In multi-component systems, degradation, maintenance, renewal and operational mode change continuously the operating conditions. The identification of the onset of abnormal conditions from signal measurements taken in such evolving environments can be quite challenging, due to the difficulty of distinguishing the real cause of the signal variations. In this work, we present a method for fault detection in evolving environments that uses a Sparse Autoencoder-based Deep Neural Network (SAE-DNN) and a novel procedure that remarkably reduces the computational burden for setting the values of the hyperparameters. The method is applied to a synthetic case study and to a bearing vibration dataset. The results show that it is able to accurately detect faults in multi-component systems, outperforming other state-of-the-art methods.
Citation: Pinciroli, L.; Baraldi, P.; Ballabio, G.; Compare, M.; Zio, E. Optimization of the Operation and Maintenance of renewable energy systems by Deep Reinforcement Learning. Renewable Energy 2022, 183, 752-763. https://doi.org/10.1016/j.renene.2021.11.052
Equipment of renewable energy systems are being supported by Prognostics & Health Management (PHM) capabilities to estimate their current health state and predict their Remaining Useful Life (RUL). The PHM health state estimates and RUL predictions can be used for the optimization of the systems Operation and Maintenance (O&M). This is an ambitious and challenging task, which requires to consider many factors, including the availability of maintenance crews, the variability of energy demand and production, the influence of the operating conditions on equipment performance and degradation and the long time horizons of renewable energy systems usage. We develop a novel formulation of the O&M optimization as a sequential decision problem and we resort to Deep Reinforcement Learning (DRL) to solve it. The proposed solution approach combines proximal policy optimization, imitation learning, for pre-training the learning agent, and a model of the environment which describes the renewable energy system behavior. The solution approach is tested by its application to a wind farm O&M problem. The optimal solution found is shown to outperform those provided by other DRL algorithms. Also, the approach does not require to select a-priori a maintenance strategy, but, rather, it discovers the best performing policy by itself.
The life cycle of wind turbines depends on the operation and maintenance policies adopted. With the critical components of wind turbines being equipped with condition monitoring and Prognostics and Health Management (PHM) capabilities, it is feasible to significantly optimize operation and maintenance (O&M) by combining the (uncertain) information provided by PHM with the other factors influencing O&M activities, including the limited availability of maintenance crews, the variability of energy demand and corresponding production requests, and the long-time horizons of energy systems operation. In this work, we consider the operation and maintenance optimization of wind turbines in wind farms woth multiple crews. A new formulation of the problem as a sequential decision problem over a long-time horizon is proposed and solved by deep reinforcement learning based on proximal policy optimization. The proposed method is applied to a wind farm of 50 turbines, considering the availability of multiple maintenance crews. The optimal O&M policy found outperforms other state-of-the-art strategies, regardless of the number of available maintenance crews.
In the International Thermonuclear Experimental Reactor, plasma is magnetically confined with Superconductive Magnets (SMs) that must be maintained at the cryogenic temperature of 4.5 K by one or more Superconducting Magnet Cryogenic Cooling Circuits (SMCCC). To guarantee cooling, Loss-of-Flow Accidents (LOFAs) in the SMCCC are to be avoided. In this work, we develop a three-step methodology for the prompt detection of LOFA precursors (i.e., those combinations of component failures causing a LOFA). First, we randomly generate accident scenarios by Monte Carlo sampling of the failures of typical SMCCC components and simulate the corresponding transient system response by a deterministic thermal-hydraulic code. In this phase, we also employ quick-running Proper Orthogonal Decomposition (POD)-based Kriging metamodels, adaptively trained to reproduce the output of the long-running code, to decrease the computational time. Second, we group the generated scenarios by a Spectral Clustering (SC) employing the Fuzzy C-Means (FCM), in order to identify the main patterns of system evolution towards abnormal states (e.g., a LOFA). Third, we develop an On-line Supervised Spectral Clustering (OSSC) technique to associate time-varying parameters measured during plant functioning to one of the prototypical groups obtained, which may highlight the related LOFA precursors (in terms of SMCCC components failures). We apply the proposed technique to the simplified model of a cryogenic cooling circuit of a single module of the ITER Central Solenoid Magnet (CSM). The framework developed promptly detects 95% of LOFA events and around 80% of the related precursors.