Paper Conference
Proceedings of eSim 2022: 12th Conference of IBPSA-Canada
![]() ![]() ![]() ![]() |
A comparison of machine learning functions for time-series prediction in buildings
Blair Birdsell 1, Ralph Evins 21 Energy in Cities Lab, Victoria, Canada2 Energy in Cities Lab, Victoria, CanadaAbstract: This work undertakes a comprehensive comparison of RNNs and CNN-based ResNets for both multivariable day-ahead and annual predictions with specific focus on their application in buildings. This varied comparison of two types of nets under two types of prediction conditions documents and describes methods that will be impactful to building designers and operators. The comparison shows that both CNN-based ResNets and RNNs are suitable for short-term forecasting, however, the study establishes RNN’s strength in forecasting sequences. This has implications when trying to predict the short-term behaviour of building systems.
When the annual performance was estimated, CNN-ResNets showed a distinct advantage in accurately predicting the expected monthly minimum and maximum values. This result can contribute to better decision making in the design and planning process. Keywords: Machine learning. Building optimization. Modeling & Simulation. Smart Buildings. IoT.Paper:esim2022_238