A Dynamic Deep Learning Model for Performance Prediction and Fault Detection in Geothermal Power Plants

Abstract:

Automation and fault detection in geothermal power plants require reliable and accurate predictive models. While physics-based simulation models are quite useful and have traditionally been applied to predict the behavior of geothermal power plants, they are typically too complex to construct, involve many uncertain parameters, and are time-consuming to run. Data-driven models present an attractive alternative for leveraging the measured data in instrumented plants to derive flexible and efficient predictive tools for real-time applications, including automation and fault detection. Recent advances in data science and predictive analytics have led to a host of promising tools for extracting complex patterns from data and using them to develop statistical predictive models.

We present a dynamic deep learning model for performance prediction, real-time fault detection, and control (optimization) of geothermal power plant operations. The developed model consists of three components: (i) an encoder for capturing data patterns and correlations to enable dimensionality reduction and latent-space representation, (ii) a latent-space dynamical model based on fully connected neural networks to compactly describe the evolution of the latent states, and (iii) a decoder that maps the resulting predictions in the latent space back to the original data space. The developed dynamic neural network architecture is used to capture important time-varying response of power plants based on past measurements of several variables, including the control variables and the corresponding system responses. The encoder and decoder structures serve as mapping functions between the original data space and the low-dimensional latent space, while the latent space is used to capture the system dynamics using its dominant modes. The weights in the resulting dynamic neural network architecture are trained using collected time series data from a real power plant. Once trained, the model is used to predict the dynamic response of the power plant, which can be used for efficient model predictive control and fault detection applications. We present the developed workflow in detail and demonstrate its performance using field data from binary cycle geothermal power plants.