A Temporal Features-Enhanced Mixture-of-Experts Approach for Indoor Temperature Prediction

Kanxuan He, Hongshan Guo

Keywords: Time Series Forecasting, Building Thermal Environment, Gradient Boosting Machines, Smart Building, Mixture-of-Experts, ICML

TL;DR: We propose a Mixture-of-Experts framework featuring temporal smoothing methods, achieving a MAE of 1.1839 across the whole validation set.

Abstract: This study presents a flexible modeling pipeline for indoor temperature prediction that leverages a Mixture-of-Experts (MoE) framework built upon Light Gradient Boosting Machine (LightGBM) models. The approach incorporates a set of temporal feature-enhanced experts using methods such as Moving Average (MA) and Exponentially Weighted Moving Average (EWMA) to embed temporal trends. A model selector is trained to assign dynamic soft weights to each expert at every time step based on contextual features, enabling the final prediction to be a weighted combination of all experts' outputs. The Soft-MoE framework achieved a mean absolute error (MAE) of 1.1839$^\circ$F across all rooms over the entire validation period. Notably, during periods with pronounced diurnal temperature fluctuations, the EWMA-enhanced expert reduced MAE by 45.4\% compared to the base mode. The proposed MoE framework demonstrates strong adaptability to diverse temporal dynamics and is readily applicable in real-world building environment control systems. The complete Jupyter Notebook is available at: https://drive.google.com/file/d/1Os6GDuHBo0CpUwvMVGwr5qiuXFoEawaB/view?usp=sharing.

Submission Number: 34