Towards Effective Datasets for Training Data-driven Models for Smart Grid Security Assessment

Due to the decentralisation of energy resources and the electrification of the heat and transport sectors, it is becoming increasingly more difficult for network operators to supervise and manage smart grids. Research points at the development of decentralised energy management tools (e.g., based on artificial intelligence) to develop strategies for the individual decision-making of numerous agents (e.g., smart electric vehicle charging), that aggregated respect the physical constraints of the grid. To train these tools it is therefore necessary to foresee problematic operational states (e.g., under/overvoltage, exceeding the rated powers of line/transformer) which could damage equipment, trigger protections, and cause service disruptions. While this can be done through power flow simulations (i.e., solving a system of non-linear physical equations with a numerical solver [1]), their high computational cost hinders computing speed. An alternative is using a traditional power system simulator to generate labelled datasets in an off-line stage, and then training a data-driven model to act as a surrogate for a fraction of the computational cost of the conventional tools [2]. In this manner, the computing time associated with the training of the AI-based energy management tools is greatly reduced. We compared three state of the art data-driven approaches to identify if an operational point (i.e., electricity demand at each consumption point) is classified as "safe" or "unsafe" (i.e., absence/presence of congestions) [3]. To this end, we propose and evaluate novel data generation strategies, and compare them to the standard random generation approach in the literature, highlighting the inadequacy of the latter when applied to low voltage networks. This is tested on the IEEE European Low-Voltage Test Feeder, with reports on computational times for training and inference, as well as significant improvements in classification performance using alternative data generation strategies. This work is conducted as part of the ANR "EDEN4SG" project under grant agreement ANR-22-CE05-0023.