Fatigue damage reduction in hydropower startups with machine learning

Abstract

As the global shift towards renewable energy accelerates, achieving stability in power systems is crucial. Hydropower accounts for approximately 17% of energy produced worldwide, and with its capacity for active and reactive power regulation, is well-suited to provide necessary ancillary services. However, as demand for these services rises, hydropower systems must adapt to handle rapid dynamic changes and off-design conditions. Fatigue damage in hydraulic machines, driven by fluctuating loads and varying mechanical stresses, is especially prominent during the transient start-up of the machine. In this study, we introduce a data-driven approach to identify transient start-up trajectories that minimize fatigue damage. We optimize the trajectory by leveraging a machine learning model, trained on experimental stress data of reduced-scale model turbines. Numerical and experimental results confirm that our optimized trajectory significantly reduces start-up damage, representing a meaningful advancement in hydropower operations, maintenance, and the safe transition to higher operational flexibility.

Fig. 1. Overview of the workflow steps.

a iTF-DNN is a deep learning model trained on the first dataset of start-up trajectories to predict stress. b The optimal start-up trajectory search is based on the damage estimated from the model’s stress prediction. c The optimized trajectory is tested in the reduced-scale model. d Final validation of the optimized trajectory in terms of the damage.

Fig. 2. iTF-DNN components.

a The total stress prediction combines the predictions of the trend stress model with the model for oscillations. The trend part of iTF-DNN is trained from the observed measurements. The residuals (b) of the trend part of iTF-DNN are transformed with an STFT to the time-frequency domain (c) and further used to train the oscillation part of iTF-DNN.

Fig. 3. Comparison of damage predictions from iTF-DNN with damage based on measurements.

Damage for sensor 1 (a) and sensor 2 (b) computed based on the real measurements and predicted from the iTF-DNN. The model was trained on the Classic, 2Slopes, and Linear trajectories, while the BEP trajectory was excluded from the training data.

Fig. 4. Phase-space plots of the optimization setup and resulting optimized trajectory.

a Data from the first experimental campaign used to train iTF-DNN, along with the disallowed region defining the constraint applied in start-up optimization. b The final optimized trajectory, shown alongside the stress model uncertainty.

Fig. 5. Optimized and existent trajectories collected during the second experimental campaign.

Comparison in the phase space (a) and over time (b).

Fig. 6. Start-up data of the first experimental campaign.

Summary of the data used for training and validation, plotted by start-up type.