Comparison of two data-driven approaches for daily river flow forecasting

Abstract

Ongoing research on the use of data-driven techniques for rainfall-runoff modelling and forecasting has stimulated our desire to compare the effectiveness of transparent and black-box type models. Previous studies have shown that models based on Artificial Neural Networks (ANN) provide accurate black- box type forecasters: whilst Gene Expression Programming (GEP: Ferreira, 2001; 2006) provides transparent models in which the relationship between the independent and the dependant variables is explicitly determined. The study presented in this paper aims to advance our understanding of both approaches and their relative merits as applied to river flow forecasting. The study has been carried out to test the effectiveness of two forecasting models: a GEP evolved equation and a model that uses a combination of ANN and Genetic Algorithms (GA). The two approaches are applied to daily rainfall and river flow in the Blue Nile catchment over a five year period. GeneXproTools 4.0, a powerful soft computing software package, is utilised to perform symbolic regression operations by means of GEP and in so doing develop a rainfall-runoff forecasting model based on antecedent rainfall and river flow inputs. A transparent model with independent variables of antecedent rainfall and flow to forecast river discharge could be achieved. The ANN model is developed with the assistance of a GA: the latter being used in the selection of the ANN inputs from a pre-determined set of external inputs. The rainfall and flow data for the first four years was used to develop the model and the final year of data was used for testing. The paper describes the methods used for the selection of inputs, model development and then compares and contrasts the two approaches and their suitability for river flow forecasting. The results of the study show that the GEP model is a useful transparent model that is superior to the ANN-GA model in its performance for riverflow forecasting.