Short-term Power Forecasting Model Based on Dimensionality Reduction and Deep Learning Techniques for Smart Grid

Abstract

This paper evaluates the performance of different feature extraction or dimensionality reduction techniques for the applications of short-term power forecasting using smart meters' data. The number and data type of input features are crucial to the performance of power forecasting models. The performance of the machine learning models decreases with the increase in the number of input features. That is, the machine learning models tend to overfit, and the forecasting accuracy is reduced. The performance of the feature extraction or dimensionality reduction techniques has been evaluated in the context of the forecasting applications with models involving Artificial Neural Networks (ANN), Long Short-Term Memory (LSTM), and Linear Regression (LR). The application is day-ahead forecasting on a real and open dataset of energy utilization by households in England. The obtained results depict the importance of dimensionality reduction techniques for higher accuracy and faster training times. While linear Principal Component Analysis (PCA) is a preferred dimensionality reduction technique for faster training times, kernel PCA, Non-negative Matrix Factorization (NMF), Independent Component Analysis (ICA) and Uniform Manifold Approximation and Projection (UMAP) yield better accuracies.