The following report provides concrete results on how time series analysis techniques, especially those based on artificial intelligence, can help the electro-energy system, supporting the improvement of planning and operation processes envisaged by utilities. The main goal is to demonstrate that data-driven approaches can help utilities in their daily task of providing increasingly resilient networks. The activity of this year focused on concrete use cases and the application of data-driven techniques to solve problems that are usually addressed with huge human resources (e.g. image and/or video analysis), or with algorithms that provide sufficient results to comply with regulatory requirements (e.g. on-load tap changer) but are not efficient from an economic and/or performance point of view, or employ short-term forecasting analyses that do not take into account possible context changes due, for instance, to emergency situations, such as pandemics and environmental disasters.
A pipeline has been developed to analyze daily active power profiles of secondary substations in Milan, capable of taking as input any type of time series dataset and providing as output the classification of time profiles into different clusters. The activity on the analysis of videos from the LANPRIS experiment for the recognition of surface discharges on high voltage insulators continued using Computer Vision techniques and Deep Learning techniques.
The analysis of data on the Milan MV distribution network operated by Unareti during the Covid-19 epidemic was extended to the end of 2020, integrating the analysis of reactive power and the incidence of faults on the distribution network , with the aim of investigating a possible link between the characteristics of the interruptions and the trend of electrical or meteorological variables (e.g. heat waves).
A concrete example of the application of machine learning techniques to the resolution of a practical problem considering the management of the on-load tap changer (OLTC) of a primary substation was addressed.
The algorithm was evaluated by simulating the Vobarno MV network operated by Unareti.
With regard to forecasting models based on data streaming, the occupancy forecasting model of electric vehicle charging systems was studied in depth, by analyzing the influence of some model parameters on the forecast result, and the architecture for a system to predict faults on the distribution network that integrates continuous data flows with semantic information on the network was prepared. Graph modeling of energy networks and the application of graph analytics and graph machine learning techniques have been used for classification and regression operations on entire energy networks. Graph-based powerflow resolution algorithms were explored in depth, to facilitate a more extensive comparison between traditional methods and innovative graph-based methods.