Embedding Building Operation Cycles into Transformer Models for Indoor Temperature Prediction

Shundong Li, Nan Ma

Keywords: indoor temperature prediction, time series modeling, PatchTST, rotary position embeddings

Abstract: Reliable forecasting of indoor thermal conditions is essential for optimizing HVAC control, reducing energy consumption, and ensuring occupant comfort in buildings. However, accurate long-term prediction of indoor temperature remains a major challenge due to temporal dependencies and variable building dynamics. In this study, we evaluated a range of models on the Smart Buildings Control Suite benchmark, spanning classical statistical approaches (spline regression, XGBoost) to advanced sequence architectures (encoder–decoder LSTM, PatchTST, Time-LLM). We further developed a domain-adapted variant of PatchTST that integrates Rotary Positional Embeddings (RoPE) aligned with building operation cycles, such as daily and weekly schedules. Results on a six-month validation window showed that classical baselines capture short-term dynamics but fail to maintain consistent accuracy over multi-week windows, while attention-based transformers substantially outperform recurrent and boosted-tree models. Most importantly, our RoPE-augmented PatchTST achieved the lowest mean absolute error (MAE=1.76 °F) in two-week forecasts across the six-month validation period, highlighting the importance of embedding building operation-specific temporal schedules into sequence models. These results indicate that domain-aware transformers can support reliable long-horizon indoor temperature prediction and thermal comfort management, ultimately facilitating more sustainable and energy-efficient building operation.

Submission Number: 45