Very Short-Term Feeder Load Forecasting Using Deep Learning for Real-Time Stability Monitoring in Smart Power Grids

Accurate forecasting of aggregate power demand from substation feeders is essential for active monitoring and control of the smart power grid. The rise in grid integration of renewable power sources necessitates accurate forecasting of the grid's stochastic characteristics. In this study, the historical power parameters measured at the outgoing point of a distribution substation feeder were employed for time series forecasting of very short-term aggregate feeder active power demand. Three deep learning algorithms, temporal convolutional network (TCN), long short term memory (LSTM) and gated recurrent unit (GRU) were employed, and the genetic algorithm (GA) was used for the models' hyperparameter tuning. Three forecast horizons including 1 (5 minutes), 3 (15 minutes) and 6 (30 minutes) steps ahead were evaluated. The proposed models were assessed for generalizability using three diverse real-world datasets with diverse customer profiles, demand patterns, and voltage levels. The performance of the deep learning models was evaluated using R-squared, root mean squared error (RMSE), mean absolute error (MAE), and weighted mean absolute percentage error (wMAPE). The results showed that the load type, the forecast horizon, and the rate of demand fluctuation significantly influence the performance of very-short-term feeder load forecasting (LF). The TCN model showed superior performance with a 97.5% R2-score (6 steps forecast), compared to the LSTM's (94.3%) and GRU's (94.8%), especially as feeder load demand fluctuation and forecast horizon increases. It also showed faster learning of the relationship between time series inputs and target variables.