Study on Electricity Load Forecasting Using Large Language Models presented at the 56th EBES Conference
On July 4, at the 56th Eurasia Business and Economics Society Conference (56th EBES Conference), held in Istanbul in a hybrid format, Junior Research Fellow at the International Laboratory of Intangible-driven Economy Evgeniya Shenkman presented the study "Medium-Term Electricity Load Forecasting Using Large Language Models", co-authored with Anastasiya Kireecheva. The research was conducted as part of the Russian Science Foundation project No. 25-18-00539, "Comparative Analysis of AI-Based Agents and Real Individuals in Economic Decision-Making."
The 56th EBES Conference took place from July 2 to 4 at Istanbul Ticaret University, bringing together researchers in economics, finance, management, and sustainable development. Evgeniya Shenkman's presentation was delivered during the "Economic Development, Sustainability and Innovation" session, which focused on contemporary research in economics and digital technologies.
The study explores the use of large language models (LLMs) for medium-term regional electricity load forecasting. The authors investigated whether modern LLMs can improve forecasting accuracy without additional training (zero-shot forecasting) compared with classical statistical methods, and compared the performance of Russian and international language models.
The analysis was based on monthly planned electricity consumption data for the Unified Energy System of the Urals. To evaluate forecasting performance, the researchers compared the outputs of Claude, Gemini, GigaChat, and Yandex Cloud AI Alice with those of the classical SARIMA model. They also examined how different prompt designs—from a basic prompt to prompts enriched with expert and domain-specific context—affect forecasting accuracy.
The results showed that forecast quality depends strongly on prompt design. Providing information about the temporal structure of the data and the application domain significantly improved the accuracy of the best-performing models. At the same time, the Russian language models did not outperform the more general-purpose international models.
The authors also demonstrated that, with appropriately designed prompts, LLM-based forecasts can in some cases outperform the classical SARIMA model. In addition, the study proposes an approach to uncertainty estimation based on repeated LLM forecasts, making it possible to compare the resulting forecast distributions with the confidence intervals produced by statistical forecasting methods.
