Exploring Machine Learning Approaches for Precipitation Prediction: Post Processing of Daily Accumulated North American Forecasts

This paper delves into the application of machine learning (ML) and deep learning (DL) models to enhance the precision of 24-hour precipitation forecasts. Leveraging an extensive North American dataset that includes precipitation data sourced from Numeric Weather Prediction models from various Canadian, American, and European weather agencies. In our approach, we integrate meteorological attributes such as cloud cover, wind patterns, geographical location, elevation, and more to enhance the post-processing of NWP models. Our methodology extends beyond conventional ML models by incorporating state-of-the-art techniques, such as Graph Convolutional Neural Networks, designed to exploit spatial dependencies, using Haversine distances. The trained ML models were able to achieve significant improvements over the baseline, including a 15% reduction in Mean Absolute Error, a 5% decrease in Root Mean Squared Error, a substantial 45% reduction in Median Absolute Error, and a remarkable 50% decrease in Relative Bias. It is noteworthy that Convolutional Neural Networks and Gradient Boosting Regressor exhibit strong performance in predicting lower precipitation levels whereas XGBoost and Graph Convolutional Neural Networks excelled in predicting heavier precipitation.