In-context learning enhanced large language model for robust distribution system state estimation

Existing data-driven distribution system state estimation (DSSE) methods face significant challenges in capturing useful information from massive zero injections that are prevalent in practical large-scale distribution systems. If zero-injection nodes are not explicitly utilized, limited real-time measurements can lead to a low observability problem. These methods are also vulnerable to bad data under heterogeneous data sources from different measurement units, such as advanced metering infrastructure (AMI), supervisory control and data acquisition (SCADA), and pseudo-measurements. This paper presents a proof-of-concept study of a large language model (LLM)-based DSSE method to address these challenges. The zero injections are transformed into textual content and are extracted through the self-attention mechanism of LLMs. The quantized low-rank adapter (QLoRA) and in-context learning (ICL) are utilized for efficient fine-tuning and quick adaptation, minimizing extensive weight adjustments across varied operational conditions. These strategies not only enhance the model’s scalability but also improve its adaptability and robustness to various situations. In particular, the self-attention mechanism allows the proposed method to deal with bad data effectively. The developed LLM-based method is evaluated against various data-driven approaches and the conventional weighted least squares (WLS) method on a realistic 2135-node Dominion Energy distribution feeder, which contains 60.98% zero-injection nodes. Specifically, incorporating zero-injection information reduces the voltage-magnitude mean absolute error (MAE) by 41.67% (from 0.0012 p.u. to 0.0007 p.u.), and under 10% bad data, the proposed method maintains a low MAE of 0.0049 p.u., compared with 0.0677 p.u. for the WLS method. These simulation results demonstrate the effectiveness and advantages of the proposed method under diverse measurement conditions and topology changes. • An LLM-based DSSE framework incorporating massive zero-injection nodes is proposed. • Zero-injection nodes supplement insufficient real-time measurements to enhance distribution system observability. • QLoRA and in-context learning enables efficient adaptation across practical large-scale distribution systems. • Consistently superior MAE outperforms WLS and multiple data-driven baselines under noise, missing/bad data, and topology changes. • Validated on a Dominion Energy 2135-node unbalanced three-phase feeder.