Using Artificial Neural Networks for Generating Probabilistic Subseasonal Precipitation Forecasts over California

Ensemble weather predictions from global forecast systems require statistical postprocessing in order to remove systematic errors and to obtain reliable probabilistic forecasts. Many traditional postprocessing methods are based on statistical models that make parametric assumptions about the forecast distribution and/or the relationship (e.g. linearity) between predictors and predictands. Two recent papers have demonstrated for ensemble temperature and wind speed forecasts that more accurate predictions can be obtained using artificial neural networks (ANNs) for statistical postprocessing. For more complex weather variables like precipitation accumulations, one may expect an even larger potential for improvement but also a number of challenges related to the statistical peculiarities of precipitation (e.g. mixed discrete/continuous nature of the forecast distribution and heteroscedasticity of the forecast uncertainty).

In this presentation we propose a statistical post-processing approach for precipitation forecasts that is built around an artificial neural network (ANN). Both forecasts and observations are normalized in a way that is appropriate for precipitation amounts, and this normalization makes it possible to share model parameters across a larger domain. This way, data can be pooled across many grid points and a sufficiently large training data set can be assembled to afford the increased flexibility of an ANN while still being applicable in the context of subseasonal forecasting where the signal-to-noise ratio is typically very small, and large training data sets are required to avoid overfitting.

The proposed method is demonstrated with week-2, week-3, and week-4 forecasts of precipitation accumulations over California. The probabilistic forecasts obtained with this approach are shown to be reliable and to have positive skill (relative to climatology) for all lead times.