Applicability of Machine Learning to Short-Term Prediction of Changes in the Low Voltage Electricity Distribution Network

Abstract

Low voltage electricity distribution network actively maintains the stability of its key parameters, primarily against the predictable regularity of seasonal changes. This makes long-term coarse prediction practical, but it hampers the accuracy of a short-term fine-grained one. Such predictability can further improve the stability of the network. This paper presents the outcome of research to determine whether Machine Learning (ML) algorithms can improve the accuracy of the prediction of next-second values of three network parameters: voltage, frequency and harmonic distortions. Four ML models were tested: XGBoost Regressor, Dense neural networks (both one and two layer) and LSTM networks, against static predictors. Real data collected from the actual network were used for both training and testing. The challenging nature of this data is due to the network executing corrective measures, thus making parameter values return to their means. This results in non-normal distribution with strong long-term memory impact, but with no viable correlation to use for short-term prediction. Still, results indicate improvements of up to 20%, even for non-optimized ML algorithms, with some scope for further improvements.