Optimisation des contraintes mécaniques au démarrage des trubines hydroélectriques à l'aide de mesures indirectes

Abstract

Estimation of mean level and variation range of mechanical stress in hydraulic turbine blades are essential to obtain an accurate assessment of its residual fatigue reliability. Also, small changes in the control parameters during transient regimes (e.g. start-up and shutdown) can significantly affect the behavior of the mechanical stresses on the blades. Consequently, optimized start-up sequence can minimize the strain on the turbine, enhance its reliability under fatigue and increase life expectancy. In order to find the control pattern which optimises the strain on the blades during start-ups, operators run series of tests measuring directly, on the turbine, the magnitude of the strain during the transient for different wicket gates opening patterns. The goal of this work is to demonstrate that instead of measuring the strain during every start-up, once a consistent and accurate dynamic model has been identified, one can predict the strain on the turbine and use those predicted signals to find the optimal start-up. The method used consists of identifying a model representing the turbine with one start-up, validating the model using other star-ups, and then using the model to predict the strain and find an optimal start-up using indirect measurement. The model is identified using signals of at least one strain gauge combined with the signal of other sensors strongly correlated to the strain and located on the parts surrounding the runner (such as torsion gauges on the shaft) and can be used as a more economical and specially more robust alternative than strain gauge. As an example, sensors on the shaft which are not in contact with water cost less to instrument and will last longer in an extended measurement campaign. In order to choose the best model set and then choose the best model in that set, a thorough comparison was done between ARX, ARIX, ARMAX and ARIMAX model sets. For each model, several orders were tested. Using the root mean square error, the ARMAX model appears to represent the dynamic system at best. The model’s output (predicted signal) is then compared to the measured signal from the strain gauge. For further validation purposes, a comparison is also done between the predicted signals and the measured signals for the other start-ups.

Date	31 May 2016
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Antoine Tahan (Supervisor) & Martin Gagnon (Co-supervisor)