Predicting annual illuminance and operative temperature in residential buildings using artificial neural networks

Abstract

Simulation-based multi-objective optimizations is proving to be an effective approach for supporting the building design and finding a balance between daylight availability, thermal comfort and energy performance. This article investigates the potential for applying ANNs to predict both annual daylight illuminance and operative temperature, in order to reduce the overall simulation time. The main findings from deploying a simulation model is the importance of multi sensor-node calculations including long- and shortwave radiation for operative temperature. Operative temperature is usually calculated for the room center, in contrast to daylight where illuminance is calculated for a grid of sensor-nodes. The results show a significant difference in operative temperature at different locations in the room where shortwave radiation has greatest impact on the results. It is therefore important to address operative temperature in the same way as daylight illuminance, using a grid of sensor-nodes when exploring multi-objective optimization performance. However, these calculations are computational demanding and increase simulation time by 2000%. A fully connected neural network is developed with five hidden layers and five different neuron structures. In general, the ANN models are showing promising results which may be integrated in a multi-objective design workflow. The results show significant time saving potential by implementing ANNs. The overall time is reduced by 96% by using ANN models for predicting annual temperature and illuminance values.