الفهرس | Only 14 pages are availabe for public view |
Abstract ”f I > Abstract Forecasting of the annual quantity of available water has always been a problem, not only for the Nile River, but for all rivers of the world. This forecasting is crucially essential for project planning, dam operating rules, and optimum utilization of the river water. Inflow forecasting is a key input to the process of reservoir management. Accurate forecasts increase the benefits such as more energy generation, more effective drought mitigation. and better flood protection. owadays, the Artificial Neural Networks (ANND have received a great deal of attention and are being touted as one of the computational tools ever developed. It has been applied widely in different fields. Because of its ability to imitate human-like brain functions, ANNT can potentially constitute powerful tool for hydrological forecasting. Therefore, this thesis is directed to utilize the available seasonal univariate neural network forecasting model in the multivariate dimension. In other words, the available seasonal ANNT forecasting model is based only on the previous natural inflows at Aswan. However, in the multivariate case, the natural flow at Aswan is related to the natural flow at the main upstream stations, Malakal, Khartoum and Atbara as well as at Aswan itself. The flows at the upstream stations along with the previous flows at Aswan have been used to train the multivariate model to forecast the current natural flow at Aswan. The contribution of the past inflows of the three upstream stations to forecast the inflows at Aswan has been considered in three different manners. The idea is, for every upstream station, how many months in past should be incorporated to predict the current inflow at Aswan. Both univariate and multivariate neural network models have been trained over the same historical period. For both models, validation using the data of recent years, has been done. The thesis contains description of the major characteristics of the Nile River Basin, specially its main watersheds, tributaries and discharges at the different gaging stations. It also reviews the previous studies related to the applications of the ANNT in hydrology. The thesis briefs the concepts of Artificial Neural Networks. Different architectures, transfer functions, single and multiple layers of neurons, and training procedures are also high lighted in a brief way. Finally, the performance of the neural network ability in forecasting drought, flood, and average year against some other techniques namely; Corridor and Regression, have been tested . and presented. As a global conclusion and results of the study, the seasonal neural network models have many parameters but can potentially generate the most accurate forecasts. However, their performance depends on the stage of training and, perhaps, on the size of the training data set. The discrepancies found between the desired and simulated output by the univariate and multivariate ANNT models should not be attributed only to the multivariate approach; the data sets used should take some (if not all) of the blame. The thesis recommends to carry out further research to investigate and compare results using different neural network models structures. Another model structures with different learning rates, hidden nodes, and number of layers are needed to be studied. |