Comparative study on stream flow prediction using the GMNN and wavelet-based GMNN

Flood flow forecasting is essential for mitigating damage in flood-prone areas all over the world. Advanced actions and methodology to optimize peak flow criteria can be adopted based on forecasted discharge information. This paper applied the models of the integrated wavelet, multilayer perceptron (MLP), time-delay neural network (TDNN), and gamma memory neural network (GMNN) to predict hourly river-level fluctuations, including storage rate change variable. Accordingly, the researchers initially used the discrete wavelet transform to decompose the water discharge time-series into low- and high-frequency components. After that, each component was separately predicted by using the MLP, TDNN, and GMNN models. The performance of the proposed models, namely wavelet-MLP, wavelet-TDNN, and wavelet-GMNN, was compared with that of single MLP, TDNN, and GMNN models. This analysis affirms that precision is better in the case of integrated models for forecasting river reach levels in the study region. Furthermore, multiple inputs-multiple outputs (MIMO) networks (MIMO-1 artificial neural network (ANN) and MIMO-2 ANN), along with multiple inputs-single output (MISO) ANN were employed for obtaining flow forecasts for several sections in a river basin. Model performances were also evaluated using the root mean squared error having less than 10% of the average mean value, with the coefficient of correlation being more than 0.91 and with the peak flow criteria showing the chances of flash floods being low to moderate with values not more than 0.15.