Failure warning system for individual electric power components without component-specific sensor data

With the expansion of machine learning, many failure detection algorithms have been developed. However, existing models are not applicable in the field of power systems, because practical limitations of grid operators are not considered. For instance, part of the models rely on componentspecific sensor data. Owing to the size of power systems among others, it is unrealistic to install sensors on all grid components. We investigate the possibility to predict the failure of a grid component without using component-specific sensors. To this goal, an automatization algorithm is developed, that selects pre-processing methods of input data and optimal hyperparameters of machine learning models. It leverages commonly available operational data to generate warnings when there is a risk of failure of a specific grid component. Autoencoders provide satisfactory results when suitable choices are made concerning the performance criteria, and the data pre-processing, including the number of lags and sliding window steps and the use of binary variables. Finally, we discuss ways to improve the algorithm, notably methods for data augmentation, cross validation, and strategies to cope with the trade-off between false alarms and failure detection.