Analyse de la valeur des composantes de la chaîne de prévision hydrologique en contexte de gestion de réservoirs hydroélectriques

Abstract

Hydrological forecasting has become an essential tool for water management, and amongst other things in a context of hydropower reservoir management. Through time, hydrological forecasting went from deterministic (the best possible forecast) to probabilistic (estimates of a range of possible outflows) to account for uncertainty present in the various elements composing the hydrological forecasting chain. However, despite the economical and social importance of water management and the efforts and resources spent to improve its performance, the relative value of the various elements composing the hydrological forecasting chain is not well defined. This gap in the knowledge base concerning the value of these elements, when compared to one another, does not allow for researchers and hydropower managers to allocate time and resources optimally to improve the overall value of the hydrological forecasting chain. This study aims to establish the relative value of five elements of the hydrological forecasting chain in a context of hydropower reservoir management, namely the hydrological model, the hydrometeorological input dataset, the objective function for model calibration, the calibration score, and the bias and dispersion present in the members of the ensemble. To achieve this goal, a full factorial experiment design has been developed, in which various ensemble streamflow predictions are being generated by changing the elements of the hydrological forecasting chain one by one. These ensembles are then passed on to a test bench that was provided by our industrial partner, Rio Tinto, that simulates a simplified version of the Lac-Saint-Jean reservoir and the Isle-Maligne hydropower station, in Québec, Canada. The test bench includes a decision-making optimization algorithm that manages the hydropower reservoir based on the management rules similar to those used by the industrial partner. To determine the relative value of the studied elements, the average generated profits are regrouped by elements (or combination of elements), and their variance is compared. To assess if the results can be generalized to other systems, the exercise has been repeated and the test bench’s parameters, such as the buyout cost of energy, the storage capacity of the reservoir, and the presence or absence of a minimal energy production, have been modified. Results of the current study have shown that for this system, the studied elements have similar values and that the interactions between them explain most of the variance. To improve the overall performance of the system, multiple elements must be improved at once. In fact, interactions between three and four elements represent 71.2% of the total variance, while the main effects (without interactions) have a marginal impact. Also, the results of this study show that the impact of bias and dispersion on the average generated profit’s variance is almost zero. However, a third of the variance can be explained by three elements, namely the calibration score, the hydrometeorological dataset and the objective function. Beyond the actual results, this study proposes a framework that will allow doing an exhaustive analysis of the value of the elements of the hydrological forecasting chain, and also of the whole hydropower reservoir management chain, for systems of varying properties and complexity.

Date	19 May 2022
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Richard Arsenault (Supervisor)