AI-Driven Virtual Power Plant Scheduling: CUDA-Accelerated Parallel Simulated Annealing Approach

Abstract

The proliferation of distributed energy resources in smart cities calls for scalable and timeefficient optimization of virtual power plants. This study introduces a GPU-accelerated Multiple-Chain Simulated Annealing (MC-SA) framework that employs dual-level parallelization to enable real-time VPP scheduling. By improving computational speed and responsiveness, the method advances resilient, adaptive, and sustainable urban energy management.

What are the main findings?

Developed a fully GPU-accelerated Monte Carlo Simulated Annealing (MC-SA) framework that integrates multi-chain algorithmic parallelism with prosumer-level decomposition for scalable VPP scheduling.

Achieved over 10× speedup and near real-time runtimes on 1000-prosumer scenarios, while ensuring strict feasibility through a projection-based constraint-handling mechanism.

What are the implications of the main findings?

Enables real-time metaheuristic optimization for smart grid applications, supporting intraday market responsiveness and grid-aware dispatch decisions.

Provides a transferable, GPU-compatible parallelization strategy for distributed optimization and control across large-scale smart-city infrastructure.

Efficient scheduling of virtual power plants (VPPs) is essential for the integration of distributed energy resources into modern power systems. This study presents a CUDA-accelerated Multiple-Chain Simulated Annealing (MC-SA) algorithm tailored for optimizing VPP scheduling. Traditional Simulated Annealing algorithms are inherently sequential, limiting their scalability for large-scale applications. The proposed MC-SA algorithm mitigates this limitation by executing multiple independent annealing chains concurrently, enhancing the exploration of the solution space and reducing the requisite number of sequential cooling iterations. The algorithm employs a dual-level parallelism strategy: at the prosumer level, individual energy producers and consumers are assessed in parallel; at the algorithmic level, multiple Simulated Annealing chains operate simultaneously. This architecture not only expedites computation but also improves solution accuracy. Experimental evaluations demonstrate that the CUDA-based MC-SA achieves substantial speedups—up to 10× compared to a single-chain baseline implementation while maintaining or enhancing solution quality. Our analysis reveals an empirical power-law relationship between parallel chains and required sequential iterations (iterations ∝ chains^{−0.88±0.17}), demonstrating that using 50 chains reduces the required number of sequential iterations by approximately 10× compared to single-chain SA while maintaining equivalent solution quality. The algorithm demonstrates scalable performance across VPP sizes from 250 to 1000 prosumers, with approximately 50 chains providing the optimal balance between solution quality and computational efficiency for practical applications.