River flow forecasting : statistical models research

River flow forecasting is a widely studied topic, mixing together hydrological and mathematical tools to obtain reliable river flow predictions. This thesis work is based on Hydromax, a real-time river flow forecasting application that uses conceptual and statistical models. The conceptual model provides a basis for the hydrological physics, and computes the part of the entire rain volume that directly impacts the river system. The statistical model uses as input the effective rainfall provided by the conceptual model, and computes river flow predictions with no regard to hydrological processes. It can also calculate long-term predictions based on rainfall forecasts. This thesis explores new statistical methods, with an approach involving the experimentation of new river flow and rainfall transformations in order to reach a better accuracy for river flow forecasts. The performance of each new model is reported with quantitative and practical criteria, and compared to current models. Finally, an optimal model is presented. The latter is a mixed model combining linear and squared shifted logarithmic river flows on one side, and shifted logarithmic rainfalls on the other. For an improved river flow forecasting, the choice of the best model should depend on the river regime and the forecasting horizon. Models based on transformed river flows will enable a better forecasting accuracy when confronted with low river flow conditions. For long-term forecasting, especially when rainfall forecasts are perturbed with measurement noise, models using transformed rainfalls will be preferred.