A multimodel regression-sampling algorithm for generating rich monthly streamflow scenarios

Abstract

This paper presents a multimodel regression-sampling algorithm (MRS) for monthly streamflow simulation. MRS is motivated from the acknowledgment that typical nonparametric models tend to simulate sequences exhibiting too close a resemblance to historical records and parametric models have limitations in capturing complex distributional and dependence characteristics, such as multimodality and nonlinear autocorrelation. The aim of MRS is to generate streamflow sequences with rich scenarios while properly capturing complex distributional and dependence characteristics. The basic assumptions of MRS include: (1) streamflow of a given month depends on a feature vector consisting of streamflow of the previous month and the dynamic aggregated flow of the past 12 months and (2) streamflow can be multiplicatively decomposed into a deterministic expectation term and a random residual term. Given a current feature vector, MRS first relates the conditional expectation to the feature vector through an ensemble average of multiple regression models. To infer the conditional distribution of the residual, MRS adopts the k-nearest neighbor concept. More precisely, the conditional distribution is estimated by a gamma kernel smoothed density of historical residuals inside the k-neighborhood of the given feature vector. Rather than obtaining the residuals from the averaged model only, MRS retains all residuals from all the original regression models. In other words, MRS perceives that the original residuals put together would better represent the covariance structure between streamflow and the feature vector. By doing so, the benefit is that a kernel smoothed density of the residual with reliable accuracy can be estimated, which is hardly possible in a single-model framework. It is the smoothed density that ensures the generation of sequences with rich scenarios unseen in historical record. We evaluated MRS at selected stream gauges and compared with several existing models. Results show that (1) compared with typical nonparametric models, MRS is more apt at generating sequences with richer scenarios and (2) in contrast to parametric models, MRS can reproduce complex distributional and dependence characteristics. Since MRS is flexible at incorporating different covariates, it can be tailored for other potential applications, such as hydrologic forecasting, downscaling, as well as postprocessing deterministic forecasts into probabilistic ones.