What question did this study set out to answer?

This research aims to improve short-term temperature forecasting accuracy using a novel hybrid stacking ensemble model.

June 8, 2026Open Access

Short-Term High-Frequency Temperature Forecasting Using Hybrid Stacking Ensemble

Key Points

This research aims to improve short-term temperature forecasting accuracy using a novel hybrid stacking ensemble model.
Utilized a high-frequency multivariate sensor dataset from Max Planck Institute, recorded every 10 minutes.
Developed a preprocessing pipeline that included feature engineering and data normalization.
Compared performance of SARIMA, Prophet, individual machine learning models, and hybrid stacking ensemble on a 7-day test set.
Hybrid Stacking Ensemble achieved MAE=0.255°C, RMSE=0.329°C, and R²=0.949.
It significantly outperformed both individual machine learning models and traditional time-series models.
Reduced error of the best-performing model by 26%, demonstrating near-perfect linear agreement with observed temperatures.

Abstract

Weather forecasting is valuable for agriculture through crop production, renewable resource management such as solar or wind farms, and disaster management for preparation and response. It enables an understanding of future conditions so that decisions can be made at the right time. To make well-founded decisions about the long-term use of limited resources, it is essential to know accurately when these resources will be available. In this research, we used a high-frequency multivariate sensor dataset recorded every 10 minutes from Max Planck Institute weather station. This dataset includes 20 atmospheric parameters, such as temperature, humidity, pressure, and wind direction. We developed an appropriate preprocessing pipeline to prepare raw data for analysis. Our pipeline included converting timestamps into date/time format, engineering circular wind-direction features, interpolating missing values, splitting the training and test sets chronologically, and normalizing the features using Min--Max scaling. Using the processed dataset, we compared time-series models (SARIMA and Prophet) with tree-based machine learning models (Random Forest, XGBoost, and CatBoost), as well as a hybrid stacking ensemble that combined the strengths of these models. Performance was assessed on a 7-day test set using Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE), and \ (R²\). Our results showed that the Hybrid Stacking Ensemble was the most effective overall model, achieving \ (MAE=0. 255\, ^C\), \ (RMSE=0. 329\, ^C\), and \ (R²=0. 949\). It significantly outperformed both the individual machine learning models and the baseline models, such as SARIMA and Prophet. Hybrid Stacking Ensemble reduced the error of the best-performing individual model by \ (26\%\), while maintaining near-perfect linear agreement with the observed temperatures.

Read Full Paperexternally

AI से पूछें

Bookmark

View Full Paper