Exploring Machine Learning Approaches to Precipitation Prediction: Post Processing of Daily Accumulated North American forecasts

Abstract

This thesis presents recent work on exploring machine learning (ML) and deep learning (DL) models to improve the accuracy of 24 hour precipitation forecasts. Leveraging a comprehensive North American dataset of precipitation values from Numerical Weather Prediction (NWP) models and secondary meteorological features, the research showcases the need of ML techniques in post-processing NWP precipitation predictions. The evaluation reveals remarkable performance improvements over baseline model, with certain ML models achieving a 15% reduction in Mean Absolute Error (MAE), a 5% decrease in Root Mean Squared Error (RMSE), a 45% reduction in Median Absolute Error (MdAE), and a 50% decrease in Relative Bias (RB). Convolutional Neural Networks (CNN) and Gradient Boosting Regressor (GBR) emerged as top performers, demonstrating their proficiency in accurately predicting daily precipitation.