Gas Turbine Anomaly Prediction using Hybrid Convolutional Neural Network with LSTM in Power Plant

Abstract

The fault and anomaly of real-time performance gas turbine data are difficult to predict because of the complexity of feature data and dynamically time series. In the case of real performance gas turbine, the complexity of the physical model is hard to interpret. In deep learning, the Convolutional Neural Network (CNN) is used to perform the identification of data with great feature extraction. But, since CNN is poorly accurate for time-series data, the prediction for gas turbine anomaly could be hardly optimized. Another neural network method that can interact with time-series data is Recurrent Neural Network (RNN), especially, the Long Short-Term Memory (LSTM) that can deal with the vanishing gradient problem in traditional RNN. This paper aims to develop hybrid CNN-LSTM as a proposed method to predict gas turbine anomaly more accurately than single CNN. The accuracy of the single CNN method is 81.33%. With the addition of LSTM in the same CNN architecture, the accuracy of hybrid CNN-LSTM is 91.79%. The accuracy of model data is significantly increased by adding LSTM layer after the convolutional and pooling layer.