Presentation Information
[2Yin-B-46]Renewable energy power generation prediction by lightweight general-purpose time series modelsEffectiveness and challenges of lightweighting time series foundation models for real-world problems
〇Shogo Masaya1, Yohei Nishitsuji2 (1. INPEX Corporation, 2. Sumitomo Corporation)
Keywords:
General-purpose time series model,Lightweight,Simulation,Energy
In recent years, the paradigm of foundation models—most notably, large language models—has expanded into the field of time series forecasting, resulting in active research on time series foundation models. However, these models face several challenges: they require massive training datasets and significant computational resources, and their high inference latency makes them unsuitable for forecasting tasks that demand real-time responsiveness. To address these issues, lightweight general-purpose time series forecasting models have been proposed to reduce computational overhead. This study focuses on renewable energy forecasting, such as wind power, as a time-sensitive task to validate the effectiveness of these lightweight models. Numerical experiments suggest that, although performance varies depending on specific conditions, lightweight general time series models can achieve predictive accuracy comparable to current industry standards, such as light gradient boosting machine. This presentation discusses the effectiveness and challenges of this approach.