Effect of fragment-based perturbations on the disaggregation performance of a VLB streamflow generator

► An attempt to improve the disaggregation of a robust annual streamflow generator is made. ► The current approach applies perturbations subjectively based on a triangular distribution to counter smoothing. ► A formulation based on fragments and design to recover flow structure after smoothing is developed and tested. ► This formulation found not to improve disaggregation due to inadequate matching of years. In order to adequately incorporate measures of reliability in water resources planning, synthetically generated streamflow sequences are often applied in addition to historic time series. In spite of decades of research on parametric stochastic generation methods, their practical use has not been prevalent and some practitioners have been applying much simpler non-parametric bootstrap methods. The traditional bootstrap however does not provide any flows other than those in the historic record and consequently obtains synthetic sequences of low variability. This considerably limits its value as an effective stochastic generator. The Variable Length Block (VLB) bootstrap was recently developed to get rid of this limitation and to enable a more realistic replication of multi-annual flow variability by using blocks of variable length. VLB replicates annual and monthly statistics reasonably but over-estimates the minimum monthly flows and also tends to over-estimate the skewness of monthly flows. VLB also has a tendency to under-estimate serial correlation towards the middle of the year and to over-estimate correlation closer to the end and the beginning of the year. This study presents an attempt to remove these deficiencies of the VLB disaggregation scheme. The disaggregation uses a weighted average of historic fragments (monthly flow/annual flow) obtained from historic matching years with perturbations superimposed to counter the smoothing effect of averaging. The perturbations are subjectively obtained from a triangular distribution and perturbations derived to specifically recover the monthly flow structure (after smoothing) are tried out in the new scheme. These perturbations are obtained as weighted difference of pairs of fragments from historic matching years. Disaggregation using these perturbations is tested on a 5-site stochastic data generation problem. Sixty-eight years of naturalized monthly flows are used and 101 synthetic sequences are generated simultaneously at the 5 sites. The fragment-based perturbations are found to result in gross unde