Deep Learning for Prediction and Fault Detection in Geothermal Operations

Abstract

Automation, control and real-time surveillance of geothermal power plants require robust and accurate predictive models. While physics-based models require complex modeling, calibration and computational efforts, advances in cost-effective monitoring data acquisition systems have motivated the development of data-driven predictive analytics as a flexible, efficient, and easy-to-deploy alternative for real-time application. We present a dynamic neural network model for performance prediction, control and real-time monitoring of geothermal power plants. Measurements collected from geothermal power plants are often high dimensional, highly cross-correlated and auto-correlated. In addition, the measurements are often influenced by disturbances and subsequent adjustments of operational settings. Therefore, we present a neural network architecture that is able to represent the dynamics and driving forces behind the data in a low dimensional latent space, while taking into account the system responses affected by control changes. Our developed model consists of three parts: (i) an sliding-window encoder for capturing data correlations to enable latent-space representation, (ii) fully connected neural network layers tailored for describing evolution of the latent states, accounting for effects of control changes and making predictions, and (iii) a decoder which maps the prediction results in the low-dimensional latent space back to the original data space. The resulting dynamic neural network architecture is trained to maximize prediction accuracy. Once the model is trained, it can be used for future predictions of important time-varying responses of the power plant based on observed measurements and referenced control settings. The prediction results can be used for applications such as model predictive control and fault detection. We apply our proposed neural network model to field data from a binary cycle geothermal power plant to demonstrate its prediction and fault detection performances.