Long lead-time flood forecasting using data-driven modeling approaches

Abstract:

In spite of numerous structure measures being taken for floods, accurate flood forecasting is essential to condense the damages in hazardous areas considerably. The need of producing more accurate flow forecasts motivates the researchers to develop advanced innovative methods. In this study, it is proposed to develop a hybrid neural network model to exploit the strengths of artificial neural networks (ANNs). The proposed model has two components: i.) Dual – ANN model developed using river flows; and ii.) Multiple Linear Regression (MLR) model trained on meteorological data (Rainfall and Snow on ground). Potential model inputs that best represent the process of river basin were selected in stepwise manner by identifying input-output relationship using a linear approach, Partial Correlation Input Selection (PCIS) combined with Akaike Information Criterion (AIC) technique. The presented hybrid model was compared with three conventional methods: i) Feed-forward artificial neural network (FF-ANN) using daily river flows; ii) FF-ANN applied on decomposed river flows (low flow, rising limb and falling limb of hydrograph); and iii) Recursive method for daily river flows with lead-time of 7 days. The applicability of the presented model is illustrated through daily river flow data of Bow River, Canada. Data from 1912 to 1976 were used to train the models while data from 1977 to 2006 were used to validate the models. The results of the study indicate that the proposed model is robust enough to capture the non-linear nature of hydrograph and proves to be highly promising to forecast peak flows (extreme values) well in advance (higher lead time).