SeqCast: Sequence-Based Forecasting of Temperature Distribution in Smart Buildings

In this work, we introduce SeqCast, a deep learning-based forecasting system for short-term temperature distributions in smart buildings, leveraging 5-minute resolution time series data.SeqCast separately encodes historical context and decodes future sequences using exogenous variables such as air flow rates and CO2 concentration measurements.The model is trained end-to-end on multivariate building-level data, with evaluation based on both mean squared error (MSE) and KL divergence to assess prediction accuracy and distributional alignment.We compare SeqCast to a baseline vanilla LSTM model and find that the Seq2Seq-Encoder-Decoder-LSTM architecture shows significantly stronger performance on both metrics.Our results highlight the ability of sequence models to generalize across diverse temperature distribution patterns, supporting their role in intelligent building energy management.