Two-Stage Active Distribution Network Voltage Control via LLM-RL Collaboration: A Hybrid Knowledge-Data-Driven Approach

A research team from Tsinghua University has published a groundbreaking paper proposing a hybrid AI architecture to solve critical voltage control challenges in active distribution networks (ADNs). As distributed photovoltaics (PVs) create unpredictable power flows and voltage instability, the team's 'Two-Stage Active Distribution Network Voltage Control via LLM-RL Collaboration' framework leverages the distinct strengths of two AI paradigms. A large language model (LLM) agent, utilizing models like GPT-4, handles the day-ahead stage by interpreting coarse forecasts and grid codes to schedule major equipment. A reinforcement learning (RL) agent then takes over for intra-day fine-tuning, using precise measurements to command PV inverters for reactive power control.

The technical innovation lies in the dynamic collaboration and specialized training pipelines for each agent. The LLM agent employs a self-evolution mechanism to improve its scheduling strategies, while the RL agent uses a pretrain-finetune pipeline for efficient learning. This hybrid knowledge-data-driven approach marries the LLM's reasoning and semantic understanding of grid regulations with the RL's optimization prowess for real-time control. Comprehensive tests show the method outperforms traditional approaches, offering a more practical and efficient path to managing the renewable energy transition by ensuring power quality and preventing blackouts.