OCC-Mamba: A Mamba-based deep learning approach for indoor occupancy prediction

Irfan Qaisar, Kailai Sun, Dianyu Zhong, Xu Yang, Qianchuan Zhao

Abstract: Accurate occupancy prediction is critical for occupant-centric control aimed at improving building energy efficiency and indoor comfort. Current machine learning and deep learning methods (e.g., MLPs, random forests, LSTMs, and Transformers) often struggle to capture both long-term occupancy trends and short-term fine-grained dynamics, limiting their effectiveness in real-world applications. To address these challenges, we present OCC-Mamba, a novel deep learning model designed specifically for building occupancy prediction. The model leverages recent advances in sequence learning to efficiently capture temporal dependencies while remaining lightweight and scalable. By combining state-space modeling with gated attention mechanisms, OCC-Mamba effectively models diverse occupancy patterns and generalizes across buildings and sensing conditions. OCC-Mamba is validated on three heterogeneous real-world datasets (Honeycomb, ROBOD, and ASHRAE 10) and achieves consistently lower errors and higher scores compared to baseline models, including MLP, LSTM, Transformer, and Mamba variants. Beyond occupancy prediction, the practical significance of OCC-Mamba is demonstrated through its integration into EnergyPlus simulations, which highlights its potential to support occupant-centric HVAC control and energy-efficient building operation. The source code is available at: https://github.com/irfanqaisar92/OCCMamba.