Application of Neural Network to Flood Forecasting, an Examination of Model Sensitivity to Rainfall Assumptions

This paper describes the development of a back-propagation Neural Network model for predicting flood and its application to a short response catchment. Common operational flood forecasting is based on traditional physically based and conceptual methods. These methods, despite being based on robust physical laws, have limitations. Data-driven models are plausible alternatives to physically based methods for certain flood forecasting applications. However, there is still a need for further demonstration of their ability in flood forecasting in order to build enough confidence for their application in practice. Among the aims of developing forecasting models, is utilizing them as a decision support system. To ensure applicability of the system for real-world application, limitations of the model should be outlined. Initial simulations conducted on a small-response time catchment outlined sensitivity of the accuracy of the model to rainfall and the way it is addressed. In this study, uncertainties associated with unseen portion of rainfall at the time that the actual forecasting is carried out, during real-world flood event are emulated. Four scenarios are considered for this study, rainfall is assumed known, rainfall is naively predicted, rainfall is treated as hidden variable and rainfall is predicted using axillary ANN. The study shows that a proposed ANN is adequately skilled for short-term flood predictions; however variability in rainfall within span of few hours limits reliability of predictions as time horizon increases. In addition the study proposes several directions that may improve the forecasts despite inherit limitations. In particular, subsequent to qualitative performance analysis, it was observed that optimization goal defined for ANN, Root Mean Square Error (RMSE) is not fully aligned with purpose of a decision support system and hence can be pursued as a potential research directions.

[1]  Kuolin Hsu,et al.  Improved streamflow forecasting using self-organizing radial basis function artificial neural networks , 2004 .

[2]  Adam P. Piotrowski,et al.  Optimizing neural networks for river flow forecasting – Evolutionary Computation methods versus the Levenberg–Marquardt approach , 2011 .

[3]  Ayalew Kassahun,et al.  Quality assurance in model based water management - review of existing practice and outline of new approaches , 2005, Environ. Model. Softw..

[4]  Faming Liang,et al.  Explicitly integrating parameter, input, and structure uncertainties into Bayesian Neural Networks for probabilistic hydrologic forecasting , 2011 .

[5]  K. P. Sudheer,et al.  Methods used for the development of neural networks for the prediction of water resource variables in river systems: Current status and future directions , 2010, Environ. Model. Softw..

[6]  K. Beven.  Rainfall-Runoff Modelling: The Primer , 2012 .

[7]  Giulia Napolitano,et al.  An exploration of neural networks for real-time flood forecasting , 2011 .

[8]  Simon Li,et al.  Uncertainties in real‐time flood forecasting with neural networks , 2007 .

[9]  Ron Kohavi,et al.  A Study of Cross-Validation and Bootstrap for Accuracy Estimation and Model Selection , 1995, IJCAI.

[10]  Peter A. Vanrolleghem,et al.  Uncertainty in the environmental modelling process - A framework and guidance , 2007, Environ. Model. Softw..

[11]  Chandranath Chatterjee,et al.  Uncertainty assessment and ensemble flood forecasting using bootstrap based artificial neural networks (BANNs) , 2010 .

[12]  Francesco Serinaldi,et al.  Impact of EMD decomposition and random initialisation of weights in ANN hindcasting of daily stream flow series: An empirical examination , 2011 .