Machine Learning Based Improvement of Precipitation Type Forecasts

Accurately predicting precipitation types (p-type) such as rain, snow, ice pellets, and freezing rain during the cool season is a critical challenge for weather forecasters. Misclassifications can lead to significant public safety issues and economic losses. This study aims to develop machine learning (ML) models for predicting classification of p-type, and determine their relative robustness. These ML models include Multinomial Logistic Regression (MLR), Random Forest (RF), Multiclass Artificial Neural Network (ANN), and Quantile Random Forest (QRF). These models are assessed through hindcast experiments using reforecast data from the Global Ensemble Forecast System (GEFS) v12 and quality-assured mPING reports. These experiments cover cool seasons from 2017 to 2019 and the lead time range of 24 to 120 hours. A three-fold cross-validation approach was employed to ensure model reliability (leave one year out) in which each scheme is applied to the GEFS reforecasts to produce probabilistic p-type forecasts, to determine the comparative skills of each scheme in predicting p-type.

The experiment shows that QRF and RF achieve the highest accuracy of 60% for lead times out to 120 hours. MLR and ANN achieve an accuracy of about 50% and 42% respectively for the same lead time range. The QRF algorithm is recommended for further development since it strikes the best balance between accuracy and reliability both for shorter as well as longer lead times. Future work involves evaluating model performance against p-type data from the GEFS real-time archive.

Keywords: Classification, precipitation type, MLR, RF, ANN, QRF, GEFS, mPING