[3E2-GS-2g] Machine learning

11:00 AM - 11:15 AM JST(2:00 AM - 2:15 AM UTC)

[3E2-GS-2g-01] Performance Comparison of Time Series Foundation Models Using TEPCO Power Grid Regional Supply and Demand Data

〇Kazuki Kiyoshige¹, Jin Saito¹, Makoto Tasaki¹, Akihito Takeuchi¹ (1. Daiwa Institute of Research Ltd)

Keywords:

TSFM,Electricity Demand Forecasting,Evaluation & Validation

Pretrained generative models, exemplified by large language models, have become increasingly significant in both academic and commercial domains. Time-series foundation models (TSFMs) have been released at a rapid pace; however, their comparative performance and conditions for effective application have not been sufficiently characterized. To address this gap, we present a systematic and reproducible benchmark for comparing the performance of major publicly available TSFMs. We use recently published regional electricity supply and demand data from TEPCO Power Grid, thereby eliminating risks of contamination with, or leakage from, pretraining data. The evaluation spans combinations of input context length and forecast horizon. In addition to standard error metrics, we employ visualization and regression analysis to examine how extending the forecast horizon or adjusting the context length affects accuracy, inference latency, and operational costs. The results reveal model-specific sensitivities, optimal trade-offs between context length and forecast horizon, and distinct behaviors in long-range forecasting. We also provide practical implications for real-world deployment and for guiding the selection and design of TSFMs.