WCD-23. Generation of calibrated Weeks 3-4 precipitation accumulation forecasts using machine learning techniques

Abstract
Statistical post-processing is a tool to ameliorate systematic biases and ensemble dispersion errors inherent in raw forecasts produced from numerical weather prediction models. For this new Joint Technology Transfer Initiative (JTTI) project, we are developing an ensemble-based post-processing methodology to improve the forecast skill of weeks 3-4 precipitation accumulation over the contiguous United States. We will first demonstrate the skill of a more traditional statistical approach in which we fit 20 years of past model-observation pairs to a state-of-the-science precipitation-specific probability distribution. Next, we will explore if neural network (NN)-based methods, with the ability to model nonlinear interactions, can yield more skillful forecasts than the conventional approach. We will test the importance of including various large-scale weather predictors such as total column water and 500mb geopotential height as well as lower frequency processes such as ENSO, MJO, and QBO into the NN models. We anticipate that the latter will have particularly useful predictive information for precipitation forecasting at these extended forecast horizons (i.e., sub seasonal timescales). The goal is to eventually transition the NN-based post-processing algorithms into operations at the Climate Prediction Center to aid in the generation of their Week3-4 precipitation outlook maps which depict probabilities of above- and below-normal precipitation amounts. This poster will focus on results from the beginning stages of the project.