Recursively updating the error forecasting scheme of a complementary modelling framework for improved reservoir inflow forecasts

- Implement an adaptive Kalman filter in a complementary modelling framework.
- Investigate convergence and consistency of the filter.
- Demonstrate the ensuing forecasting skill improvement over extended forecast lead-times.

SummaryReservoir inflow forecasting is an integral element of hydropower systems operation and is of paramount importance to hydropower producers. Effective forecasts directly impact power production scheduling, which in turn effects the revenues earned from power production. In this study, we implement a filter updating procedure (GU-COMP) that updates the gains on the error-forecasting component of a complementary forecasting framework (COMP) that also comprises a conceptual model with time invariant parameters. The GU-COMP procedure is applied for forecasting hourly flows of the Krinsvatn catchment (207 km2; located in Norway) over a forecast lead-time of 24 h. The gain coefficients are considered as the only state variables and are updated daily using the error observed between measured and forecasted flows at the catchment outlet. The performance is rated based on evaluation of filter performance (i.e. convergence and consistency), and relative assessment of forecasting skills using the root mean square error (RMSE) and the percentage volume error (PVE) metrics. Bracketing close to 95% of the innovation sequences within two standard deviations from the mean, the filter is found to be well behaved. The RMSE and PVE metrics agree that GU-COMP outperforms COMP in reducing the forecast errors, and significantly altering distributional characteristics of the PVEs in the spring and summer seasons. It is also noted that the relative forecast accuracy enhancement diminishes for forecast lead-times beyond 20 time steps (hours).