Author: Louise Arnal – University of Saskatchewan
Title: A reproducible data-driven workflow for probabilistic seasonal streamflow forecasting over North America
Abstract: Seasonal streamflow forecasts are critical for many different sectors – e.g., water supply management, hydropower generation, and irrigation scheduling. Initial hydrological conditions (e.g., snow storage and soil moisture) are an important source of hydrological predictability on seasonal timescales. Snowmelt is the main source of runoff generation in high-latitude and/or high-altitude headwaters basins across North America, and the basins downstream. As a result, data-driven forecasting from snow observations is a well-established approach for operational seasonal streamflow forecasting in the USA and Canada. The aim of this work is to benchmark the skill of probabilistic seasonal streamflow forecasts across North America. To this end, we developed a reproducible data-driven workflow and implemented it for basins with a nival or glacial regime across North America. The workflow uses snow water equivalent measurements from the Canadian historical Snow Water Equivalent dataset (CanSWE), the Natural Resources Conservation Service (NRCS) manual snow surveys, and the SNOTEL automatic snow pillows in the USA. These datasets are gap filled using quantile mapping based on neighboring snow and precipitation stations. Principal component analysis is then used to define a small set of orthogonal predictor variables. These principal components are used as predictors in a regression model to generate ensemble hindcasts of streamflow volumes for nival and glacial basins across North America. Preliminary results for 93 nival basins and 17 glacial basins across Canada suggest that this forecasting method has the ability to provide skilful hindcasts (i.e., better than streamflow climatology) during the snowmelt season with up to 2-3 months lead. The results of this study provide a reference against which alternative forecasting methods (e.g., process-based forecasting models or machine learning approaches) can be assessed in the future.