Power plant turbine power trend prediction based on continuous prediction and online oil monitoring data of deep learning

Abstract

Power output is an important property of steam turbines and more accurate trend prediction is essential for understanding the operation of power plants and anomaly detection in equipment. Equipment power is strongly influenced by human factors and achieving accurate trend prediction is often impossible. In the current research, a continuous prediction model was developed on the basis of deep learning to establish the relationships among oil online monitoring parameters and equipment power. The model used long short term memory (LSTM) method to develop a trend prediction model for oil wear state parameters. Trend prediction results were applied as a test set to establish a deep belief networks (DBN) prediction model for predicting device power. In modeling process, continuous prediction model was optimized via feature selection method on the basis of ridge regression with L2 regularization, recursive feature elimination (RFE) and differential evolution (DE) algorithms for the elimination of subjective factor to decrease cumulative error of forecasting. Comparative experimental results showed that LSTM-RFE-DE-DBN continuous prediction model outperformed LSTM-RFE-DE-BPNN and LSTM-DBN. The developed model realized continuous prediction and applied lubricating oil wear status with objective factors to perform the power prediction of power plant turbines with subjective factors.