Explainable Machine Learning for Real-Time Frequency Stability Assessment

The ongoing decarbonisation of electrical power systems has led to the widespread integration of new technologies, including renewable generation, high voltage direct current (HVDC) interconnectors, batteries, and electric vehicles. This transition introduces a two-fold challenge for secure and stable system operation: i) renewables are typically stochastic, and when coupled with evolving load profiles, they introduce uncertainty, leading to a vast number of scenarios requiring assessment; and ii) the system is becoming increasingly complex, making conventional methods based on solving differential algebraic equations computationally prohibitive for near-real-time applications. Existing near-real-time dynamic security assessment tools evaluate multiple contingencies using parallel time-domain simulations but remain restricted in the number of scenarios that can be explored within operational time constraints. As a result, operators are often limited to assessing a subset of scenarios or relying on simplified models to identify operational boundaries related to system stability, which can lead to suboptimal dispatches and, in the worst case, unforeseen blackouts. In this work, we investigate the application of a machine learning (ML)-based methodology to investigate frequency stability-related case studies on the 36-zone reduced GB power system model. In particular, we study two representative days: one in which the system exhibits significant stability margin, suggesting operators have scope to push the system closer to the limits for reasons of cost and carbon efficiency; and another case exhibiting stressed conditions, exemplifying the capability of the ML model to identify rare scenarios that could result in stability violations. The ML-based approach improves situational awareness, understanding and decision support by capturing detailed dynamics at low computational cost. The proposed ML framework provides near-real-time visibility of (locational) frequency stability metrics such as the nadir, enabling the dynamic security assessment of a large number of scenarios. Importantly, we go beyond the notion of ML as a mere black-box predictor through explainability, thereby building the necessary trust for real-world applications. The dataset is created through probabilistic sampling, considering uncertainties in day-ahead forecasts of wind generation and demand, as well as simulating a disturbance (the loss of the largest generation in this case). The inputs to the ML models are the expected or current operating conditions, and the output is the predicted minimum frequency nadir on the system (across all locations), following the disturbance for which the model has been trained.