On the use of neural networks for dynamic stress prediction in Francis turbines by means of stationary sensors. (June 2021)

• Record Type: Journal Article
• Authors: Presas, Alexandre; Valentin, David; Zhao, Weiqiang; Egusquiza, Mònica; Valero, Carme; Egusquiza, Eduard
• Is Part Of: Renewable energy. Volume 170(2021)
• Page Start: 652
• Page End: 660
• Publication Date: 2021-06
• Subjects: Hydropower -- Francis turbine -- Fatigue -- Stress -- Neural network -- Reliability; Renewable energy sources -- Periodicals; Power resources -- Periodicals; Énergies renouvelables -- Périodiques; Ressources énergétiques -- Périodiques; 333.794
• DOI: 10.1016/j.renene.2021.02.013
• Languages: English
• ISSNs: 0960-1481

Abstract

Abstract: Nowadays, one of the major mechanical issues of hydraulic turbines and particularly Francis turbines are the failures produced by fatigue. Due to the massive entrance of new renewable energies such as wind or solar, hydraulic turbines have to withstand off-design conditions and multiple transients, which greatly increase the risk of fatigue failures. Fatigue damage and crack propagation models are based on the static and dynamic stresses on the turbine blades. Therefore, an accurate and realistic determination of these stresses is of paramount importance although it is yet a challenging task. Numerical simulations have still limitations when predicting static and dynamic stresses in the most harmful conditions such as deep part load conditions and transients, which are highly stochastic. The installation of strain gauges on the blades gives accurate stress measurement but it involves long and very expensive measurement campaigns, as the turbine runner is submerged, confined and rotating. In this paper we propose a neural network-based method to determine the magnitude of static and dynamic stresses based on the measurements of stationary sensors, which greatly reduces the complexity and costs of strain gauge testing. Inputs of the neural networks are selected based on previous experience monitoring Francis turbines. The trained network can be implemented in advanced monitoring systems that could continuously evaluate the stress level of the turbine and determine the risk of possible fatigue damage. Highlights: Development of a method to predict stresses in a turbine with stationary sensors. Method uses relevant signal indicators as inputs of an Artificial Neural Network. The most relevant stationary sensors for the prediction have been determined. Fatigue assessment of the turbine can be performed. Complex strain gauge testing in Francis turbines could be simplified.