Daily scale streamflow forecasting in multiple stream orders of Cauvery River, India: Application of advanced ensemble and deep learning models

Abstract

Accurate forecasts of streamflow (Qflow) are crucial for optimal management of water reservoir systems and preparing for catastrophic events such as floods. Although several machine learning (ML) techniques have been documented to forecast Qflow with notable accuracy, no model could do so with the utmost certainty since Qflow data can be noisy and sometimes incomplete. Despite these challenges, there has been significant progress in the field of Qflow forecasting. Hence, novel techniques need to be explored or applied to forecast Qflow with improved accuracy and reduced data dependency. Recently, optimized input selection techniques were shown to increase the performance of forecast models, although they have yet to be thoroughly investigated in Qflow forecasting applications. This study aims at the application of deep-learning and advanced ML techniques, namely the (i) Convolutional Neural Network (CNN), (ii) Random Forest (RF), and (iii) Gradient Tree Boosting (GTB), to forecast daily Qflow in multiple stream orders of the Cauvery River, India. The performance of these models was compared with that of conventional multilayer perceptron (MLP), a class of feedforward artificial neural network (ANN). Antecedent daily Qflow time series were used as model inputs, with input selection guided by Akaike’s Information Criteria (AIC) and Bayesian Information Criteria (BIC). The models were also calibrated using inputs selected from commonly used cross-correlation function for comparative evaluation with those models developed using AIC and BIC. The results show that applying AIC and BIC to select the ideal model input parameters marginally increase the model’s performance. The deep-learning and advanced ML techniques clearly outperformed the conventional ANN models in forecasting one- and three-day lead daily Qflow at multiple gauging stations considered along the Cauvery River. It was found that the deep learning (CNN) model had superior forecasting accuracy with respect to all 3 stations for the 1st Input-Output (I/O) combination. An NNSE, WI and KGE indices of 0.85 or higher indicates that the forecast is highly efficient.