Short-term time-series forecasting of smart grid states using unsupervised clustering-assisted LSTM

The increasing penetration of renewable energy introduces significant spatiotemporal variability in both supply and demand patterns, complicating real-time grid state forecasting and dispatching in smart grids. In this paper, we propose an unsupervised clustering-assisted Long Short-Term Memory (LSTM) framework to improve short-term forecasting of power grid operational states from a supply-demand balance perspective. Load, solar irradiance, and wind speed are selected as key features and normalized using min-max scaling. The normalized time-series data are segmented into daytime and nighttime windows to capture diurnal operational differences, and further classified into typical and extreme scenarios based on thresholding. A K-Medoids clustering algorithm is applied to identify representative operational states. An LSTM model is then trained to learn the state change between cluster-labeled states. The proposed method is validated on synthetic grid operation data, demonstrating improved forecasting accuracy and robustness under operational uncertainties, particularly during extreme conditions. © 2025 IEEE.