Benchmarking Forecasting Models for Long-Horizon Prediction of Temperature Distribution in Smart Buildings

Vaibhav Sourirajan

Keywords: smart bulidings, temperature forecasting, transformer

TL;DR: We benchmark Seq2Seq LSTM, Transformer, and Kalman Filter models for one-week-ahead temperature forecasting in smart buildings, highlighting trade-offs between accuracy and complexity.

Abstract: Accurate long-term forecasting of temperature distribution in buildings is critical for optimizing control strategies, improving energy efficiency, and maintaining occupant comfort. In this work, we benchmark three approaches for forecasting one-week ahead temperature distributions in a smart building. We evaluate the following models: (1) SeqCast, a Seq2Seq Encoder Decoder LSTM; (2) a Transformer-based direct forecasting model trained with curriculum learning; (3) a Robust Kalman Filter, a lightweight baseline grounded in classical state-space modeling. All methods are evaluated using mean absolute error across the prediction horizon. Our results show that the Transformer model significantly outperforms both SeqCast and the Robust Kalman Filter. Our study highlights the trade-offs between model complexity, interpretability, and forecasting performance in the context of building-level time series forecasting.

Submission Number: 36