Building Electrical Load Forecasting through Neural Network Models with Exogenous Inputs

As buildings become key actors in the economic and sustainable operation of future electrical grids and smart cities, reliable models which capture the underlying electrical energy consumption become an important factor for robust control algorithms. Current ubiquitous field devices supported by complex data infrastructures allow generation, storage and online analysis of large quantities of data for deriving usable black-box models of building energy patterns. The paper presents an approach to model the energy consumption of medium and large sized buildings using Non-linear Autoregressive Neural Networks with eXogenous Input (NARX). We show that the chosen network architectures offers good performance for time series prediction from historical values and external input signals such as outdoor temperature in comparison to a baseline approach. Model evaluation and validation are carried out on public dataset for replicable research outcomes.