Seasonal Prediction with Error Estimation of Columbia River Streamflow in British Columbia

Abstract

Large-scale climatological states [tropical Pacific sea surface temperatures (SST), Pacific-North American (PNA) atmospheric teleconnection and Pacific Decadal Oscillation (PDO)] and local precipitation data are used to predict the April–August Columbia River streamflow at Donald, British Columbia, Canada. Using predictors up to the end of November in the preceding year, forecasts of the April–August streamflow were made by multiple linear regression (MLR) under a jackknife scheme. A correlation skill of 0.52 is attained using PDO, PNA and SST as predictors, with PDO being the strongest and SST the weakest. When local precipitation is added among the predictors, PDO becomes redundant, and MLR with precipitation, PNA and SST as predictors attained a correlation skill of 0.70. Feedforward neural-network models were used for nonlinear regression, but the results were essentially identical to the MLR predictions, implying that the detectable relationships in the short, 49-sample record are linear. A bootstrap process estimates the relative errors of the MLR predictions.