Enabling Time-series Foundation Model for Building Energy Forecasting via Contrastive Curriculum Learning
Rui Liang, Yang Deng, Donghua Xie, Fang He, Dan Wang
TL;DR
This paper tackles building energy forecasting (BEF) by adapting existing Time-series Foundation Models (TSFMs) to BEF tasks using limited real data and abundant simulated data. It introduces Contrastive Curriculum Learning (CCL), which uses a contrastive difficulty measure and a linear training scheduler to order BEF samples from easy to hard during adaptation, leveraging both real and simulated data. Empirical results on BDG, Building-900K, and UCI datasets show substantial zero-shot and few-shot improvements over direct fine-tuning, with several deployments achieving CV-RMSE below engineeringly acceptable thresholds and surpassing some state-of-the-art BEF models. The findings highlight the practical potential of curriculum learning to tailor generalized foundation knowledge to a specific domain like BEF, enabling more efficient and effective cross-domain adaptation.
Abstract
Advances in time-series forecasting are driving a shift from conventional machine learning models to foundation models (FMs) that are trained with generalized knowledge. However, existing FMs still perform poorly in the energy fields, such as building energy forecasting (BEF). This paper studies the adaptation of FM to BEF tasks. We demonstrate the shortcomings of fine-tuning FM straightforwardly from both the perspectives of FM and the data. To overcome these limitations, we propose a new \textit{contrastive curriculum learning}-based training method. Our method optimizes the ordering of training data in the context of TSFM adaptation. Experiments show that our method can improve the zero/few-shot performance by 14.6\% compared to the existing FMs. Our code and new TSFM will be available at <Anonymous Github Repo>.