Thermal model parameter estimation of a hydroelectric generator using machine learning

Abstract

The synchronous generator have a central role in the power system, and is governed by multi-physics behaviour. A digital twin of the synchronous generator can lead to more efficient utilization in a more safe and reliable way. In this report, mechanistic thermal models of a synchronous generator and a heat exchanger will be defined in order to assess the feasibility of parameter estimations of mechanistic models using machine learning and recurrent neural networks (RNNs). Skagerak Energi provides with operational data that has been used with the mechanistic models and machine learning algorithms in this project. The data was collected using Azure Databricks, from a 12 MVA synchronous generator located at Grunnaai. The results from parameter estimations shows that a recurrent neural network has the ability to estimate the three model parameters defined in the heat exchanger model with great success. Optimization of the model parameters with respect to prediction error showed similar results, with the machine learning approach being slightly more accurate. The neural network could not find good parameter solutions to the six model parameters in the generator model. However, the RNNs consistently predicted parameters in a narrow range of values. This may indicate that the RNN was trained on training data not representative to the measurements. The optimization with respect to the prediction error showed great accuracy in model parameter estimation.