River-Flow Forecasting Using Higher-Order Neural Networks

Abstract

In this paper, we propose a novel neural modeling methodology for forecasting daily river discharge that makes use of neural units with higher-order synaptic operations (NU-HSOs). For hydrologic forecasting, conventional rainfall-runoff models based on mechanistic approaches in the literature have shown limitations attributable to their overparameterization and complexity. With the use of neural units with quadratic synaptic operation (NU-QSO) and cubic synaptic operation (NU-CSO), as suggested in this paper, the refined neural modeling methodology can overcome the intricacy and inefficiency of conventional models. In this paper, neural network (NN) models with NU-HSO are compared with conventional NNs with neural units with linear synaptic operation (NU-LSO) for forecasting river discharge. This study was conducted using 1- to 5-day lead time forecasting in the Mahanadi River basin at the Naraj gauging site to evaluate the effectiveness of the higher-order neural networks (HO-NNs). Performance indices for the prediction of daily discharge forecasting indicated that NNs with NU-CSO and NNs with NU-QSO achieved better performance than NNs with NU-LSO even with a lower number of hidden neurons. Thus, this study shows that HO-NNs can be effective in hydrologic forecasting.