A Neural Network for Flood Prediction: A Case Study

Ednah Onyari and Masengo Ilunga


Flood modelling, artificial neural network, extreme events, minimum Redundancy Maximum Relevancy


Reliable estimation of streamflow is important in water resource planning and management. This is particularly fundamental in the management of extreme events like floods and droughts. This article investigates the influence of input variables in predicting floods using a multi-layer perceptron neural network (MLP-ANN). The recurrent method that employs antecedent flows and rainfall is used to predict streamflow. Three models are developed, trained and validated for the Olifants River in South Africa. The variable inputs into Model 1 were only antecedent streamflow, Model 2 had antecedent streamflow and rainfall, and Model 3 had selected antecedent streamflow and rainfall using the minimum Redundancy Maximum Relevancy (mRMR) method. The available data were split into training, testing, and verification. The results show that Model 3 predicted streamflow reasonably well as compared to Model 1 and 2. A flood event that occurred in the year 2000 was used to indicate the accuracy with which new events could be predicted. The event was predicted with a correlation coefficient of 92% by Model 3. It is concluded that MLP-ANN can reliably predict streamflow and input variables determine model performance.

Important Links:

Go Back