Deep learning for stochastic precipitation generation – deep SPG v1.0

Abstract

We present a deep-neural-network-based single-site stochastic precipitation generator (SPG), capable of producing realistic time series of daily and hourly precipitation. The neural network outputs a wet-day probability and precipitation distributions in the form of a mixture model. The SPG was tested in four different locations in New Zealand, and we found it accurately reproduced the precipitation depth, the autocorrelations seen in the original data, the observed dry-spell lengths, and the seasonality in precipitation. We present two versions of the hourly and daily SPGs: (i) a stationary version of the SPG that assumes that the statistics of the precipitation are time independent and (ii) a non-stationary version that captures the secular drift in precipitation statistics resulting from climate change. The latter was developed to be applicable to climate change impact studies, especially studies reliant on SPG projections of future precipitation. We highlight many of the pitfalls associated with the training of a non-stationary SPG on observations alone and offer an alternative method that replicates the secular drift in precipitation seen in a large-ensemble regional climate model. The SPG runs several orders of magnitude faster than a typical regional climate model and permits the generation of very large ensembles of realistic precipitation time series under many climate change scenarios. These ensembles will also contain many extreme events not seen in the historical record.