Detecção de falhas em sistemas fotovoltaicos a partir da análise dos sinais de geração

Abstract

The production of energy through photovoltaic panels is a well-known technology, and recently its cost has reduced to the point where it is feasible for domestic users to install systems for personal use in their homes, smallholdings, micro-enterprises, and the like. However, installations still present a high cost. For this reason, monitoring energy generation and performing maintenance are necessary to ensure a long-lasting and optimised system operation. Various systems have been developed in an attempt to monitor performance and predict failures, using several techniques and monitoring various variables. This work observes the attempt to detect and predict failures through AI software based on the use of an AutoEncoder. Neural networks are complex elements in software development, and AutoEncoders behave in a specific and different manner compared to other types of neural networks. Therefore, a study on how neural networks function, how they "learn" and how their results are evaluated was presented. Since the theme of this project is AutoEncoders, greater emphasis was given to them, defining what they are, what differentiates them from other networks, their architecture and operation, as well as evaluating articles in the field of fault detection using AutoEncoders in other types of signals. Additionally, the operation of a photovoltaic energy generation system, its installation, maintenance, measurement of the amount of energy generated, and the quality of this generation are also presented. The importance of preventive and predictive maintenance to maximise the efficiency and lifespan of photovoltaic systems was also discussed. The detection system project began with obtaining a DataSet containing measurements of the generation level along with various variables about the moment of this generation. The second stage involved creating the training and testing sets, as well as developing the AutoEncoder and defining the performance evaluation metric, which is the Kullback-Leibler (KL) divergence. The next stage focused on training the model with different numbers of layers and configurations. Finally, the configurations and their respective errors were graphically displayed to facilitate analysis and comparison. Moreover, the feasibility of large-scale application and potential future improvements to the AutoEncoder model were analysed. We concluded that the results were satisfactory when evaluating the divergence of the data after the deconstruction and reconstruction by the AutoEncoder, comparing them with the original dataset data, demonstrating the method's effectiveness in fault detection. [resumo fornecido pelo autor]