Can Large Language Models Predict Electricity Demand? A Comprehensive Comparison of 14 Models on Belgian Grid Data

This study systematically compares the performance of statistical models, machine learning, deep learning, and large language models (14 configurations in total) on the electricity load forecasting task of the Belgian grid, aiming to reveal the capability boundaries of LLMs in time-series prediction. Using nearly 10 years of Belgian grid data, key findings include: Time-LLM (an architecture adapting GPT-2 via a reprogramming layer) outperforms traditional XGBoost and LSTM; directly prompting GPT-4o for prediction yields poor results; the ensemble model (XGB+LSTM+Time-LLM) achieves the best performance.

Electricity load forecasting is a core issue in the energy industry; accurate short-term forecasting is crucial for grid dispatch, trading, and renewable energy integration. Traditional methods include statistical models (ARIMA, Prophet) and machine learning models (XGBoost, LSTM). However, with the rise of LLMs, we need to answer: Can these text models be directly applied to numerical time-series prediction? This study comes from a master's project at the University of Hull, using over 395,000 15-minute interval load data from the Belgian grid between 2015 and 2025 to compare 14 model configurations.

Dataset and Preprocessing

The data comes from the Belgian Elia public portal, aggregated into hourly data resulting in approximately 99,000 records. Preprocessing steps include: linear interpolation to fill 0.19% missing values; constructing calendar, lag (t-1/t-24/t-168), and rolling statistical features for XGBoost; standardization using StandardScaler for LSTM and Time-LLM (fitted only on the training set).

Model Lineup

• Statistical Baselines: Naive Persistence, ETS, ARIMA, Prophet
• Machine Learning: XGBoost
• Deep Learning: Two-layer LSTM (128 units)
• LLM Methods: Time-LLM (frozen GPT-2 + reprogramming layer), GPT-4o zero-shot/few-shot

Evaluation Protocol

Split by time into 70% training /15% validation /15% test; metrics include MAE, RMSE, sMAPE, MASE.

24-hour Forecasting Results (MAE/MW)

Model	MAE	MASE
Ensemble Model	263	0.49
Time-LLM	271	0.50
XGBoost	277	0.51
GPT-4o Zero-shot	481	0.89

48-hour Forecasting Results (MAE/MW)

Model	MAE	MASE
Ensemble Model	299	0.55
Time-LLM	317	0.59
XGBoost	315	0.59
GPT-4o Zero-shot	535	0.99

Key Insights

• Time-LLM performs best (among single models), direct GPT-4o yields poor results;
• XGBoost is strong, highlighting the significant value of feature engineering;
• The ensemble model is optimal, reflecting the value of diversity.

1. Hybrid Strategy is Optimal: The ensemble of XGBoost, LSTM, and Time-LLM achieves the best results;
2. LLMs Require Adaptation: Direct use of GPT-4o is impractical, Time-LLM-like architectures are feasible;
3. Feature Engineering Remains Important: XGBoost's performance demonstrates the value of domain knowledge;
4. Statistical Models as Baselines: Prophet and others are still useful in scenarios with limited data or where interpretability is needed.

Limitations

• Only uses Belgian grid data; generalizability needs verification;

Future Directions

• Explore the impact of different LLM backbone networks on time-series adaptation;
• Optimize the design of few-shot prompts for GPT-4o;
• Validate conclusions on more datasets.