Neural Network Architectures for Electricity Consumption Forecasting

Abstract

Advancing Data Science has become a central goal in the era of Big Data, Cloud Computing, and the Internet of Things. Data science includes a multitude of methods, algorithms, and systems that are used to extract hidden patterns and knowledge from data. This paper is centered around machine learning (ML) models, and their implementation and applicability in the field of the energy informatics. Four neural networks (NNs) are proposed for computational experiments on electricity consumption forecasting. The NNs considered are back propagation neural network (BPNN), fully recurrent neural network (fRNN), long short-term memory network (LSTM), and gated recurrent unit (GRU). The four models are implemented using PyTorch, an open-source ML framework. By comparing their predictive performance, we find that LSTM and GRU attain lower errors than the other two models. More precisely, GRU presents better performance for the daily forecast whereas LSTM outperforms the rest for the monthly forecast. Insights about PyTorch are also provided, as it represents a recent addition to state of the art technologies, in terms of implementation.