Advanced analytics and machine learning algorithms for predictive maintenance on Medium Voltage (MV) distribution network

Advanced analytics and machine learning algorithms for predictive maintenance on Medium Voltage (MV) distribution network The world of the energy production is quickly changing thanks to the advent of renewable sources of energy and the birth of new entities in the energy production chain like the prosumers. This is making the Distribution System Operator (DSO) more focused on what concerns the data they get from the distribution network in order to improve the system management. In this optics, the work of the thesis focuses on the analysis of SCADA events regarding Medium Voltage lines recorded from 2010 to 2016 in order to apply predictive maintenance. The idea is challenging since this data were not gathered with the purpose of apply machine learning algorithms on them. In particular, after a first significant phase of data exploration and data shaping fundamental for the thesis to be developed and for a future work on these data, two approaches of machine learning have been tried: 1) Cosine Similarity Algorithm Once defined a pre-fault window as the period of time immediately preceding an interruption on the line, the pre-fault windows of all the permanent faults were investigated in order to find a similarity among these time windows. To perform this task, the cosine similarity algorithm has been used which returns values among -1 (opposite vectors) and 1 (equal vectors) for the similarity. This algorithm evaluates through a vector multiplication the similarity between two vectors representing the events occurred in the pre-fault window. This approach gave poor results: indeed, looking at the pre-fault window of event of the same type, 25% of them had similarity equal to 0 (the lowest since we had only positive vectors) and only a 20% of them shown a correlation higher than 0.6. 2) Rule Mining Approach The Rule Mining approach is based on the concepts of pattern mining and rule extraction. The objective of the algorithm is to find the correlation between previously unknown patterns of recorded events and line interruptions in the dataset that can be useful in the prognosis of the line. Running this analysis on the events happening before a fault, the alleged cause code and the component affected by the interruption, this analysis returned high values of correlation among specific causes and specific components. It was then decided to pair this information as premises and look for the correlation of these pairs with the other events. Given the extraction of these rules, experts of the distribution lines defined which correlation among events was expected confirming then the goodness of the algorithm and at the same time new unknown rules were extracted. Moreover, this approach was also used to go from a prognosis problem (predict through the SCADA events the probability of an interruption) to a diagnosis problem (given SCADA events recorded, say which kind of interruption occurred on the line and on which component).