Deep learning based stochastic rainfall generator and forecasting for high frequency multiple-site time series

Código

I-EBHE0053

Autores

Antoine Chapon, Taha B.M.J. Ouarda, Nathalie Bertrand

Tema

WG 2.3: Near-term (annual to decadal) forecasts of water availability

Resumo

Stochastic rainfall generators are a valuable tool in flood risk assessment. This study proposes a stochastic generator for hourly rainfall at multiple sites. The occurrences and temporal intermittency of nonzero rainfall values are modeled by a Hawkes process, which is a class of point processes accounting for the dependence of occurrences. The rainfall amount of nonzero values is distributed according to an extended generalized Pareto (EGP) distribution. Both the Hawkes process and its EGP mark are powered by deep learning, with an architecture based on long-short-term memory (LSTM) layers. Despite being semi-parametric, the EGP retains an extreme value behavior in its upper tail, allowing extrapolation of extremes beyond the range of observations. The intensity of the Hawkes process and the EGP are both nonstationary in space and time, with spatial covariates and influence of the history of past events. The model can be applied to station observations or gridded datasets and can handle a high number of locations thanks to the evaluation of the likelihood via stochastic gradient descent. The model is applied to a gridded reanalysis dataset of 25 years of hourly rainfall in mainland France, and its performances are evaluated for both generations of long time series and short-term forecasting. The model can reproduce spatial and temporal patterns of hourly rainfall relevant for flood risk assessment, such as the spatial and temporal autocorrelation, the duration of events, and the marginal distribution.