Data-driven models for atmospheric air temperature forecasting at a continental climate region.

Area of Science:

• Meteorology and Climatology
• Data Science and Machine Learning

Background:

• Atmospheric air temperature is a critical parameter influencing hydrology, agriculture, and climate change.
• Accurate forecasting of air temperature is essential for effective future planning and climate change assessment.

Purpose of the Study:

• To forecast short- and mid-term (daily, weekly) air temperature over North America using various data-driven approaches.
• To evaluate the performance of Support Vector Regression (SVR), Regression Tree (RT), Quantile Regression Tree (QRT), ARIMA, Random Forest (RF), and Gradient Boosting Regression (GBR).

Main Methods:

• Applied six data-driven models including SVR, RT, QRT, ARIMA, RF, and GBR to time-series data from 2000-2021.
• Utilized autocorrelation and partial autocorrelation functions for input selection and statistical measures for model evaluation.
• Compared two data splitting scenarios: a standard calibration/validation split and a randomized training/testing split.

Main Results:

• Support Vector Regression (SVR) demonstrated superior accuracy for daily temperature forecasting (RMSE = 3.592°C, R = 0.964).
• Both Regression Tree (RT) and SVR performed well for weekly temperature predictions.
• A randomized data split significantly improved model performance, suggesting climate change impacts on temperature patterns.

Conclusions:

• Data-driven methodologies, particularly SVR and RT, are effective for high-resolution daily and weekly air temperature forecasting.
• The duration, dispersion, and volatility of historical data significantly impact predictive model efficacy.
• Randomized data splitting enhances model performance, highlighting the dynamic nature of climate change effects on temperature.