The report describes the activities carried out to optimize DELFI (DEep Learning for False voltage dips Identification), a Data Analytics application created for the advanced analysis of Power Quality data using Deep Learning algorithms. The aim was to improve the performance of the model underlying the application, which classifies voltage sags based on different types of event: TRUE, FALSE, TRUE+FALSE, where ‘false’ voltage sags are not ‘real’ disturbances in the grids but are drops in the measured voltage due to the saturation of the measuring transformers in the primary cabin. The activity carried out allowed us to improve the performance of the initial model, which we revised with regard to: (i) optimisation of parameters (hyper-parameters); (ii) volume of data used for training the Deep Learning algorithm; (iii) choice of the type of classification algorithms adopted.

The effectiveness of the model was also checked against the temporal trends of the RMS of the voltages associated with the events that, unlike the waveforms, are normally stored by default in the monitoring systems. The activity led to the implementation of a final model (Model 4), based on the automated Bayesian optimisation method, as a result of a multi-stage evolution of the starting model. The obtained model, appropriately trained on a Training Set consisting of 80% of the images of RMS sequences associated with the events and taken from the database made available for the activity, provided a performance of 94%, understood as the percentage of correctly identified events on the total available number.

The result is also significant with reference to the voltage dip immunity curve adopted for class 3 equipment: the events below the curve are correctly classified with a performance of 100% for TRUE and TRUE+FALSE events and a performance of 92% for FALSE events. The level of success achieved is significant compared with the 80% performance offered by the criterion currently used in the QuEEN monitoring system and adopted in national monitoring. In the future, the identified classification model may be implemented in the QuEEN system, following intensive testing on input data acquired directly from the system itself and appropriate feasibility checks.