Rainfall-Riverflow Modelling for Flood Forecasting in Oyun River Basin using Artificial Neural Network and Hec-Ras

Abstract

This study deals with rainfall-riverflow modeling using Artificial Neural Network (ANN) and Hydrologic Engineering Centre’s River Analysis System (HEC-RAS) for flood forecasting. ANN is a multiprocessor computer system with Simple processing neurons, a high degree of interconnection, simple scalar messages and adaptive interaction between neurons. ANN architecture and transfer functions have become as essential as the network training algorithm. HEC-RAS is a hydraulic model that computes water surface profiles and maps the flood inundation area by solving hydraulic equations of open channel flow in a downstream direction. The aim of this study was to develop a rainfall-riverflow model using ANN and HEC-RAS models for flood forecasting in Oyun River Basin, Kwara State. The specific objectives were to: develop a rainfall-riverflow models of Oyun River using ANN; optimize the developed model using finite neuron pattern architecture; forecast one to seven lead day of Oyun riverflow using best performing neuron pattern architecture, and map the inundation area using HEC-RAS model. The hydro-meteorological data used was collected from the Meteorological Unit under the Land and Water Engineering Department of the National Centre for Agricultural Mechanization (NCAM), Ilorin. located about 20 km from Ilorin the Kwara State capital. It has an estimated average terrain elevation of 470 m above sea level and lies between Latitude 9o50ˈ and 8o24ˈNorth and Longitudes 4o38ˈ and 4o03ˈ East. The hydro-meteorological data were divided into three sets: the training set, the validation set and testing set. The training set was based on a historical data consisting of 80% of the total data, the validation set was based on historical data consisting of 13% of the total data, the test set was based on 7% of the total data, two other validation data sets (extreme and out-of-range data sets) were employed to evaluate the ANN capability of modeling uncertainty flood at extreme and out-of-range values. A total of six artificial neural network candidate models were trained, validated and tested. Each of the classes has three candidate sub-models with the first model having only one hidden layer, the second model having two hidden layers, and the third model haven three hidden layers. The optimization results showed that the Quadrupled Feed-Forward Architecture (QFFNNA) model with 3 hidden layers performed best with R2 value of 99.80%, 97.55% and 98.14% both for model training, validation and testing respectively, than other ANN models. The Doubled Feed-Forward Neural Network Architecture (DFFNNA) models had R2 values that ranged from 81.59% to 97.62%, 79.96% to 95.47% and 80.92 to 94.09 for model training, validation and testing, respectively. The MSE and RMSE values ranged from 0.03 to 0.27 and 0.17 m3/s to 0.52 m3/s, respectively. Therefore, the simulation mode from QFFNNA was used for artificial neural network non-linear autoregressive exogenous moving (ANN-NARXM) average updating procedure to forecast one to seven lead day forecast of Oyun River. It was further observed that the R2 values of the ANN-NARXM decreases as the lead day forecast increases. However, the R2 values of the ANN-NARXM model were unstable after the first three lead day riverflow forecast. A stand-alone model interface coupling between the ANN and HEC-RAS models was developed. The parameters of the HEC-RAS model were calibrated and used to map the flood inundation area of Oyun River basin. The HEC-RAS results shows the flood inundation area of 1121.7 m2 with flood water depth that varies from 0.6 m to 2.8m. The results obtained from this study will provide large-scale information on the optimization of artificial neural network model using finite neuron pattern architecture for the rainfall-riverflow modeling which will serve as a guide to governments and agencies for better policy and decision making for flood control and water resources management.