A Few-Shot LLM Framework for Extreme Day Classification in Electricity Markets

TL;DR

This paper introduces a few-shot LLM-based framework for predicting electricity price spikes, demonstrating comparable or superior performance to traditional models, especially with limited data, highlighting LLMs' potential in data-scarce energy forecasting.

Contribution

The paper presents a novel few-shot classification method using LLMs for electricity spike prediction, effective with limited historical data, unlike traditional supervised models.

Findings

Achieves comparable performance to SVM and XGBoost with limited data

Outperforms traditional models when data is scarce

Demonstrates LLMs' potential in energy price spike classification

Abstract

This paper proposes a few-shot classification framework based on Large Language Models (LLMs) to predict whether the next day will have spikes in real-time electricity prices. The approach aggregates system state information, including electricity demand, renewable generation, weather forecasts, and recent electricity prices, into a set of statistical features that are formatted as natural-language prompts and fed to an LLM along with general instructions. The model then determines the likelihood that the next day would be a spike day and reports a confidence score. Using historical data from the Texas electricity market, we demonstrate that this few-shot approach achieves performance comparable to supervised machine learning models, such as Support Vector Machines and XGBoost, and outperforms the latter two when limited historical data are available. These findings highlight the potential of LLMs as a data-efficient tool for classifying electricity price spikes in settings with scarce data.