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

2026-02-20 · Source: cs.LG updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Energy Markets & Policy · Depth: Advanced, long

Summary

A new few-shot classification framework leverages Large Language Models (LLMs) to predict real-time electricity price spikes. This approach aggregates system state information, including demand, renewable generation, weather forecasts, and recent prices, into statistical features. These features are then formatted as natural-language prompts and fed to an LLM, along with general instructions, to determine the likelihood of a spike day and report a confidence score. Evaluated using historical data from the Texas electricity market (ERCOT) from 2021-2024, the LLM-based method, utilizing gpt-4.1, achieved performance comparable to supervised machine learning models like Support Vector Machines (SVM) and XGBoost. Crucially, it significantly outperformed these traditional models when limited historical data (e.g., two months instead of three years) were available, demonstrating its data-efficient classification capabilities.

Key takeaway

For research scientists developing predictive models in data-scarce domains, this LLM-based few-shot classification framework offers a compelling alternative to traditional supervised methods. You should consider integrating LLMs into your workflow, especially when historical data is limited, as this approach demonstrated superior stability and performance compared to SVM and XGBoost under such conditions. Explore converting your domain-specific numerical data into structured natural language prompts to harness LLM capabilities for classification tasks.

Key insights

LLMs can effectively classify electricity price spikes with limited data by interpreting natural language prompts of system conditions.

Principles

• LLMs excel in few-shot learning scenarios.
• Textual feature representation enables LLM inference.
• Similarity search enhances prompt generation.

Method

The framework preprocesses electricity system data into engineered features, generates natural-language prompts using embedding-based similarity search (FAISS) and Maximal Marginal Relevance (MMR) for few-shot examples, then queries an LLM (gpt-4.1) for spike prediction and confidence scores.

In practice

• Use LLMs for classification with scarce labeled data.
• Convert numerical features into natural language prompts.
• Employ FAISS for efficient example retrieval in prompts.

Topics

• Few-Shot Learning
• Electricity Price Forecasting
• Large Language Models
• Extreme Event Classification
• Data-Efficient AI

Best for: Research Scientist, AI Researcher, AI Scientist, Machine Learning Engineer

Related on AIssential

• LLMs on Tabular Data with Limited Semantics: Evidence from Industrial Car Retrofit Prediction — LLMs serve best as complementary components to strong tabular machine learning baselines, not replacements, for industrial data with limited semantics.
• A Quantitative Experimental Repeated Measures Study of Training Dynamics in a Small Llama Style Language Model Under a Compute-Aware Token Budget — Compute-constrained LLM training trajectories reveal non-monotonic degradation and instability, challenging endpoint-only evaluation.
• Streamlined Constraint Reasoning via CNN Pattern Recognition on Enumerated Solutions — CNNs can detect solution patterns to guide LLM-driven constraint synthesis, significantly accelerating hard constraint problems.
• Spectra: Rethinking Optimizers for LLMs Under Spectral Anisotropy — LLM gradients exhibit a "spike-tail" anisotropy where dominant linguistic structures suppress long-tail semantic learning.
• Electricity price forecasting across Norway's five bidding zones in the post-crisis era — LightGBM excels in Norwegian electricity price forecasting, but external factors are vital for accuracy under market stress.
• Using a Local LLM as a Zero-Shot Classifier — Locally hosted LLMs can effectively perform zero-shot classification on short, semantically complex texts where traditional methods fail.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.LG updates on arXiv.org.