In the context of the solar PV boom and the AI revolution, failure prediction and minimization become of great importance for the massification of renewable energies. Machine learning provides analysis capabilities unmatched by previous technologies, which is why it is increasingly being used in a vast number of fields. In the present study, a brief introduction to solar inverters is made, followed by a contextualization of different machine learning algorithms, methodologies, and state of the art in relation to failure prediction. A summary is made for a set of classification machine learning algorithms that were tested for failure prediction in solar inverters. The input dataset corresponds to failure data, together with temperature measurements of several components for the Riverstart solar power plant in Indiana, USA, property of EDP Renewables. Preprocessing of data was done, removing missing entries and scaling temperatures to avoid data stretching. An alternative dataset that considered time series data was produced considering two window-sizing methods: Partial Autocorrelation Function and Spectral Analysis. The results of the modified dataset were worse than the original one’s, reason why it was not further considered. The results show that the most appropriate algorithm for the classification task was the Artificial Neural Network, with a recall for the failure class of R = 0.81, while the worst performing ones were K Nearest Neighbors (R = 0.18), Decision Trees (R = 0.18), and Random Forests (R = 0.15). Other algorithm’s results were Logistic Regression: R=0.58; Gradient Booster Classifier: R = 0.50; and Support Vector Machines: R = 0.58.
Nel contesto del boom del fotovoltaico e della rivoluzione dell’IA, la previsione e la minimizzazione dei guasti diventano di grande importanza per la massificazione delle energie rinnovabili. Il Machine Learning (ML) fornisce capacità di analisi senza pari rispetto a tecnologie precedenti, motivo per cui viene sempre più utilizzato in un vasto numero di campi. In questo studio, si fa una breve introduzione agli inversori solari, seguita di una descrizione di diversi algoritmi di ML, metodologie e stato dell’arte in relazione previsione dei guasti. Si fa un riassunto di una serie di algoritmi di classificazione ML che sono stati testati per la previsione dei gusti negli inversori solari. Il dataset di input corrisponde ai dati di guasto, insieme alle misurazioni della temperatura di diversi componenti per l’impianto fotovoltaico di EDPR Riverstart in Indiana. Si è fatto un processamento dei dati, rimuovendo quelle mancanti e ridimensionando le temperature per evitare deformismi. Si è prodotto un dataset alternativo considerando i dati come serie temporale, seguendo due metodi diversi per il dimensionamento della finestra temporale: PACF e Analisi Spettrale. I risultati dell’dataset modificato sono stati peggiori di quelli dell’originale, motivo per cui non è stato ulteriormente considerato. I risultati mostrano che l’algoritmo più appropriato per il compito di classificazione e stata la Rete Neurale Artificiale, con un recall per la classe guasto di R = 0.81, mentre i peggiori sono stati KNN (R = 0.18), Decision Trees (R = 0.18), e Random Forests (R = 0.15). I risultati di altri algoritmi sono stati Logistic Regression: R=0.58; Gradient Booster Classifier: R = 0.50; e SVM: R = 0.58.
Predictive maintenance of solar inverters: a study on failure prediction using machine learning
Rosati Sanhueza, Felipe
2022/2023
Abstract
In the context of the solar PV boom and the AI revolution, failure prediction and minimization become of great importance for the massification of renewable energies. Machine learning provides analysis capabilities unmatched by previous technologies, which is why it is increasingly being used in a vast number of fields. In the present study, a brief introduction to solar inverters is made, followed by a contextualization of different machine learning algorithms, methodologies, and state of the art in relation to failure prediction. A summary is made for a set of classification machine learning algorithms that were tested for failure prediction in solar inverters. The input dataset corresponds to failure data, together with temperature measurements of several components for the Riverstart solar power plant in Indiana, USA, property of EDP Renewables. Preprocessing of data was done, removing missing entries and scaling temperatures to avoid data stretching. An alternative dataset that considered time series data was produced considering two window-sizing methods: Partial Autocorrelation Function and Spectral Analysis. The results of the modified dataset were worse than the original one’s, reason why it was not further considered. The results show that the most appropriate algorithm for the classification task was the Artificial Neural Network, with a recall for the failure class of R = 0.81, while the worst performing ones were K Nearest Neighbors (R = 0.18), Decision Trees (R = 0.18), and Random Forests (R = 0.15). Other algorithm’s results were Logistic Regression: R=0.58; Gradient Booster Classifier: R = 0.50; and Support Vector Machines: R = 0.58.File | Dimensione | Formato | |
---|---|---|---|
2024_04_Rosati.pdf
non accessibile
Dimensione
3.83 MB
Formato
Adobe PDF
|
3.83 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/217823