Application of the neural network technique in the estimation of inflows into the Mtera and Kidatu reservoirs.

The Mtera and Kidatu Reservoirs are located in central Tanzania, within the Rufiji Basin, and they supply water to generators that are responsible for 85% of power production in the country. These reservoirs are of significant importance and they require efficient management. Sound reservoir operation and management require planning which involves the evaluation of the ability to meet current and future demands of the purposes which they serve. This ability is best determined with knowledge of the inflow and the hydrologic losses. Between the years 1991 any 1993. water levels at the Mtera reservoir fell very low, close to the minimum supply level of 690.5 meters above sea level. These were years of prolonged drought and is alleged that the low levels were as a result of decline in the inflow. This study is an investigation into this allegation and is centered on the application of the Neura Network technique in the estimation of these inflows. Assessment of the suitability of the Neural Network model in the estimation of inflow was made by comparing its performance with other river flow modeling techniques i.e., Baseline model, Multiple Input Simple Linear Model and Multiple Input Linear Perturbation Model. Model calibration was done using historical records of flow in the main rivers that supply water to these reservoirs. Comparison of the models was made using the model efficiency criteria and the index of volumetric fit. The Neural Network Model performed better than the other model having efficiencies of over 80% in the estimation of inflow. The Neural Network model also preserved the observed volumes quite well having index of volumetric fit values close to unity. Based on these results .the Neural Network Model was considered to be a suitable model for the estimation of inflow into the Mtera and Kidatu reservoirs and was used for that purpose. Estimation of losses from the reservoir was then carried out using the water balance approach. The results were inconclusive, as some of the estimates were negative. Investigation into the problem revealed many errors in the water level and outflow data which strongly suggested that poor quality data was responsible for the inability to trace water loss using the water balance approach.