Machine learning for condition monitoring in hydropower plants using a neural network
Abstract: The hydro power industry stands for new challenges due to a more fluctuating production fromwind and solar power. This requires more regulation of the production in the hydro powerstations, which increases maintenance demands. An oil leakage has not only consequencessuch as downtimes and maintenance costs, but also an environmental impact. Skellefte ̊aKraft is working towards reaching a condition based maintenance. Therefore, the purpose ofthis master thesis is to develop a model using a feedforward neural network to predict the oillevel in the control system of a Kaplan turbine and map which sensor signals that are required.The thesis will cover data from two hydro power stations, Grytfors and B ̊atfors, each ofwhich has two units, G1 and G2. Due to limitations of the database Skellefte ̊a Kraft areusing, the data has minute resolution and covers two months, December and January. Themodel is developed in MATLAB using their Deep Learning toolbox and the neural networkfeedforwardnet. Before training and testing the model, an optimization was done. Grytforshas a full range of sensor signals while B ̊atfors has half the amount and therefore, the datafor Grytfors was used in the optimization. A grid search was done to optimize the hyperpa-rameters using cross validation. To map which input parameters that are required a featureselection was done.From the result of the feature selection, power, accumulator levels 1 and 2 and pressurewere chosen as the input parameters for Grytfors. For B ̊atfors, all of the the existing sensorsignals were used instead. The model is then trained and tested for the two different powerstations. For Grytfors, the predicted oil level follows the pattern of the real oil level but thetest error is around 15-20 liter. Four different tests were done for B ̊atfors. The two firstfor unit 1, the third for unit 2 and the fourth on both units to investigate the potential of ageneral model for one power station. For B ̊atfors, the first two tests have test errors of around4-6 liters. The third and fourth tests have test errors of around 1.5 liter. In the first twotests, the December data contains a potential refill sequence and in the third test, for unit2, the data contains start and stop sequences. The results showed the importance of havingcomprehensive training data.
AT THIS PAGE YOU CAN DOWNLOAD THE WHOLE ESSAY. (follow the link to the next page)