Ph. D. Project
AI-based predictive maintenance for a photovoltaic power plant
2023/05/02 - 2026/05/01
Solar photovoltaic (PV) power plants are one the most promising solutions for generating renewable electricity using sunlight. To maximize their energy production, PV plants are maintained according to a maintenance program with both preventive and corrective operations. Due to the lack of adequate decision support tools, preventive maintenance actions are often planned according to the age of the equipment to be maintained. An inadequate maintenance schedule can lead to major disadvantages (e.g. unnecessary preventive maintenance operations, high failure frequency, random/reduced performance, additional maintenance/operating costs). In this context, Feedgy would like to move toward predictive maintenance so that maintenance decisions will be made based on the prediction of the health state of the PV plant's equipment under a specific operation profile. The deployment of this emerged maintenance strategy is based on the two main processes: (1)-fault/failure prediction of the PV's equipment (e.g., inverters, circuit-breakers, photovoltaic panels, etc.) and (2)-optimization of the predictive maintenance decisions for the equipment.
To answer these industrial challenges, the first objective of this thesis is to develop prognostic approaches using Machine Learning to predict the failure date of the critical equipment of a PV plant. Indeed, in the literature, several prognostic approaches have been proposed to predict the failure date of an equipment/system. However, the existing approaches are mostly proposed for production systems or machine tools. They are not directly applicable to specific equipment such as those of a PV plant. Moreover, these equipment's ageing/degradation process is very sensitive to their location and the weather condition. Therefore, the development of such a prognostic approach for predicting the faut/failure of a PV plant's equipment remains an important scientific challenge to be solved.
The second objective of this thesis is to propose/develop a predictive maintenance strategy for a fleet of PV plants based on the prediction results obtained from the previous step. The developed strategy should consider the different properties of the PV plants (location, configuration ...), the operating conditions as well as specific constraints related to both the production planning and the logistic support. A multi-criteria optimization (e.g. minimizing maintenance total costs and/or maximizing production power) with specific constraints will be developed, based on conventional optimization algorithms or reinforcement learning algorithms, to search for the optimal maintenance schedule. Actually, several decision support models (maintenance strategies and optimization algorithms) have been developed in the literature. However, the development and implementation of adequate solutions on an industrial scale still lack foundations, methods and tools.
The major scientific originalities of this thesis are the following:
• Develop advanced prognostic algorithms to predict the date of faults/failure of the key equipment of a PV plant. The developed algorithms must allow exploiting both the available data and the industry knowledge (towards a prognosis with a relevant degree of confidence).
• Develop an advanced decision support model for predictive maintenance based on the prognostic results.
• Develop a multi-criteria optimization approach providing an adapted maintenance solution considering the properties and the operating conditions of the PV plants and the constraints related to production scheduling and logistic supports.
• Develop a model to evaluate the effectiveness of the proposed predictive maintenance solution.
• Implement the proposed models in commercial software.
Photovoltaic power plants, machine learning, predictive maintenance, prognostics, RUL, system reliab
Eco-Technic systems engineering