real-world heterogeneity where devices from different manufacturers have distinct control logic?
The algorithm abandons idealized homogeneity assumptions by mathematically identifying and aggregating control graphs that are essentially isomorphic but may have different execution order expressions. It compresses effective execution sequence variants from 20 to 1, which directly reduces the execution order heterogeneity faced by the unified control kernel under SIMT architecture. This aggregation eliminates instruction flow divergence and load imbalance that simple grouping methods cause, as validated on scenarios with thousands of heterogeneous devices.
What is the measured impact of static specialization on instruction count and kernel performance, and why does this matter for scalability?
Static specialization eliminates dynamic graph parsing logic, reducing the number of execution instructions per control step by 91.8%. This leads to a 2.7x kernel-level speedup. The reduction in instruction redundancy frees floating-point resources and enables more predictable memory access, which is critical for scaling to larger grids because it prevents the GPU's high compute throughput from being wasted on interpretation overhead. The method demonstrates excellent scalability, as shown by the nearly 3x end-to-end speedup on a scenario with thousands of devices.
How does the proposed method compare to full-GPU or CPU-GPU-FPGA co-design approaches in terms of numerical stability and engineering flexibility?
Full-GPU approaches force sparse, stiff network equations onto the GPU, leading to low sparse matrix solving efficiency and poor numerical stability, which fails to meet high-fidelity simulation accuracy requirements. CPU-GPU-FPGA co-designs offer speed and bandwidth advantages but suffer from high FPGA development thresholds, high hardware costs, long compilation cycles, and lack of flexibility for frequently changing topologies. The proposed CPU-GPU heterogeneous architecture with graph clustering and automatic code generation maintains numerical stability and engineering generality while achieving a nearly 3x speedup, making it the best choice for real-world grids with diverse and evolving device models.
What are the limitations of the current static specialization, and what future work is proposed to address them?
The current static specialization leads to parameter fixation, which limits dynamic tuning of control parameters. Future work aims to extend the structural identification and static specialization from the component level to the case level, enabling batch parallel simulation for large-scale N-1 fault scanning. By promoting graph isomorphism and differentiated compilation to the case level and decoupling topological logic from operating parameters, it is expected to achieve automatic clustering and incremental code generation for massive similar cases, avoiding full modeling and compilation for each N-1 scenario or parameter adjustment. This will further exploit GPU throughput for concurrent simulation of hundreds or thousands of fault cases.
Related Chinese Research & Cross-Citations
Impedance Modeling and Parameter Optimization Method for Wind Farms Considering Station-Level Control
Large-scale wind farms integrated into weak grids are susceptible to broadband oscillations, a problem that existing impedance models and control parameter optimization methods fail to address systematically because they neglect station-level control and frequency coupling effects. This paper proposes a station-level control strategy based on available power allocation and an adaptive compass search (ACS) algorithm for optimizing station control parameters. A sequence impedance method incorporating frequency coupling effects establishes the aggregated wind farm impedance, and the grid-connected multiple-input multiple-output (MIMO) system is decoupled into positive- and negative-sequence single-input single-output (SISO) impedance models. A high-precision wind farm impedance model that accounts for station-level control is constructed, and the Nyquist criterion evaluates the suppression effect of station control on broadband oscillations. The ACS algorithm optimizes station control parameters to enhance the adaptability of wind farm impedance to external grid impedance, thereby reducing oscillation risk. RT-LAB platform simulations validate the impedance modeling method and control parameter optimization. Results demonstrate that considering station-level control yields a high-precision wind farm impedance model. Compared with genetic algorithm (GA), particle swarm optimization (PSO), grey wolf optimizer (GWO), and sparrow search algorithm (SSA), ACS is more suitable for optimizing wind farm control parameters. Under specific weak-grid conditions, ACS-optimized station control parameters effectively improve grid-connected stability of the wind farm.
Active and Reactive Power Coordinated Voltage Support Control Method for High-Inertia Energy-Storage Synchronous Condenser
The high-inertia energy-storage synchronous condenser (SC-HI-ES) can provide reactive power support by adjusting excitation current and active power support by actively varying rotor speed. However, the absence of a prime mover makes its control capability difficult to quantify, preventing coordinated active and reactive power support for grid voltage. This paper analyzes the factors influencing SC-HI-ES power control capability, derives rotor current equations considering stator and rotor circuit constraints, and establishes the power controllable range (PCR) under rotor current constraints and speed variations. The relationship between the PCR and voltage support power demand is parsed, and a coordinated active-reactive power voltage support control method is proposed. Case studies validate the method's effectiveness. Results show that the proposed method enhances voltage support and avoids rotor current over-limit. Compared with methods ignoring speed variation or controlling only reactive power, the proposed method dynamically calculates PCR-based power demand under current speed, and through coordinated active and reactive power control, maximally satisfies the point of common coupling voltage target while ensuring rotor current does not exceed allowable values. This provides effective technical support for SC-HI-ES application in new-type power systems. Frequency control capability characterization and frequency-coordinated control are identified as future research directions.
Hardware-in-the-Loop Test Platform for Photovoltaic High-Frequency Controllers Based on UREP + FPGA
Real-time constraints prevent CPU-based electromagnetic transient simulators from accurately sampling high-frequency PWM signals at microsecond step sizes, leading to insufficient HIL test precision and even instability of the closed-loop system. This paper introduces a PWM averaging technique to construct a novel UREP (electromagnetic transient real-time simulator) + FPGA hardware-in-the-loop test platform. To achieve accurate PWM acquisition, the FPGA's nanosecond-level clock resolution increases the number of sampling points within a single switching period from a few points under microsecond simulation steps to several thousand points. The FPGA implements fixed-step PWM averaging, converting discrete switching states into continuous duty cycles that serve as control signals for the UREP-side inverter average model. This technique avoids frequent topology updates and state matrix reconstructions triggered by switching events, reduces high real-time computational resource consumption, improves numerical stability of the in-loop system, and enhances HIL test accuracy under large-step real-time constraints. Simulation cases and industrial-grade controller HIL tests verify the platform's feasibility and accuracy. The platform provides an efficient, reproducible verification scheme for photovoltaic controller control strategies and offers a feasible technical path for domestic electromagnetic transient real-time simulation platforms to conduct high-frequency controller HIL tests.
Design and Implementation of a Distributed Simulation Architecture for Modern Power Systems Based on Data Distribution Service
The integration of high-penetration renewable energy and power electronic converters has intensified the dynamic complexity of modern power systems, imposing stringent demands on simulation accuracy and computational efficiency. This paper proposes a distributed simulation architecture based on Data Distribution Service (DDS) that leverages dispersed computing resources to enhance scalability and efficiency while preserving numerical fidelity. The power system model is mathematically decoupled into multiple independently solvable basic subsystems, and a generic data transmission interface is designed using DDS. A time-consumption balancing scheme groups and distributes these subsystems across multiple devices, and a two-layer synchronization strategy enables efficient parallel simulation. The architecture is validated on a two-area four-machine system, the IEEE New England 10-machine 39-bus system, and the WECC 29-machine 179-bus system. Compared with centralized simulation, the average error on the two-area four-machine system remains below 1.05%. Simulation efficiency improvement reaches approximately 10% on the 10-machine 39-bus system and about 48% on the 29-machine 179-bus system. The results confirm high accuracy across different system scales and demonstrate that efficiency gains become more pronounced as system size increases, validating the architecture's scalability and compatibility. The proposed framework offers a promising pathway for large-scale power system simulation and supports future integration with edge computing and cross-platform deployment.
Heterogeneous Weighted Graph Partitioning and Decoupling Optimization Strategy for Electromagnetic Transient Parallel Simulation of AC/DC Distribution Networks
The increasing penetration of power electronic devices in AC/DC distribution networks imposes stringent computational demands on electromagnetic transient (EMT) parallel simulation. Conventional transmission-line delay decoupling methods are ill-suited to the strong electrical coupling characteristic of such networks. This paper proposes a non-delay decoupling parallel simulation acceleration framework based on heterogeneous weighted graph partitioning. An empirical computational cost evaluation model for each parallel decoupling stage is established, and a heterogeneous weighted graph model is constructed to precisely characterize the simulation computational complexity of AC/DC distribution network components, mapping matrix dimensions of device mathematical models to graph node weights. A multi-objective graph partitioning scheme is formulated that simultaneously balances partition computational overhead and minimizes the number of tie-line variables, complemented by an optimal partition number screening strategy. Simulation validation is conducted on three large-scale AC/DC distribution network composite test cases: IEEE 34-node, IEEE 123-node, and European Low Voltage (European LV) systems, all retrofitted with DC sections. Results demonstrate that the proposed empirical computational cost model achieves a simulation time fitting goodness-of-fit R² > 0.97, indicating high predictive accuracy. Under optimal partition configuration, the proposed method attains parallel speedup ratios of 12.11–15.77, significantly outperforming conventional natural partitioning schemes and effectively enhancing the EMT parallel simulation efficiency of AC/DC distribution networks.
Hierarchical Hybrid Non-Delay Decoupling Parallel Method for Electromagnetic Transient Simulation of DC-Collector Offshore Wind Farms
Real-time electromagnetic transient (EMT) simulation of DC-collector offshore wind farms is constrained by microsecond time steps, high model order from cascaded power electronic converters, and the inability of conventional decoupling methods to handle complex series-parallel topologies without introducing artificial delays. Existing non-delay decoupling methods, such as multi-area Thevenin equivalence (MATE) and compensation method, rely on branch tearing and are ill-suited for systems with numerous common-bus partitions, leading to excessive link variables and singular admittance matrices. This paper proposes a hierarchical hybrid non-delay decoupling parallel method that integrates MATE and its dual (node-tearing) approach through a layered architecture. The method constructs a mixed equivalent model tailored to DC-collector offshore wind farms, enabling flexible selection between unified and hierarchical solution modes for link variables based on operating conditions. A complete non-delay decoupling simulation workflow is established and validated on a DC series-parallel grid-connected offshore wind farm test case implemented in MATLAB. Results demonstrate that the proposed method reduces solution matrix dimensions and achieves significant speedup without compromising accuracy. For a 96-turbine wind farm, the hybrid decoupling model achieves an 11.99x speedup over the detailed model, compared to 8.77x for series-only decoupling and 6.22x for parallel-only decoupling. Mean absolute errors (MAE) for key variables remain below 0.1011, and root mean square errors (RMSE) below 0.6318, confirming high fidelity. The method enhances parallel simulation performance and offers a generalizable solution for real-time EMT simulation of large-scale offshore wind farms.