Generating Feature Sets for Day-Ahead Load Demand Forecasting Using Deep Neural Network

Abstract

The performance of load demand forecasting plays a vital role in economic operation and planning in the power sector. There are several methodologies in the literature for predicting load. However, there is still an essential need to develop more accurate load forecast method. The performance of these methods can be improved by using an effective machine learning methods by selecting informative feature sets. In this paper, at first, we choose the effective time lags based feature by using auto-correlation and cross-correlation. Then, more robust features have been extracted by using Principal Component Analysis (PCA) and Autoencoder (AE) based Deep Neural Network (DNN). Extracted features are provided as an input to the Artificial Neural Network (ANN) model. ANN with Levenberg-Marquardt (LM) training algorithm has been used for day-ahead load forecasting (DALF) using the extracted features. The proposed method is AE based DNN for features extraction followed by ANN with LM training algorithm. The proposed method has been compared with ANN and PCA-ANN. The performance evaluation for DALF has been analyzed on two different substations of New England Independent System Operator (NE-ISO) dataset. Each dataset is analyzed for two separate cases. The performance of the proposed approach is better than ANN and PCA-ANN method.