Power Grid Frequency Forecasting from μPMU Data using Hybrid Vector-Output LSTM network

The instantaneous balance of electrical supply and demand on the power grid is indicated by the power grid frequency, making it a pivotal variable for power system controls. Accurate frequency forecasting could enable new faster means of frequency management that enhance power system stability. A hybrid vector-output Long Short-Term Memory (LSTM) neural network has been studied using microsynchrophasor data to predict trajectories. The objective of this research is to evaluate the effectiveness of very short time horizon frequency prediction using this method. The proposed model has been trained with over and under-frequency operational limit excursion events as well as normal condition state, with the goal of minimising prediction errors. Training and testing have been conducted using 390,000 datapoints covering 65 frequency events obtained from a distribution grid connected solar farm in England. The results demonstrate this method can provide useful grid frequency projections and shed light on underlying behaviour. Index Terms—Electrical grid frequency, power system stability, time series forecasting, long short term memory