Recurrent neural networks for electrical load forecasting to use in demand response

Abstract

Electric load forecasting is a fundamental technique to understand end-user behavior and therefore a crucial factor in the design of demand response (DR) programs. Load forecasting will also identify the appropriate design of DR programs. In this chapter, a range of different machine learning applications are studied to represent the influential factors for electrical load demand forecast in a DR context, with a variety of different data scenarios, temporal and technical scenario. This chapter explores and compares the load prediction analysis through basic recurrent neural networks (RNNs); Vanilla RNN, gated recurrent units (GRU), and long short-term memory (LSTM), using principal component analysis (PCA). It is found that PCA can be used to reduce the number of principal components for Vanilla RNN, GRU, and LSTM networks. Reducing the number of principal components using PCA is one of the techniques that is used in dimensionality reduction. Reduction in dimensionality will relieve the computational burden. In this work, the dimensionality reduction improves the predictive output. It is observed that for electric load demand forecasting, the preferred technique is GRU, trained with a principal component. The performance is evaluated through mean absolute percentage error (MAPE), which is relatively lower than other techniques.