Hydro-climatologie globale pour la prevision des crues du nil au moyen de fonctions de transfert avec bruit et de reseaux de neurones artificiels (French and English text).

摘要

The principal objective of this research is to improve forecasting models of the cumulative volume of natural inflows entering the large reservoir of the High Aswan Dam, located on the Nile River in Egypt. This has been accomplished by statistical/stochastic modeling of the teleconnection between the natural inflows resulting from precipitation in tropical regions, and the indices of climatic variability. Two forecasting models have been built: the first is based on a transfer function with noise (TFN) while the second exploits artificial neural networks (ANN). Inputs to both models are the sea surface temperatures (SST) in specific regions as well as the cumulative volumes of natural inflows of previous years. The forecast is implemented with a three-month lead-time before the occurrence of the Nile flood peak; this enables a better planning of the future monthly withdrawals from the reservoir.; The results obtained from the models presented in this thesis are very satisfactory and appear to be significantly superior to those obtained from previously published or practically implemented models. The models explain up to 63% of the streamflow variability, with correlation coefficients between forecasted and observed streamflows exceeding 0.85. Mean absolute percentage errors are typically of the order of 6%.; The first aspect is related to a better choice of the predictor of the flood, which is based on recent climatological studies. Two indices of the climatic variability are used: The first is representative of the phenomenon coupling the El-Niño ocean current with the Southern Oscillation (ENSO). The second, which for the first time has been exploited within the framework of this research, is obtained by averaging the SST anomalies in a specific region of the Indian Ocean. The use of this variable, which turns out to be a good predictor of the Nile flood, allows the refinement of the forecasts obtained with models where only the ENSO index is used as the exogeneous variable.; The second innovative aspect concerns the choice of more appropriate models related to the streamflow forecasting using climatic predictors. Transfer functions with noise (TFN) and artificial neural networks (ANN) are used for the first time to directly forecast streamflows using climatic indices. The forecasting performance of these models is markedly superior to those of linear regression models commonly used in teleconnection studies between streamflows and climatic indices. (Abstract shortened by UMI.)