CO-BUILD Smart Buildings Competition: An Empirical Comparison of HVAC Temperature Prediction Models

Sohei Arisaka, Eikichi Ono, HIROYASU MIURA, Yutaka Shoji, Yangayang Li, Kuniaki Mihara

Keywords: smart building, time series

Abstract: This paper presents a comprehensive comparison of temperature prediction models for HVAC systems using the CO-BUILD Smart Buildings Competition dataset. We evaluate five modeling approaches—Naive Mean, Light Gradient Boosting Machine (LightGBM), Time-series Dense Encoder (TiDE), Time Series Foundation Model (TimesFM), and a Multimodal Large Language Model—across prediction horizons from 5 minutes to 2 weeks. Through exploratory data analysis, we identify key building characteristics, device relationships, and operational patterns that inform our preprocessing pipeline, which includes timezone conversion, missing data handling, and feature selection incorporating both direct VAV measurements and cross-device CO$_2$ influences. Our results demonstrate that LightGBM achieves superior short-term performance (up to 3 hours), while TiDE proves effective for longer horizons. TimesFM accurately predicts weekly temperature patterns in a zero-shot setting, and a multimodal LLM exhibits unique reasoning capabilities, successfully forecasting temperature shifts during operational transitions. This study provides practical insights for model selection in building energy management systems.

Submission Number: 41