With a total installed capacity of 37.2 GW, Hydro-Québec is the largest electricity producer in Canada. More than 99% of that power is generated from renewable sources, but the immense majority comes from water: Hydro-Québec comprises 61 hydroelectric generating stations as well as 28 reservoirs with a combined storage capacity of more than 176 TWh. Keeping these assets in prime condition is essential in order to be able to provide customers with reliable clean energy, but this represents significant expenses.
Transitioning from preventive to predictive maintenance is a promising solution to drive down costs. In this sense, Hydro-Québec is seeking ways to leverage the vast amounts of sensor data it is already producing and storing on OSIsoft PI servers. The development of a digital twin prototype for the speed governor of a selected generating unit is part of this ongoing effort. This digital twin will consist of models connected to the data streams that will automatically run in deferred time following certain events (start-up, set point change, etc.) and provide additional decision support regarding maintenance needs.
Firstly, a Simulink model capable of simulating the nominal behaviour of the actual speed governor will enable fault detection through the comparison of real and simulated signals. Alerts will be logged when predefined thresholds are exceeded. Secondly, a classification model will be created using the MATLAB Classification Learner App. This will complement the fault detection aspect with a diagnosis, allowing to identify which defect is present. The classifier will be trained with synthetic data generated from the Simulink model, to which defects will have been introduced beforehand. This approach will allow to overcome the lack of data in the presence of defects, which are relatively rare in speed governors. Both models will be deployed as dynamic-link libraries (DLL) using MATLAB Compiler and Simulink Compiler.