CR 68:73-89 (2016)  -  DOI: https://doi.org/10.3354/cr01383

ABSTRACT: This paper proposes an artificial neural network (ANN) approach to forecasting future precipitation through spatial downscaling and constructing new intensity-duration-frequency (IDF) curves that take climate change into consideration using a temporal downscaling method. For spatial downscaling, the bat neural network (BatNN) was developed and benchmarked with the traditional scaled conjugate gradient neural network (SCGNN). Both downscaling models were trained with observed precipitation series from Kuching and selected predictors from 1961 to 1990 (base period). Evaluations from goodness-of-fit metrics and QQ-plots from 1991 to 2010 showed that BatNN outperformed its benchmark in terms of predicting accuracy. However, both models showed an underestimation of extreme precipitation events. Subsequently, future precipitation forecasts made by BatNN were used as the basis for the construction of new IDF curves. A scaling-GEV (general extreme value) approach was used to temporally downscale the predicted daily annual maximum precipitation quantile into sub-daily quantiles. The feasibility of the approach was validated with observed precipitation and the Gumbel distribution method. Predicted future IDF curves for the 2020s, 2050s and 2080s showed an increase in precipitation extremes of 19% relative to the base period.

KEY WORDS: Artificial neural network · Statistical downscaling · Precipitation forecast · Intensity-duration-frequency curve · IDF curve