Forecasting Building Temperature Time Series with Exogenous Variables: ICML Co-Build Challenge

Ruslan Gokhman

Keywords: Time Series Forecasting, NLinear, DLinear, Smart Buildings, Exogenous Variables, Linear Models, Energy Management

Abstract:

We tackle the problem of forecasting building temperature time series using exogenous variables (weather and
setpoints) from the ICML Co-Build Challenge. We evaluate NLinear and DLinear models—efficient, interpretable
linear forecasting baselines. Our best model achieves a mean absolute error (MAE) of 0.22498 and mean squared
error (MSE) of 0.43481 on 10,563 samples of the official test set, reliably predicting temperature dynamics. Due
to technical issues, we were unable to evaluate on the full test set, which is larger. The entire study is anchored
around performance on this official benchmark, as required by the challenge.

Submission Number: 43