Hydrological Drought Forecasting Using Nonlinear ANN

From the last decades, the world is experiencing the adverse effects of climate change like heat waves, flash flood, drought, etc., affecting the countries worldwide socioeconomically and creating crisis against food security. The current coping method relies on crisis management (a reactive approach), which could result in an ill-timed and expensive temporary fix. It has become essential to have a prior forecast for these types of natural disasters so that the effects of these disasters can be reduced to the minimum, as forecast helps in the planning and management of water resource systems. Several linear (ARMAX, SARIMA) and nonlinear techniques (nonlinear regression, ANN) are used for drought forecasting. The study uses an artificial neural network (ANN), which excels at modelling and forecasting nonlinear and nonstationary time series data. At the T. Agraharam gauge station, a streamflow forecasting model for the Krishna River was created using a three-layer neural network with a back-propagation algorithm. The daily stream flow data of 15 years (1991–2005) was used for the T. Agraharam gauge station of the Krishna River basin. The ANN network’s optimal weights for this investigation were obtained using historical data in the ANN tool of MATLAB 2016. The ANN architecture with a 6–4–1 node is selected as the best network with the least RMSE (0.110) and high correlation (0.78) with six weeks of antecedent time. This study uses a neural network model to model a streamflow following \(Q_{t} \, = \,{\text{f}}\left( {Q_{t} \, - \,n} \right)\) . A feed-forward back-propagation is used in the neural network and the training algorithm adopted is trainscg. The forecasted values obtained ranged from 0.05 to 0.95, so it is again transformed back into actual values by using the function of post-processing. The proposed methodology frequently underestimates peak flows while fitting data in the low-flows range. Over the entire range of data, a good agreement between the measured and anticipated values is seen. The model is effective at low-flow rates, but it underestimates peak flow rates; therefore, this model performance is highly suitable for drought. It can be used as a watch system for droughts as it predicts six weeks in the future and can play an effective role in drought planning and management.