A hybrid statistical-dynamical forecast of seasonal streamflow for a catchment in the Upper Columbia River basin in Canada

Taylor Swift-LaPointe^*Rachel H WhiteValentina Radic

• Department of Earth, Ocean and Atmospheric Sciences, Faculty of Science, University of British Columbia, Vancouver, British Columbia, Canada

We explore a hybrid statistical-dynamical approach as a methodology for potentially improving total seasonal streamflow volume forecasts at a key lake reservoir in the Upper Columbia River basin, a region vital for hydroelectric power generation in British Columbia. Seasonal streamflow forecasts in this basin at early or mid-winter initialization times often exhibit limited skill due to the lack of snowpack information in the initial conditions. Our method integrates temperature and precipitation data from the ECMWF seasonal forecasts (SEAS5) with a Long Short-Term Memory (LSTM) neural network. To our knowledge, this is the first time an LSTM has been used specifically for predicting total seasonal streamflow volume in this basin. When forced with reanalysis data (ERA5), the LSTM model performs substantially better at predicting total seasonal streamflow when trained and applied at a monthly timescale, as compared to the more typical daily timescale used in previous streamflow LSTM applications. In the case study region, when forecasts are initialized on 1 January, only three months of meteorological forecast skill are needed to achieve strong predictive skill of total seasonal streamflow (R2 > 0.7), attributed to accurate representation of snowpack build up in the winter months. The hybrid forecast, with the LSTM forced by SEAS5 data, tends to underestimate seasonal volumes in most years, primarily due to biases in the SEAS5 input data. While bias correction of the inputs improves model performance, no skill beyond that of a forecast with average meteorological conditions as input is achieved. The effectiveness of the hybrid approach is constrained by the accuracy of seasonal meteorological forcings, although the methodology shows potential for improved predictions of seasonal streamflow volumes if seasonal meteorological forecasts can be improved.