EXPLAINABLE AND CONTEXT-AWARE MACHINE LEARNING FOR COMMON AND IMBALANCED ATMOSPHERIC FORECASTING

Date

2024-11-14

Author

Şenocak, Ali Ulvi Galip

Metadata

Show full item record

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Item Usage Stats

51
views

53
downloads

Cite This

Machine learning-based post-processing can enhance physics-based numerical weather prediction models’ forecasts, yet challenges remain in handling extreme events and ensuring model explainability. This dissertation presents two novel approaches to address these limitations. First, a global context-aware density estimation model that considers cumulative data distribution and climate features through Koppen Climate Zones, improving forecasts for four atmospheric surface parameters (precipitation, temperature, wind, and relative humidity) is developed. Here, for instance, the MAE for common precipitation events improved from 2.42 mm/day to 0.03 mm/day. Second, a regional explainable two-stage quantitative precipitation forecast approach that combines classification (binary and multi-class) with regression, providing explainability through model-wide predictor importance and instance-level explanations, is developed. The two-stage approach reduced RMSE by 10.50% and increased correlation with observations by 7.50% compared to the best physics-based model, IFS. The global analysis reveals the effectiveness of different loss functions, with mean squared error outperforming mean absolute error, and the positive impact of focal loss inclusion on secondary losses. The explainability analysis of the regional approach highlights the importance of seasonality-related parameters and demonstrates the advantages of multi-class precipitation intensity classification as a first stage. Together, these approaches advance the field by addressing both the challenges of quantitative extreme event forecast (i.e., long-tailed, imbalanced, datasets) and the need for explainable machine learning models in weather forecasting. The combination of context-aware modeling and explainable artificial intelligence provides a comprehensive framework for improving weather forecasts while maintaining transparency and reliability across different climate zones and weather conditions.

Subject Keywords

Context-aware Machine Learning, Numerical Weather Prediction, Postprocessing, Long-tailed Regression, Explainable Artificial Intelligence, Kernel Density Estimation

URI

https://hdl.handle.net/11511/112676

Collections

Graduate School of Natural and Applied Sciences, Thesis