Multi-Band Fusion Framework Using Hybrid Predictors for Indoor Temperature Forecasting

Accurate indoor temperature prediction is essential for practical building control applications. However, real-world implementation is often constrained with limited input availability. To better translate the competition task into real-world scenarios, we reformulated the task by restricting external input to easily accessible weather data, reducing reliance on internal sensor measurements. This limited-data setting necessitates leveraging the auto-regressive nature of indoor temperature sequences, but rolling predictions inevitably accumulate errors. To address this, we propose a Multi-Band Fusion (MBF) framework for multi-room indoor temperature forecasting. MBF integrates models capturing different temporal frequencies: a Temporal Convolutional Network (TCN) for short-term, auto-regressive predictions, and a Light Gradient Boosting Machine (LightGBM) model for long-term trends using external weather conditions and static room features. Online calibration fuses the low-frequency LightGBM baseline with mid- and high-frequency components from TCN, producing forecasts that preserve both stability and short-term variations. On the whole time span of test set across all rooms, MBF achieved an average MAE of 2.386, representing a 24\% improvement over the baseline. Notably, in periods of high indoor temperature variability, MBF improved accuracy by 6–8\% compared to pure LightGBM, demonstrating enhanced robustness under dynamic thermal conditions. For short-term 6-hour prediction, sole TCN achieved average MAE of 1.199 on test set. The complete result analysis notebook is available at: https://drive.google.com/file/d/14MhoV3wSFt4vMcbsJa-LGFCut0-gGs/view?usp=sharing.