Combining semi-distributed process-based and data-driven models in flow simulation: a case study of the Meuse river basin

One of the challenges in river flow simulation modelling is increasing the accuracy of forecasts. This paper explores the complementary use of data-driven models, e.g. artificial neural networks (ANN) to improve the flow simulation accuracy of a semi-distributed process-based model. The IHMS-HBV model of the Meuse river basin is used in this research. Two schemes are tested. The first one explores the replacement of sub-basin models by data-driven models. The second scheme is based on the replacement of the Muskingum-Cunge routing model, which integrates the multiple sub-basin models, by an ANN. The results show that: (1) after a step-wise spatial replacement of sub-basin conceptual models by ANNs it is possible to increase the accuracy of the overall basin model; (2) there are time periods when low and high flow conditions are better represented by ANNs; and (3) the improvement in terms of RMSE obtained by using ANN for routing is greater than that when using sub-basin replacements. It can be concluded that the presented two schemes can improve the performance of process-based models in the context of flow forecasting.