TRACE: Unlocking the Potential of LLMs in Time Series Forecasting for Distributed Energy Resources

Accurate power forecasting is essential to the safe operation of power systems. However, the massive distributed energy resources (DERs) data is characterized by both diversity and scarcity. Solving the power forecasting problem requires striking a balance between the model’s generalization ability and its capacity for few-shot learning (FSL). Traditional deep learning approaches are primarily designed for single entities, which limits their generalization capabilities and makes them ineffective in data-scarce scenarios. To address these challenges, we propose TRACE, a novel forecasting framework built on pretrained large language models (LLMs). By tracing and aligning the chain-of-thought (CoT) reasoning patterns, this approach bridges the extensive prior knowledge embedded in LLMs with the unique characteristics of the distributed energy resources power forecasting. Moreover, we introduce a Multi-agent Consistency framework in TRACE, which achieves thought refinement through collaboration, enhancing the stability of LLMs’ predictions and their ability to handle complex computational tasks, thereby improving prediction accuracy. Experimental results on public real-world datasets demonstrate that TRACE achieves prediction accuracy on par with state-of-the-art deep learning models, while significantly outperforming other baseline models in data-scarce scenarios.