Training a Reasoning Large Language Model for Improving Power Flow Convergence

Power flow convergence is a critical and long-standing challenge in power system operations, where “black-box” data-driven approaches often lack the generalization and interpretability required for critical applications. This paper explores a new paradigm by leveraging the advanced reasoning capabilities of Large Language Models (LLMs) to address complex power system control problems, using power flow convergence as a key demonstration. We introduce a reinforcement learning framework to train a specialized LLM agent that makes sequential adjustments to system parameters. By employing a Chain-of-Thought (CoT) process, the agent emulates an expert's analytical steps, navigating the system from a non-convergent to a convergent state while providing a transparent decision-making trail. The framework features a two-stage training process: Supervised Fine-Tuning (SFT) to instill the desired reasoning format, followed by Reinforcement Learning with Verifiable Rewards (RLVR) to refine the decision-making policy. Case studies demonstrate that the trained LLM not only successfully guides non-convergent cases to a solution but also offers a high degree of interpretability by explicitly articulating its logic. Furthermore, the model exhibits strong zero-shot generalization to unseen power system topologies, showcasing the potential of reasoning-augmented LLMs as a more transparent, adaptable, and generalizable tool for power system control.