Online Time Series Prediction and Fault Detection for Geothermal Power Plants Using Recurrent Neural Networks

Abstract

Reliable prediction models facilitate performance monitoring and fault detection in geothermal power plants. With the measurements collected from power plants, data-driven models such as recurrent neural networks (RNN) can be trained learn the power plant dynamics and predict its performance. A common approach to apply RNN for prediction is to train a model offline using historical data. However, the dynamic relationship between measurements is not time-invariant due to maintenances and other variations in the system, such as changes in brine split between heat exchangers. As a result, using a pre-trained neural network for prediction can lead to poor performance under dynamically changing and non-stationary environments, and thus degrade the fault detection performance. To deal with this problem, we present a workflow that first detects changes in operating conditions from data using the dynamic inner canonical correlation analysis (DiCCA) and then adaptively updates a long short-term memory (LSTM) encoder-decoder neural network model to learn the new dynamics. We apply the updated model to fault detection and demonstrate its performance using field data collected from a binary cycle geothermal power plant.