Reservoir Level Forecasting using Neural Networks: Lake Naivasha

Abstract

Six feature groups comprising of water levels, rainfall, evaporation rate, discharges for rivers Malewa and Gilgil and one pair of time harmonics were used to develop neural network models to forecast water levels for Lake Naivasha in Kenya. Six elements were used from each feature group. Some feature groups were compressed using the Karhunen–Loeve Transform (KLT) to reduce their dimensions. The neural network models developed were able to forecast effectively the reservoir levels for the lake for four consecutive months after a given month and given data for six consecutive months prior to the month. It was found that the more the number of feature groups used, the higher the ability of neural networks to forecast accurately the reservoir levels. Data compression generally reduced the size and computation time of the models.