ALBERT

Description: In recent years, there has been an increase in the number of publications of tthe application of artificial neural networks (ANNs) has been investigated in a variety of hydrological contexts. However, how does the performance of ANNs compare with more traditional approaches? Here, we illustrate the respective error rates for a state-of-the-art evolutionary neural network (EANN), and the global GR4J and TOPMODEL approaches to streamflow prediction.The EANN demonstrates superior average performance (NRMSE = 0.4692 compared to 0.508 and 0.496 for GR4J and TOPMODEL respectively) for relatively short-range prediction (1 year ahead), but significantly underperforms for longer-range prediction (NRMSE = 0.4821 compared to 0.273 and 0.279 for GR4J and TOPMODEL respectively for two years ahead; NRMSE = 0.474 compared to 0.334 and 0.339 for GR4J and TOPMODEL respectively for three years ahead).Interestingly, for longer range prediction (2 and 3 years ahead), for which the global models yield a lower overall error, the EANN overestimates high peak flows, whereas the conceptual models underestimate high peak flows. For 2 years ahead, the EANN has an NRMSE of 0.0287 for high peak flows compared to 0.1877 and 0.0671 for GR4J and TOPMODEL respectively. For 3 years ahead, the EANN has an NRMSE of 0.0156 for high peak flows compared to 0.2480 and 0.0725 for GR4J and TOPMODEL respectively.These results suggest that the EANN may be a more reliable flood predictor despite greater overall error rates. We will be investigating how these trends hold up for longer prediction periods (5 to 15 years ahead).