FLOOD FORECASTING USING ARTIFICIAL NEURAL NETWORK MODELLING APPROACH

Abstract

Runoff forecasting is a complex, complicated and a significant hydrological phenomenon. The accurate prediction of runoff helps in the planning, designing and management of water resources, especially irrigating, producing hydropower, managing food risk, and protecting dams. The study presents the rainfall-runoff modelling using the multi-layer perceptron based artificial neural network. Artificial neural network was developed with minimum variables examined and dataset of 10 years for the Leaf River in Mississippi, US. The model also normalizes the input variables and uses the back propagation algorithm to train the network. Research study focus on the analysis of the different cases and efforts to optimise the working model. First case deals with a static network that develops with a fixed training to forecast the future analysis. Second case deals with a constant up-gradation of network model using the recent past data. In this, study analysis the performance metrics for one day, two day, three day and four day lead time. At last, performance of the developed artificial neural network models is evaluated using different relative indicators like root mean squared error (RMSE) and coefficient of determination (R-squared). The study shows the static network model has very poor performance while after incorporating the lead time effects the performance of the network improves drastically. And, with the increase of lead time model performance decreases. With the increase in lead time, the immediate condition are not known as a result it predicts with older known values which decreases the performance. The study concludes that the artificial neural network model developed for lead time of one day can precisely predict the daily runoff with 0.8348 and 0.8131 value of coefficient of determination (R-squared) for training 9 | Page WRDM, I IT Roorkee

phase and validation phase. Fourth case of our study compares the artificial neural network model with other statistical technique - multi-linear regression model (MLR). Artificial neural network model is found to be performed better for one day lead time with inputs P(t), P(t-1), P(t-2), P(t-3), P(t-4), PET, S(t-1), S(t-2), S(t-3) and S(t-4) for test size of 0.4