UrbanAI 2025 Challenge: Linear vs Transformer Models for Long-Horizon Exogenous Temperature Forecasting

Ruslan Gokhman

Keywords: Time series forecasting, temperature prediction, smart buildings, linear models, NLinear, DLinear, exogenous variables, Transformer models, UrbanAI 2025 Challenge

TL;DR: Linear models (Linear, NLinear, DLinear) outperform Transformer-family architectures for long-horizon, exogenous-only temperature forecasting in the UrbanAI 2025 Challenge.

Abstract:

We study long-horizon exogenous-only temperature forecasting using linear and
Transformer-family models. We evaluate Linear, NLinear, DLinear, Transformer,
Informer, and Autoformer under standardized train, validation, and test splits. Re
sults show that linear baselines (Linear, NLinear, DLinear) consistently outperform
more complex Transformer-family architectures, with NLinear achieving the best
overall accuracy across all splits. These findings highlight that carefully designed
linear models remain strong baselines for time series forecasting in challenging
exogenous-only settings.

Submission Number: 78