Heavy rainfall events, owing to their convective origin and high impacts, remain a key weather forecasting challenge. In this talk, Paul addressed two problems: (1) how to get useful regional guidance for such events (and thus increase forecaster situational awareness); and (2) how to downscale that regional guidance to the local scale where the high impacts are mostly realized.
To address these problems, evolutionary programming, multiple logistic regression, and decision trees were all successfully applied. Several other methods were tested and rejected, including boosted trees, bootstrap forests, random forests, K nearest neighbors, and single hidden layer artificial neural networks. The use of soil moisture and layer average relative humidity as additional predictors was also evaluated. Further modifications to the basic methods that were evaluated included using past analogs to select evolutionary program ensemble members for Bayesian Model Combination, and applying a different cost function in training (mean absolute error instead of mean square error). Maximizing evolutionary program ensemble diversity by increasing the population carrying capacity from 10,000 to 500,000 individuals was also evaluated.
Discussion of these techniques, and results from these various tests were provided. Finally, an exploration of possible future directions, including ways to leverage the spatial dimension of the data, and focusing on the nowcast range using deep learning was provided.