Comparison of two data-driven modelling techniques for long-term streamflow prediction using limited datasets

This paper presents an investigation into the efficacy of two data-driven modelling techniques in predicting streamflow response to local meteorological variables on a long-term basis and under limited availability of datasets. Genetic programming (GP), an evolutionary algorithm approach and differential evolution (DE)-trained artificial neural networks (ANNs) were applied for flow prediction in the upper uMkhomazi River, South Africa. Historical records of streamflow, rainfall and temperature for a 19-year period (1994-2012) were used for model design, and also in the selection of predictor variables into the input vector space of the model. In both approaches, individual monthly predictive models were developed for each month of the year using a one-year lead time. The performances of the predictive models were evaluated using three standard model evaluation criteria, namely mean absolute percentage error (MAPE), root mean-square error (RMSE) and coefficient of determination (R²). Results showed better predictive performance by the GP models (MAPE: 3.64%; RMSE: 0.52: R²: 0.99) during the validation phase when compared to the ANNs (MAPE: 93.99%; RMSE: 11.17; R²: 0.35). Generally, the GP models were found to be superior to the ANNs, as they showed better performance based on the three evaluation measures, and were found capable of giving a good representation of non-linear hydro-meteorological variations despite the use of minimal datasets.