Load forecasting to implement smart grid functions

Thumbnail Image



Journal Title

Journal ISSN

Volume Title


University of New Brunswick


Local power distribution networks (LDN), like Saint John Energy (SJE), need to reduce peak electricity demands during their billing cycle to lower the electricity costs and reduce the environmental footprint. The introduction of distributed energy resources to the grid is changing the load demand shape and lowering the effectiveness of the traditional load demand forecasting methods. Therefore, utilities and researchers are searching for Machine Learning (ML) methods, which can adapt better to the changes in load demand behavior. Two sets of ML-based techniques were implemented to forecast the load demand on SJE’s LDN, both are based on eXtreme Gradient Boosting (XGBoost). Two other techniques were implemented for benchmarking, they are based on Seasonal Autoregressive Moving Average (SARIMA) and Long Short-Term Memory (LSTM). The results show that the proposed XGBoost-based techniques outperform the SARIMA-based and LSTM-based benchmarks on up to 96% and 94%, respectively, according to the skill score metric. Additionally, two of them are providing practical real-world by forecasting daily the load demand of Saint John city, they are under assessment by SJE.