Machine learning enhancement of storm scale ensemble precipitation forecasts

• Authors: David John Gagne, II;Amy McGovern;Ming Xue
• Venue: Proceedings of the 2011 workshop on Knowledge discovery, modeling and simulation
• Year: 2011

Precipitation intensity forecasting is among the most challenging for meteorologists because of two main sources of complexity. First, the formation of precipitation requires the conversion of atmospheric moisture into water that falls to the ground and sometimes requires upward vertical motion. Multiple mechanisms can produce, enhance, or interfere with the precipitation processes, and those mechanisms operate on many scales. Numerical models explicitly represent grid resolvable scales but have to parameterize smaller-scale processes with physical and statistical relationships. Denser grids mean fewer parameterizations, but some processes are too small to resolve feasibly. These model resolution and parameterization limitations result in precipitation prediction errors. Second, the limited number of observations to initialize the models and error in such observations results in initial condition imperfections that grow as the model is run. Numerical model ensembles were developed to quantify the range of possible errors stemming from these sources by perturbing the initial conditions and varying the parameterizations of each model. Sorting the contributions of those error sources requires an archive of previous ensemble forecasts and post-processing to discover patterns and translate them into a calibrated probabilistic forecast. Machine learning techniques provide ways to perform both tasks with varying degrees of skill and understanding. This project compares the ability of multiple machine learning techniques to produce skilled probabilistic precipitation forecasts from a high resolution ensemble of numerical model forecasts and to discover patterns in the ensemble that reveal its strengths and weaknesses as well as where the largest influences in the forecast system lie. The application of machine learning techniques to the problem of ensemble post-processing has so far been very limited in the field of numerical weather prediction (NWP).