A Unified Benchmark for Multi-Horizon Outdoor Temperature Forecasting: Comparing Baseline, Statistical, and Machine Learning Models

Espinoza-Nima R.; Chuquihuanca-Carhuapoma R.; Vilcherrez-Vilela D.; Hidalgo-Sullón J.; Ramírez-Ordinola V. · CEUR Workshop Proceedings (2026)

Resumen

Currently, short-term outdoor temperature prediction faces difficulties associated with the quality of meteorological records, data processing, and differences in assessment procedures, which limit objective comparison between models and make it difficult to determine whether a predictive improvement responds to the algorithm used. Faced with this problem, the study aimed to compare reference, statistical, and machine learning models to predict temperature at horizons of 1, 3, 6, 12, and 24 hours. A total of 51 874 records from an automatic weather station at the National University of Piura, Peru, were analyzed. The data were normalized, subjected to physical range controls, aggregated by hour and processed against short periods of missing information. Subsequently, they were divided chronologically into 70% for training, 15% for validation and 15% for testing, avoiding incorporating future information into the predictors. Persistence, seasonal naïve, moving average, linear regression, Ridge, SARIMAX, XGBoost and a multilayer neural network model were compared, along with statistical tests and robustness analysis. Meanwhile, XGBoost obtained the lowest errors in the five horizons evaluated. Variant 3A reached RMSE values of 0.5991 °C and 1.1752 °C at 1 and 6 hours, while variant 3B recorded 0.9521 °C, 1.1972 °C and 1.2323 °C at 3, 12 and 24 hours. Therefore, it is concluded that rigorous temporal processing favors reliable predictions and allows establishing a reproducible procedure to compare different families of models under equivalent conditions.

Instituciones: Universidad Nacional de Piura