Machine Learning (ML) for real-time Dynamic Security Assessment (DSA) promises a probabilistic approach to secure lower safety margins and costs. However, future systems with a high share of renewables have low inertia and converter-interfaced devices resulting in faster dynamics. Past research on ML-based DSA used high inertia systems to study ‘the best’ ML data, features, and models building upon each other's work for decades. Seldom has ML-based research for DSA questioned whether the underlying assumptions for (and the conclusions of) these studies are still valid for low inertia systems. This work studies exemplary changes in assumptions (and conclusions) for ML-based DSA when moving from High Inertia (HI) to Low Inertia (LI) systems. The dynamical system of the LI system is brought in perspective with the most typical ML-based approaches, which are organised in sequential steps. The steps consider the generation of the training database, the data pre-processing and feature selection, the model training and validation. This work analyses each step individually for the changed assumptions in the dynamical LI system, and subsequently, a case study provides the evidence that considering a LI system to identify the ‘best’ ML approaches is important. The case studies on IEEE 14 and 68 bus systems confirm that LI systems must be optimised for security (otherwise, they result in 80% less security than HI systems). The key findings, however, are that using ML makes significantly more sense in LI systems than in HI systems as the LI dynamics are in shorter timescales (and the advantage of ML is to predict security in milliseconds) and that secure/insecure operations can be separated more straightforwardly in LI systems as ML increases the accuracy by up-to 40% towards close to 100% when using neural networks.@en

Power systems transport an increasing amount of electricity, and in the future, involve more distributed renewables and dynamic interactions of the equipment. The system response to disturbances must be secure and predictable to avoid power blackouts. The system response can be simulated in the time domain. However, this dynamic security assessment (DSA) is not computationally tractable in real-time. Particularly promising is to train decision trees (DTs) from machine learning as interpretable classifiers to predict whether the system-wide responses to disturbances are secure. In most research, selecting the best DT model focuses on predictive accuracy. However, it is insufficient to focus solely on predictive accuracy. Missed alarms and false alarms have drastically different costs, and as security assessment is a critical task, interpretability is crucial for operators. In this work, the multiple objectives of interpretability, varying costs, and accuracies are considered for DT model selection. We propose a rigorous workflow to select the best classifier. In addition, we present two graphical approaches for visual inspection to illustrate the selection sensitivity to probability and impacts of disturbances. We propose cost curves to inspect selection combining all three objectives for the first time. Case studies on the IEEE 68 bus system and the French system show that the proposed approach allows for better DT-selections, with an 80% increase in interpretability, 5% reduction in expected operating cost, while making almost zero accuracy compromises. The proposed approach scales well with larger systems and can be used for models beyond DTs. Hence, this work provides insights into criteria for model selection in a promising application for methods from artificial intelligence (AI).@en