Multi-step-ahead neural networks for flood forecasting

Abstract A reliable flood warning system depends on efficient and accurate forecasting technology. A systematic investigation of three common types of artificial neural networks (ANNs) for multi-step-ahead (MSA) flood forecasting is presented. The operating mechanisms and principles of the three types of MSA neural networks are explored: multi-input multi-output (MIMO), multi-input single-output (MISO) and serial-propagated structure. The most commonly used multi-layer feed-forward networks with conjugate gradient algorithm are adopted for application. Rainfall—runoff data sets from two watersheds in Taiwan are used separately to investigate the effectiveness and stability of the neural networks for MSA flood forecasting. The results indicate consistently that, even though the MIMO is the most common architecture presented in ANNs, it is less accurate because its multi-objectives (predicted many time steps) must be optimized simultaneously. Both MISO and serial-propagated neural networks are capable of performing accurate short-term (one- or two-step-ahead) forecasting. For long-term (more than two steps) forecasts, only the serial-propagated neural network could provide satisfactory results in both watersheds. The results suggest that the serial-propagated structure can help in improving the accuracy of MSA flood forecasts.