Abstract:
The estimation of risks associated with alternate plans and designs for water resources systems requires generation of synthetic streamflow sequences. The mathematical algorithms used to generate these sequences at monthly time scales are found lacking in two main respects: inability in preserving the dependence attributes particularly at large (seasonal to inter-annual) time lags; and, a poor representation of the observed distributional characteristics. Traditional approaches for representing such dependence consist mainly of stochastic disaggregation models. These models use generated annual streamflows that are disaggregated to monthly values while prescribing an assumed annual to monthly dependence structure. In this process, the dependence at the year boundaries and between years is not reproduced. These models are characterised based on conventional probability distributions that makes it difficult to represent “unusual” features such as asymmetry or multimodality.
Proposed here is an alternative to such conventional approaches that naturally incorporates both observed dependence and distributional attributes in the generated sequences. Use of a nonparametric framework provides a simple and effective method for reproducing the observed probability density characteristics. A careful selection of prior lags as conditioning variables imparts the appropriate short-term memory, while use of an “aggregate” flow variable defined as the aggregate flow during the past twelve months allows representation of interannual dependence in the generated sequences. The nonparametric simulation model is tested on two monthly streamflow datasets – the Beaver River near Beaver, Utah, USA, and the Burrendong dam inflows, New South Wales, Australia.
