Cumulant Learning: Highly Accurate and Computationally Efficient Load Pattern Recognition Method for Probabilistic STLF at the LV Level

This paper proposes a new load pattern recognition method for probabilistic short-term load forecasting to facilitate the management of low voltage networks and account for future load uncertainties based on large volumes of smart meter data. The proposed method, Cumulant Learning, is based on clustered loads approximated with cumulants. In this way, the size of the load data model is reduced without losing key load fluctuation patterns. For that purpose, the presented method uses deep learning (to predict the future cumulants) and the Cornish-Fisher expansion (to approximate quantiles for similar loads with high accuracy and computational efficiency). The usefulness of the proposed method is demonstrated in a case study on real smart meter data from two essentially different distribution networks in the U.K. and Australia. The case study results show that the proposed method leads to high forecast accuracy at the level of individual low-voltage consumers, with high data reduction and short execution time compared with related methods. The results also show that the proposed method is more robust to outliers and better at predicting load surges, making it suitable for quantifying future load uncertainties for higher probability values.