Deep Neural Network Architectures for Component-Based Machine Learning Model in Building Energy Predictions
Sundaravelpandian Singaravel, Philipp Geyer, Johan A. K. Suykens · Lirias · 2017
Artificial Neural Networks (ANN) are a universal approximator for any non-linear function. However, ANN approximation strongly depends on the architecture, i.e. the structure of the neurons and the training methods. This paper evaluates ANN architectures to model components that represent a building for its energy prediction. ANN architectures evaluated are one-hidden layer neural network (NN), deep NN and stacked autoencoder deep NN. The performance of these ANN architectures is assessed against Random Forest models. Results indicate that ANN methods increase the performance in percentage of coefficient of determination (R2) between 0.57% to 9.65% compared to Random Forest models. Within ANN architectures evaluated, deep NN architectures performed better for most cases.