Machine learning-based frequency security early warning considering uncertainty of renewable generation

Abstract Loss of massive generation caused by HVDC blocking or tripping large power plants is a severe threat to receiving-end grids' frequency security, especially those with a high penetration level of renewable generation. Frequency security early warning is necessary to send warning messages in time. With the warning messages, appropriate measures can be taken in advance to minimize possible losses. In this paper, a machine learning-based frequency security early warning method considering the uncertainty of renewable generation is proposed. It includes three core parts: future scenario generation, assessment model establishment, and early warning indicator establishment. In the future scenario generation part, Markov Chain Monte Carlo (MCMC) is combined with Generative Adversarial Networks(GAN) for generating numerous future scenarios reflecting possible future operation modes of the system, considering the uncertainty of renewable generation and loads. In the assessment model establishment part, the assessment model with clustering based on metric learning is applied to establish the machine learning-based frequency security assessment model. The model is continuously retrained with Domain Adaptation Metric Learning (DAML) and a transitive closure-based constraint propagation clustering approach to improve assessment accuracy. Future frequency security risk indicators are established in the early warning indicator establishment part based on future scenarios' assessment results. According to the risk indicators, future frequency security is classified into different early warning levels. The future frequency security can be expressed clearly and intuitively with the early warning levels. A simplified provincial power system of China is adopted as an example to verify the validity of the proposed early warning method.