Frequency Nadir and RoCOF Prediction Using Predisturbance Data and Artificial Neural Networks

Power system frequency stability is vital to keep a reliable electricity supply. However, the ongoing transition to cleaner electrical networks is giving place to system dynamics characterized by faster/pronounced frequency deviations and higher rates of change of frequency (RoCoFs), which impose significant challenges to grid operators and increase the risk of blackouts. Although traditional time-domain simulations have been widely used for frequency stability analysis, they are computationally intensive and in general not suitable for real-time applications. Therefore, a machine learning based alternative using artificial neural networks (ANNs) is proposed in this paper. By using predisturbance data, an ANN model is trained to predict the frequency nadir and RoCoF of a sample system before the occurrence of a given generation loss. Besides, in order to identify the most influential information (as related to the measuring points for the task), permutation feature importance is applied to the trained network. Then, a new ANN model with a reduced set of input features is built and implemented in a real-time environment for further performance verification. Offline and real-time simulation results reveal the capability of the proposed approach to support the anticipation of considered disturbances and provide relatively fast and accurate predictions of the concerned frequency metrics.