Machine learning for flood forecasting:case study: Nzoia river basin, western Kenya.
In Kenya property destruction and loss of life has occurred due to serious incidents of floods, along the Nzoia River catchment area Western Kenya. Despite having flood warning models along the Nzoia River basin; with a flood warning system at Rwambwa gauge station that sends out alerts on the river levels. These models are linear models and have overlooked the peak streamflows. A reliable intelligent nonlinear model that is capable of handling nonlinear estimation streamflow (discharge) problem is crucial in flood control operations. This research explores applicability and performance of flood forecasting models in the Nzoia River basin, Western Kenya, using two types of artificial neural network (ANNs), namely MLPANN-FF a feedforward multilayer perceptron (MLP) network and GA-ANN-FF a genetic algorithm optimized multilayer perceptron feedforward neural network model. The aim of this study is to compare the performance of these two models (MLP-ANN-FF and GA-ANN-FF) and recommend the most suitable for this problem. The historical daily rainfall, and average temperature and discharge flow, obtained from Kenya Metrological Department (KMD) were used as inputs to the two ANN models for discharge flow (streamflow) forecast for Nzoia River basin at Rwambwa river gauge. The characteristic parameters such as number of neurons within hidden layers and the selection of input variables for the MLP-ANN-FF were optimized using genetic algorithm (GA), hence yielding a GA-ANNFF model. These two models were trained, cross verified and tested with daily rainfall, average temperature, and discharge flow. The architectural topology that trained well on MLP-ANN-FF model was one with 9 input variables, 2 hidden layers and 1 discharge flow output; a 9:7:12:1 configuration setting. This was later optimized with the genetic algorithm (GA) to develop a GA-ANN-FF model that was able to optimize the input variables reducing them from 9 to 4 inputs, and reducing the number of neurons in the 2 hidden layers yielding a 4:6:4:1 GA-ANN-FF model. The conventional ANN (MLP-ANN-FF) and a GA-ANN-FF model were used as the benchmark 10% was used in testing the overall performance of the models. The results revealed that the GAANN-FF (4:6:4:1) model was able to yield better accuracy in performance for Nzoia River basin at Rwambwa River gauge, with least input variables, and number of neurons in the hidden layers though it took longer on the computation time. With a MSE of 0.021 and an r (correlation coefficient) of the desired and estimated discharge flow of 0.887 (89%), GA-ANN-FF performed satisfactory better than MLP-ANN-FF (9:7:12:1) with 9 input variables an MSE of 0.024 and r (correlation coefficient) of 0.84 (80%). The results showed that ANN integrated with GA has a better accuracy and therefore most suitable in developing flood forecast models with low MSE. This finding is important because it will eventually enable relevant agents in water resource planning and flood management and the public aware when a flood might occur and the areas that would be affected to avoid disaster caused by floods.