Simple- and modified-poor man's QPF ensembles, and a neural network approach

A simple multi-model “poor man's ensemble” can be constructed and run inexpensively and the ensemble average QPF's (quantitative precipitation forecast) conventional skill scores are likely to surpass all of its member models'. A downside of such simple arithmetic averaging is that it tends to over-predict areas of light precipitation, smooth out the finer-scale features of individual member's predicted precipitation fields, and greatly reduces the effective resolution of the QPF prediction.

In this study we look into data over a two-year period to compare the QPF skill scores of a simple, 8-model poor man's ensemble against that of individual models', and against a “posher” poor man's ensemble, in which two models with the lowest QPF scores in the preceding 30 days were excluded from the ensemble average. We also experimented with an artificial-intelligence approach to ensembling, in which neural networks are trained on a series of past QPF and precipitation analysis data to create non-linear ensemble forecasts that have shown more spatial nuance than those from the poor man's ensembles.