Training Data Generation Strategies for Data-driven Security Assessment of Low Voltage Smart Grids

Control of small-scale resources in low and medium voltage electricity networks is being decentralised, which increases the need and frequency of use of smart grid security assessment tools. This paper compares three data-driven approaches to classify if a smart grid is "safe" or "unsafe" (i.e., if grid constraints are respected) given an operational point as input: decision trees, gradient tree boosting and deep neural networks. Five novel training data generation strategies are proposed as alternatives to the standard random generation approach, aiming for data-driven models that generalise realistic scenarios better. Simulations are conducted using the IEEE European low voltage test network. Trained models are tested following trends from the literature and using realistic scenarios from the test network documentation, and electric vehicle charging patterns. Our results highlight the inadequacy of the current training data generation strategy, and offer better-performing alternatives. At last, we report on computational times dedicated to training our models, and discuss potential implications for future data-driven smart grid applications.