An Efficient Method to Correct Under-Dispersion in Ensemble Streamflow Prediction of Inflow Volumes for Reservoir Optimization

Abstract

Ensemble streamflow prediction (ESP) has been widely used to gain insight on possible future inflows to hydropower reservoirs. However underestimation of climate, model structure and initial condition uncertainty often leads to under-dispersed ESP forecasts. In this paper, we present a novel approach called “Hindcast-mode Uncertainty Estimation” (HUE) to efficiently add variability in ESP forecasts to reduce their under-dispersion. The method was tested on a Canadian catchment used by Rio Tinto – Aluminium division to produce hydropower for their aluminium smelting plants. This project was focused on correcting long-term predictions of freshet runoff volumes to optimize drawdown volumes, with up to 6 months of lead time. It was found that by adding an error term to the hydrological model’s snow water equivalent (SWE) state variable at the time of forecast in hindcasting mode, the resulting simulation could be forced to perfectly reproduce the freshet runoff volume. This error term was computed for all years on record which enabled modeling of the error’s distribution. This distribution can then be sampled from to add noise to the model’s SWE at the start of a new ESP forecast. Results show that the current winter ESP forecasts are strongly under-dispersed for the freshet runoff volume estimation and that the proposed method is able to widen the ESPs to correct the under-dispersion problem. This was validated by using Talagrand diagrams which shifted from a U-shape (prior to HUE) to a uniform distribution (with HUE). The project objectives of correcting the ESP forecast’s under-dispersion in spring runoff estimations was thus attained with minimal effort, bypassing the need to perform more complex ensemble data assimilation techniques.