Abstract:
One of the most frequent natural disasters in the world today is flooding; which occurs when a body of water exceeds its capacity due to high rainfall over a prolonged period of time. This results in increased global climate change and the balance of rainfall–runoff mechanisms used by hydrologists. Also, this leads to more extreme frequent occurences of hydrological events than those in previous years. To reduce the impact of these flood hazards, the development of early warning system for flood disaster is necessary. Recurrent Neural Networks (RNNs) have been widely used in solving a wide range of hydrological problems, such as rainfall–runoff modeling. However, it comes with some challenges in modelling long-term dependencies such as vanishing and exploding gradient problems. In examining the non-stationary and non-linearity features of the River discharge process, lots of models are being developed in order to learn patterns about such complex phenomena. In this research, a Long Short-Term Memory (LSTM) network was used to develop a model that predicts daily discharge of Lokoja River in Nigeria. The model was deployed to learn the river discharge process based on flood discharge events from 2010 to 2019. The experimental data were 3651 discharge instances. A total of 2556 (70%) of the discharge events were used as training sets and the remaining 1095 (30%) discharge events were used as the testing set. The result shows that the model is suitable for river discharge prediction and better than the physical and conceptual based models. The performance of the model gives an accuracy of 97.075%, Means Square Error (MSE) of 0.017, Root Means Squared Error (RMSE) of 0.085 and the Means Absolute Error (MAE) of 0.0173. The result of the LSTM model performance was compared with the performance of the traditional neural network and physical hybrid neural network models that combine the Self Organized Map (SOM) and Back Propagation Neural
Networks (BPNNs) using the same datasets. The results show that the LSTM performs better than the traditional neural network and the physical hybrid neural network models
