Deep Learning-Based Fault Detection and Isolation in Solar Plants for Highly Dynamic Days

Abstract

Solar plants are exposed to numerous agents that degrade and damage their components. Due to their large size and constant operation, it is not easy to access them constantly to analyze possible failures on-site. It is, therefore, necessary to use techniques that automatically detect faults. In addition, it is crucial to detect the fault and know its location to deal with it as quickly and effectively as possible. This work applies a fault detection and isolation method to parabolic trough collector plants. A characteristic of solar plants is that they are highly dependent on the sun and the existence of clouds throughout the day, so it is not easy to achieve methods that work well when disturbances are too variable and difficult to predict. This work proposes dynamic artificial neural networks (ANNs) that take into account past information and are not so sensitive to the variations of the plant at each moment. With this, three types of failures are distinguished: failures in the optical efficiency of the mirrors, flow rate, and thermal losses in the pipes. Different ANNs have been proposed and compared with a simple feedforward ANN, obtaining an accuracy of 73.35%.