Random Forests and Artificial Neural Network for Predicting Daylight Illuminance and Energy Consumption

Abstract

Predicting energy consumption and daylight illuminance plays an important part in building lighting control strategies. The use of simplified or datadriven methods is often preferred where a fast response is needed e.g. as a performance evaluation engine for advanced real-time control and optimization applications. In this paper, we developed and then compared the performance of the widely-used artificial neural network (ANN) with random forest (RF), a recently developed ensemble-based algorithm. The target application was predicting the hourly energy consumption and daylight illuminance values of a classroom in Cardiff, UK. Overall, RF performed better than ANN for predicting daylight illuminance; with coefficients of determination (R2) of 0.9881 and 0.9799 respectively. On the energy consumption testing dataset, ANN performed marginally better than RF with R2 values of 0.9973 and 0.9966 respectively. RF performs internal cross-validation and is relatively easy to tune as it has few tuning parameters. The paper also highlighted possible future research directions.