Reservoir inflow forecasting using differential evolution trained neural networks

This paper presents a study on the application of evolutionary computation and artificial intelligence techniques to forecast inflows into the Vanderkloof reservoir, South Africa for the purpose of planning and management of available water resources. A differential evolution (DE)-trained neural network (DE-NN) was developed to simulate the interaction between reservoir inflow and its causal variables such as precipitation and evaporation. The performance of the DE-NN was evaluated using two performance metrics namely mean absolute percent error (MAPE) and coefficient of determination (R2). Results from this study demonstrated that the DE-NN model was able to substantially represent inflow patterns into the Vanderkloof reservoir, thereby indicating the efficacy of the DE algorithm in producing adequate generalization on unseen datasets. These results further showcase differential evolution algorithm as a potent, viable and promising algorithm for training neural network models for use in the field of water resources management. © Springer International Publishing Switzerland 2014.