On the Use of Artificial Neural Networks for Condition Monitoring of Pump-Turbines with Extended Operation

Abstract Because of the addition of stochastic supply to the power grid introduced by new renewable energies, hydropower is required to respond rapidly to the fluctuation between power demand and supply. Consequently, hydropower units must work under extreme off-design conditions, where the machines are more prone to suffer from damages and shorter useful lives. Novel supervision and monitoring techniques which are able to compare the revenues with the remaining useful life of the turbine unit are required to cope with these new scenarios. In this paper, the upgrading of an existing monitoring system to deal with the extended operating range of a pump-turbine is discussed. Previously, the machine operating range was from 50% to 100% power and now it is from 20% to 100% power. For that purpose, the vibration signals collected from the current monitoring system have been used. First, the autoregressive mapping of the overall levels measured in the machine for all the extended operating conditions has been carried out. Back-propagation neural network was applied for the mapping. Second, the complex hydraulic phenomena that may occur in the extended operating range that can produce accelerated wear and tear have been studied. Typical phenomena are excessive turbulence, draft tube vortex rope, cavitation erosion, excessive vibration and excessive stresses in the runner. For each of these abnormal operating conditions, several features (condition indicators) were selected and mapped on the operation hill-chart using neural networks. The consequences of each abnormal operation have been analyzed with physics-based models. With the mapping, the zones where operation is not recommended can be identified and the effects estimated.