Improving Daily Reservoir Inflow Forecasts with Model Combination

A major issue in real-time management of water resources is the need for accurate and reliable hydrologic forecasts at least 24 or 48h ahead. An experiment on improving the accuracy of a conceptual hydrologic model used for daily reservoir inflow forecasting, by resorting to model combination, is presented. A robust weighted-average method is used to take advantage of three dynamically different models: a nearest-neighbor model, a conceptual model, and an artificial neural network model. At each time step, the output of each of these three models is computed, and either the absolute best result is considered or the competitive results are combined using the improved weighted-average method. The latter approach has shown a significant forecast improvement for up to 4-day-ahead prediction. Moreover, it is found that with the model combination, there is no need for postcorrection of the conceptual model forecasts. It is also found that the prediction accuracy is mainly driven by the nearest-neighbor method for the 2-day-ahead forecasts, and relatively by each model afterwards. However, none of the three models appears significantly better than the combined model approach, whatever the prediction lead time.