A novel hybrid artificial neural network - Parametric scheme for postprocessing medium-range precipitation forecasts

Abstract Many present-day statistical schemes for postprocessing weather forecasts, in particular precipitation rely on calibration using prescribed models to relate forecast statistics distributional parameters. The efficacy of such is often constrained not only by predictor-predictand relation, but also arbitrary choices temporal window and lead time range training. To address this limitation, we propose an end-to-end, computationally efficient hybrid scheme capable producing full predictive distributions accumulation without explicit stratification forecast-observation pairs season. proposed framework uses the censored, shifted gamma distribution (CSGD) as artificial neural network (ANN) estimate parameters CSGD through a unified approach. This approach, referred ANN-CSGD, allows simultaneous estimation over multiple times seasons single model incorporating latter variables predictors ANN. We test our ANN-CSGD ensemble mean forecasts 24-h totals selected river basins California, at one- seven-day times, from Global Ensemble Forecast System (GEFS). probabilistic quantitative (PQPFs) are more skillful overall than those benchmark Mixed-type meta-Gaussian (MMGD) models. PQPFs highly improve performance predicting probability (PoP) much sharper reliable higher thresholds. demonstrate how entire available training data its modified formulation, efficiently represents interactions between GEFS season/lead thus leading enhanced performance.