Short-Term Load Forecasting of Active and Reactive Powers in Small Scales using LASSO-Integrated Nonlinear Autoregressive Neural Networks

Abstract

Short-term load forecasting (STLF) is important for power system planning and optimization, especially in the dynamic environment of smart grid. Traditional load forecasting is implemented at substation levels to predict the upcoming active power and optimal system settings. In more advanced smart grid applications, e.g. the Volt-VAR Control, small-scale load forecasting opens up new opportunities in coordinating distributed resources such as distributed generation (DG) with utilities' efficiency missions. This paper proposes a STLF approach for small residential blocks with 10-12 households. The Nonlinear Autoregressive Neural Network (NAR-NN) is employed to predict hour-ahead active (P) and reactive (Q) powers with a moving window of training data. The regressor shrinkage technique, LASSO, is used to improve the selection of the regressors in the NAR-NN model by removing insignificant input features. The results show the forecasting performance could be enhanced by ~20% comparing to feed-forward Artificial Neural Networks (ANNs). The improvement in forecasting both P & Q could accommodate new smart grid applications in small scales.